ggml : remove SVE paths

.devops/intel.Dockerfile

@@ -1,8 +1,8 @@
-ARG ONEAPI_VERSION=2025.2.2-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04
 
 ## Build Image
 
-FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
 
 ARG GGML_SYCL_F16=OFF
 RUN apt-get update && \
@@ -31,7 +31,7 @@ RUN mkdir -p /app/full \
     && cp requirements.txt /app/full \
     && cp .devops/tools.sh /app/full/tools.sh
 
-FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS base
 
 RUN apt-get update \
     && apt-get install -y libgomp1 curl\

.devops/nix/package.nix

@@ -128,6 +128,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   };
 
   postPatch = ''
+    substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
+      --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+    substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
+      --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
   '';
 
   # With PR#6015 https://github.com/ggml-org/llama.cpp/pull/6015,

.devops/rocm.Dockerfile

@@ -1,8 +1,8 @@
 ARG UBUNTU_VERSION=24.04
 
 # This needs to generally match the container host's environment.
-ARG ROCM_VERSION=7.0
-ARG AMDGPU_VERSION=7.0
+ARG ROCM_VERSION=6.4
+ARG AMDGPU_VERSION=6.4
 
 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
@@ -13,8 +13,9 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-# gfx803, gfx900, gfx906, gfx1032, gfx1101, gfx1102,not officialy supported
-# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
+# gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
+# gfx906 is deprecated
+#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
 
 ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
 #ARG ROCM_DOCKER_ARCH='gfx1151'
@@ -35,10 +36,13 @@ WORKDIR /app
 
 COPY . .
 
+RUN git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
     cmake -S . -B build \
         -DGGML_HIP=ON \
         -DGGML_HIP_ROCWMMA_FATTN=ON \
+        -DCMAKE_HIP_FLAGS="-I$(pwd)/rocwmma/library/include/" \
         -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
         -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
         -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \

.github/actions/install-exe/action.yml

@@ -1,36 +0,0 @@
-name: "Install exe"
-description: "Download and install exe"
-inputs:
-  url:
-    description: "URL of the exe installer"
-    required: true
-  args:
-    description: "Installer arguments"
-    required: true
-  timeout:
-    description: "Timeout (in ms)"
-    required: false
-    default: "600000"
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install EXE
-      shell: pwsh
-      run: |
-        $ErrorActionPreference = "Stop"
-        write-host "Downloading Installer EXE"
-        Invoke-WebRequest -Uri "${{ inputs.url }}" -OutFile "${env:RUNNER_TEMP}\temp-install.exe"
-        write-host "Installing"
-        $proc = Start-Process "${env:RUNNER_TEMP}\temp-install.exe" -ArgumentList '${{ inputs.args }}' -NoNewWindow -PassThru
-        $completed = $proc.WaitForExit(${{ inputs.timeout }})
-        if (-not $completed) {
-            Write-Error "Installer timed out. Killing the process"
-            $proc.Kill()
-            exit 1
-        }
-        if ($proc.ExitCode -ne 0) {
-            Write-Error "Installer failed with exit code $($proc.ExitCode)"
-            exit 1
-        }
-        write-host "Completed installation"

.github/actions/linux-setup-spacemit/action.yml

@@ -1,20 +0,0 @@
-name: "Linux - Setup SpacemiT Toolchain"
-description: "Setup SpacemiT Toolchain for Linux"
-inputs:
-  path:
-    description: "Installation path"
-    required: true
-  version:
-    description: "SpacemiT toolchain version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup SpacemiT Toolchain
-      id: setup
-      uses: ./.github/actions/unarchive-tar
-      with:
-        url: https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v${{ inputs.version }}.tar.xz
-        path: ${{ inputs.path }}
-        strip: 1

.github/actions/linux-setup-vulkan/action.yml

@@ -1,20 +0,0 @@
-name: "Linux - Setup Vulkan SDK"
-description: "Setup Vulkan SDK for Linux"
-inputs:
-  path:
-    description: "Installation path"
-    required: true
-  version:
-    description: "Vulkan SDK version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup Vulkan SDK
-      id: setup
-      uses: ./.github/actions/unarchive-tar
-      with:
-        url: https://sdk.lunarg.com/sdk/download/${{ inputs.version }}/linux/vulkan_sdk.tar.xz
-        path: ${{ inputs.path }}
-        strip: 1

.github/actions/unarchive-tar/action.yml

@@ -1,27 +0,0 @@
-name: "Unarchive tar"
-description: "Download and unarchive tar into directory"
-inputs:
-  url:
-    description: "URL of the tar archive"
-    required: true
-  path:
-    description: "Directory to unarchive into"
-    required: true
-  type:
-    description: "Compression type (tar option)"
-    required: false
-    default: "J"
-  strip:
-    description: "Strip components"
-    required: false
-    default: "0"
-
-runs:
-  using: "composite"
-  steps:
-    - name: Unarchive into directory
-      shell: bash
-      run: |
-        mkdir -p ${{ inputs.path }}
-        cd ${{ inputs.path }}
-        curl --no-progress-meter ${{ inputs.url }} | tar -${{ inputs.type }}x --strip-components=${{ inputs.strip }}

.github/actions/windows-setup-rocm/action.yml

@@ -1,15 +0,0 @@
-name: "Windows - Setup ROCm"
-description: "Setup ROCm for Windows"
-inputs:
-  version:
-    description: "ROCm version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup ROCm
-      uses: ./.github/actions/install-exe
-      with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-WinSvr2022-For-HIP.exe
-        args: -install

.github/workflows/build-amd.yml

@@ -1,52 +0,0 @@
-name: CI (AMD)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-amd.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh',
-      '**/*.comp'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  ggml-ci-x64-amd-vulkan:
-    runs-on: [self-hosted, Linux, X64, AMD]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
-
-  ggml-ci-x64-amd-rocm:
-    runs-on: [self-hosted, Linux, X64, AMD]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          amd-smi static
-          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

.github/workflows/build-cache.yml

@@ -1,89 +0,0 @@
-name: Build Actions Cache
-
-on:
-  workflow_dispatch: # allows manual triggering
-  schedule:
-    - cron: '0 * * * *'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  ubuntu-24-vulkan-cache:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Setup Cache
-        uses: actions/cache@v4
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
-
-  ubuntu-24-spacemit-cache:
-    runs-on: ubuntu-24.04
-
-    env:
-      # Make sure this is in sync with build-linux-cross.yml
-      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Setup Cache
-        uses: actions/cache@v4
-        id: cache-toolchain
-        with:
-          path: ./spacemit_toolchain
-          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
-
-      - name: Setup SpacemiT Toolchain
-        if: steps.cache-toolchain.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-spacemit
-        with:
-          path: ./spacemit_toolchain
-          version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
-
-  windows-2022-rocm-cache:
-    runs-on: windows-2022
-
-    env:
-      # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Setup Cache
-        uses: actions/cache@v4
-        id: cache-rocm
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: Setup ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
-        uses: ./.github/actions/windows-setup-rocm
-        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}

.github/workflows/build-linux-cross.yml

@@ -253,45 +253,3 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
-
-  ubuntu-24-riscv64-cpu-spacemit-ime-cross:
-    runs-on: ubuntu-24.04
-
-    env:
-      # Make sure this is in sync with build-cache.yml
-      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
-
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Use SpacemiT Toolchain Cache
-        uses: actions/cache@v4
-        id: cache-toolchain
-        with:
-          path: ./spacemit_toolchain
-          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
-
-      - name: Setup SpacemiT Toolchain
-        if: steps.cache-toolchain.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-spacemit
-        with:
-          path: ./spacemit_toolchain
-          version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
-
-      - name: Build
-        run: |
-          export RISCV_ROOT_PATH=${PWD}/spacemit_toolchain
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DGGML_CPU_RISCV64_SPACEMIT=ON \
-                         -DGGML_RVV=ON \
-                         -DGGML_RV_ZFH=ON \
-                         -DGGML_RV_ZICBOP=ON \
-                         -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
-                         -DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake
-
-          cmake --build build --config Release -j $(nproc)

.github/workflows/build-riscv-native.yml

@@ -58,63 +58,3 @@ jobs:
             -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
-
-  # debian-13-riscv64-spacemit-ime-native: # Bianbu 2.2
-  #   runs-on: [self-hosted, RISCV64]
-
-  #   steps:
-  #     - name: Install prerequisites
-  #       run: |
-  #         sudo apt-get update || true
-  #         sudo apt-get install -y libatomic1
-  #     - uses: actions/checkout@v4
-  #     - name: Setup Riscv
-  #       run: |
-  #         sudo apt-get update || true
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 gcc-14-riscv64-linux-gnu \
-  #                 g++-14-riscv64-linux-gnu \
-  #                 ccache \
-  #                 cmake
-  #         sudo apt-get upgrade binutils -y
-
-  #     - name: Setup ccache
-  #       run: |
-  #         mkdir -p $HOME/.ccache
-  #         ccache -M 5G -d $HOME/.ccache
-  #         export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
-  #         export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
-  #         echo "$GITHUB_WORKSPACE"
-  #         echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
-  #         echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
-  #         echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
-  #         echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
-
-  #     - name: Build
-  #       run: |
-  #         cmake -B build \
-  #           -DLLAMA_CURL=OFF \
-  #           -DCMAKE_BUILD_TYPE=Release \
-  #           -DGGML_OPENMP=OFF \
-  #           -DLLAMA_BUILD_EXAMPLES=ON \
-  #           -DLLAMA_BUILD_TOOLS=ON \
-  #           -DLLAMA_BUILD_TESTS=OFF \
-  #           -DCMAKE_SYSTEM_NAME=Linux \
-  #           -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-  #           -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-  #           -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-  #           -DCMAKE_C_COMPILER_LAUNCHER=ccache \
-  #           -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-  #           -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-  #           -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-  #           -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-  #           -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-  #           -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH \
-  #           -DGGML_RVV=ON \
-  #           -DGGML_RV_ZFH=ON \
-  #           -DGGML_RV_ZICBOP=ON \
-  #           -DGGML_CPU_RISCV64_SPACEMIT=ON \
-  #           -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1
-
-  #         cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -97,7 +97,7 @@ jobs:
           ctest -L 'main|curl' --verbose --timeout 900
 
   macOS-latest-cmake-x64:
-    runs-on: macos-15-intel
+    runs-on: macos-13
 
     steps:
       - name: Clone
@@ -207,7 +207,7 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-cpu-cmake-${{ matrix.build }}
+          key: ubuntu-cpu-cmake
           evict-old-files: 1d
 
       - name: Build Dependencies
@@ -362,11 +362,11 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.16
-      #   with:
-      #     key: ubuntu-latest-cmake-rpc
-      #     evict-old-files: 1d
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-latest-cmake-rpc
+          evict-old-files: 1d
 
       - name: Dependencies
         id: depends
@@ -387,8 +387,8 @@ jobs:
           cd build
           ctest -L main --verbose
 
-  ubuntu-24-cmake-vulkan-deb:
-    runs-on: ubuntu-24.04
+  ubuntu-22-cmake-vulkan:
+    runs-on: ubuntu-22.04
 
     steps:
       - name: Clone
@@ -398,72 +398,20 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-24-cmake-vulkan-deb
+          key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
 
       - name: Dependencies
         id: depends
         run: |
-          sudo apt-get install -y glslc libvulkan-dev libcurl4-openssl-dev
-
-      - name: Configure
-        id: cmake_configure
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
-            -DGGML_BACKEND_DL=ON \
-            -DGGML_CPU_ALL_VARIANTS=ON \
-            -DGGML_VULKAN=ON
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake --build build -j $(nproc)
-
-  ubuntu-24-cmake-vulkan:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: ubuntu-24-cmake-vulkan
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo add-apt-repository -y ppa:kisak/kisak-mesa
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Use Vulkan SDK Cache
-        uses: actions/cache@v4
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
         run: |
-          source ./vulkan_sdk/setup-env.sh
           cmake -B build \
             -DGGML_VULKAN=ON
           cmake --build build --config Release -j $(nproc)
@@ -473,12 +421,11 @@ jobs:
         run: |
           cd build
           export GGML_VK_VISIBLE_DEVICES=0
-          export GGML_VK_DISABLE_F16=1
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 4200
 
-  ubuntu-24-cmake-webgpu:
-    runs-on: ubuntu-24.04
+  ubuntu-22-cmake-webgpu:
+    runs-on: ubuntu-22.04
 
     steps:
       - name: Clone
@@ -488,34 +435,16 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-24-cmake-webgpu
+          key: ubuntu-22-cmake-webgpu
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
+      - name: Vulkan SDK Dependencies
+        id: vulkan-depends
         run: |
-          sudo add-apt-repository -y ppa:kisak/kisak-mesa
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Use Vulkan SDK Cache
-        uses: actions/cache@v4
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
 
       - name: Dawn Dependency
         id: dawn-depends
@@ -558,7 +487,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev rocwmma-dev
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1130,7 +1059,7 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
@@ -1161,7 +1090,6 @@ jobs:
     env:
       # The ROCm version must correspond to the version used in the HIP SDK.
       ROCM_VERSION: "6.4.2"
-      # Make sure this is in sync with build-cache.yml
       HIPSDK_INSTALLER_VERSION: "25.Q3"
 
     steps:
@@ -1169,25 +1097,38 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      - name: Grab rocWMMA package
-        id: grab_rocwmma
+      - name: Clone rocWMMA repository
+        id: clone_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/${{ env.ROCM_VERSION }}/pool/main/r/rocwmma-dev/rocwmma-dev_1.7.0.60402-120~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
+          git clone https://github.com/rocm/rocwmma --branch rocm-${{ env.ROCM_VERSION }} --depth 1
 
-      - name: Use ROCm Installation Cache
-        uses: actions/cache@v4
+      - name: Cache ROCm Installation
         id: cache-rocm
+        uses: actions/cache@v4
         with:
           path: C:\Program Files\AMD\ROCm
           key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
-      - name: Setup ROCm
+      - name: Install ROCm
         if: steps.cache-rocm.outputs.cache-hit != 'true'
-        uses: ./.github/actions/windows-setup-rocm
-        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
+          write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
         id: verify
@@ -1220,9 +1161,8 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
             -DCMAKE_BUILD_TYPE=Release `
-            -DROCM_DIR="${env:HIP_PATH}" `
             -DGGML_HIP=ON `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_RPC=ON `
@@ -1282,12 +1222,11 @@ jobs:
       - name: Clone
         uses: actions/checkout@v4
 
-      # Disabled due to size (400MB) and always 0 cache hits
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.16
-      #   with:
-      #     key: android-build
-      #     evict-old-files: 1d
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: android-build
+          evict-old-files: 1d
 
       - name: Set up JDK
         uses: actions/setup-java@v3
@@ -1305,81 +1244,6 @@ jobs:
           cd examples/llama.android
           ./gradlew build --no-daemon
 
-  android-ndk-build:
-    runs-on: ubuntu-latest
-
-    env:
-      OPENCL_VERSION: 2025.07.22
-
-    strategy:
-      matrix:
-        include:
-          - build: 'arm64-cpu'
-            defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_CURL=OFF -D GGML_OPENMP=OFF'
-          - build: 'arm64-snapdragon'
-            defines: '--preset arm64-android-snapdragon-release'
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Install OpenCL Headers and Libs
-        id: install_opencl
-        if: ${{ matrix.build == 'arm64-snapdragon' }}
-        run: |
-          mkdir opencl
-          curl -L -o opencl/clhpp.tar.gz      https://github.com/KhronosGroup/OpenCL-CLHPP/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
-          curl -L -o opencl/headers.tar.gz    https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
-          curl -L -o opencl/icd-loader.tar.gz https://github.com/KhronosGroup/OpenCL-ICD-Loader/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
-          tar -xaf opencl/headers.tar.gz    -C opencl
-          tar -xaf opencl/clhpp.tar.gz      -C opencl
-          tar -xaf opencl/icd-loader.tar.gz -C opencl
-          sudo cp -r opencl/OpenCL-Headers-${OPENCL_VERSION}/CL         ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-          sudo cp -r opencl/OpenCL-CLHPP-${OPENCL_VERSION}/include/CL/* ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/CL
-          cd opencl/OpenCL-ICD-Loader-${OPENCL_VERSION}
-          cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -DOPENCL_ICD_LOADER_HEADERS_DIR=${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=31 -DANDROID_STL=c++_shared
-          cmake --build build
-          sudo cp build/libOpenCL.so ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-          rm -rf opencl
-
-      - name: Install Hexagon SDK
-        id: install_hexsdk
-        if: ${{ matrix.build == 'arm64-snapdragon' }}
-        env:
-          HEXSDK_VER: 6.4.0.2
-          HEXTLS_VER: 19.0.04
-        run: |
-          curl -L -o hex-sdk.tar.gz https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v$HEXSDK_VER/hexagon-sdk-v$HEXSDK_VER-amd64-lnx.tar.xz
-          mkdir hex-sdk
-          tar -xaf hex-sdk.tar.gz -C hex-sdk
-          ls -l hex-sdk
-          sudo mv hex-sdk /opt/hexagon
-          echo "HEXAGON_SDK_ROOT=/opt/hexagon/$HEXSDK_VER"                                     >> "$GITHUB_ENV"
-          echo "HEXAGON_TOOLS_ROOT=/opt/hexagon/$HEXSDK_VER/tools/HEXAGON_Tools/$HEXTLS_VER"   >> "$GITHUB_ENV"
-          echo "DEFAULT_HLOS_ARCH=64"                                                          >> "$GITHUB_ENV"
-          echo "DEFAULT_TOOLS_VARIANT=toolv19"                                                 >> "$GITHUB_ENV"
-          echo "DEFAULT_NO_QURT_INC=0"                                                         >> "$GITHUB_ENV"
-          echo "DEFAULT_DSP_ARCH=v73"                                                          >> "$GITHUB_ENV"
-
-      - name: Update CMake presets
-        id: update_presets
-        if: ${{ matrix.build == 'arm64-snapdragon' }}
-        run: |
-          cp docs/backend/hexagon/CMakeUserPresets.json .
-
-      - name: Build
-        id: ndk_build
-        run: |
-          cmake ${{ matrix.defines }} -B build
-          cmake --build build
-          cmake --install build --prefix pkg-adb/llama.cpp
-
-      - name: Test
-        id: cmake_test
-        run: |
-          echo "FIXME: test on devices"
-
   openEuler-latest-cmake-cann:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
@@ -1597,6 +1461,34 @@ jobs:
         run: |
           bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
+#  ggml-ci-x64-amd-vulkan:
+#    runs-on: [self-hosted, Linux, X64, AMD]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#
+#      - name: Test
+#        id: ggml-ci
+#        run: |
+#          vulkaninfo --summary
+#          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+#
+#  ggml-ci-x64-amd-rocm:
+#    runs-on: [self-hosted, Linux, X64, AMD]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#
+#      - name: Test
+#        id: ggml-ci
+#        run: |
+#          amd-smi static
+#          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
   ggml-ci-mac-metal:
     runs-on: [self-hosted, macOS, ARM64]
 
@@ -1623,29 +1515,3 @@ jobs:
         run: |
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  ggml-ci-arm64-cpu-kleidiai:
-     runs-on: ubuntu-22.04-arm
-
-     steps:
-       - name: Clone
-         id: checkout
-         uses: actions/checkout@v4
-
-       - name: ccache
-         uses: ggml-org/ccache-action@v1.2.16
-         with:
-           key: ggml-ci-arm64-cpu-kleidiai
-           evict-old-files: 1d
-
-       - name: Dependencies
-         id: depends
-         run: |
-           sudo apt-get update
-           sudo apt-get install -y build-essential libcurl4-openssl-dev
-
-       - name: Test
-         id: ggml-ci
-         run: |
-           GG_BUILD_KLEIDIAI=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-

.github/workflows/docker.yml

@@ -40,7 +40,7 @@ jobs:
           # https://github.com/ggml-org/llama.cpp/issues/11888
           #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
+          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
           - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
@@ -89,15 +89,12 @@ jobs:
               TYPE="-${{ matrix.config.tag }}"
           fi
           PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
-          CACHETAGS="${PREFIX}buildcache${TYPE}"
           FULLTAGS="${PREFIX}full${TYPE},${PREFIX}full${TYPE}-${{ steps.srctag.outputs.name }}"
           LIGHTTAGS="${PREFIX}light${TYPE},${PREFIX}light${TYPE}-${{ steps.srctag.outputs.name }}"
           SERVERTAGS="${PREFIX}server${TYPE},${PREFIX}server${TYPE}-${{ steps.srctag.outputs.name }}"
-          echo "cache_output_tags=$CACHETAGS" >> $GITHUB_OUTPUT
           echo "full_output_tags=$FULLTAGS" >> $GITHUB_OUTPUT
           echo "light_output_tags=$LIGHTTAGS" >> $GITHUB_OUTPUT
           echo "server_output_tags=$SERVERTAGS" >> $GITHUB_OUTPUT
-          echo "cache_output_tags=$CACHETAGS"  # print out for debugging
           echo "full_output_tags=$FULLTAGS"  # print out for debugging
           echo "light_output_tags=$LIGHTTAGS"  # print out for debugging
           echo "server_output_tags=$SERVERTAGS"  # print out for debugging
@@ -134,14 +131,11 @@ jobs:
           target: full
           provenance: false
           # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
 
       - name: Build and push Light Docker image (tagged + versioned)
         if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
@@ -156,14 +150,11 @@ jobs:
           target: light
           provenance: false
           # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
 
       - name: Build and push Server Docker image (tagged + versioned)
         if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
@@ -178,14 +169,11 @@ jobs:
           target: server
           provenance: false
           # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
 
   create_tag:
     name: Create and push git tag

.github/workflows/release.yml

@@ -75,7 +75,7 @@ jobs:
           name: llama-bin-macos-arm64.zip
 
   macOS-x64:
-    runs-on: macos-15-intel
+    runs-on: macos-13
 
     steps:
       - name: Clone
@@ -134,8 +134,6 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 's390x-z15' # z15 because our CI runners are on z15
-            os: ubuntu-22.04-s390x
           # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
           # - build: 'arm64'
           #   os: ubuntu-22.04-arm
@@ -152,7 +150,7 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-cpu-cmake-${{ matrix.build }}
+          key: ubuntu-cpu-cmake
           evict-old-files: 1d
 
       - name: Dependencies
@@ -464,7 +462,7 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
 
@@ -507,7 +505,6 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
@@ -516,15 +513,10 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
 
-          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/tcm.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/libhwloc-15.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/umf/latest/bin/umf.dll" ./build/bin
-
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-bin-win-sycl-x64.zip ./build/bin/*
 
@@ -551,12 +543,10 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      - name: Grab rocWMMA package
-        id: grab_rocwmma
+      - name: Clone rocWMMA repository
+        id: clone_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.0.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.0.0.70001-42~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
+          git clone https://github.com/rocm/rocwmma --branch develop --depth 1
 
       - name: Cache ROCm Installation
         id: cache-rocm
@@ -611,7 +601,7 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.0.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `

.github/workflows/update-ops-docs.yml

@@ -3,12 +3,10 @@ name: Update Operations Documentation
 on:
     push:
         paths:
-            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
     pull_request:
         paths:
-            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
 

CODEOWNERS

@@ -2,7 +2,7 @@
 # multiplie collaborators per item can be specified
 
 /.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @slaren @CISC
+/.github/actions/                       @slaren
 /.github/workflows/                     @CISC
 /.github/workflows/release.yml          @slaren
 /.github/workflows/winget.yml           @slaren
@@ -14,7 +14,6 @@
 /common/build-info.*                    @ggerganov
 /common/common.*                        @ggerganov
 /common/console.*                       @ggerganov
-/common/http.*                          @angt
 /common/llguidance.*                    @ggerganov
 /common/log.*                           @ggerganov
 /common/sampling.*                      @ggerganov
@@ -51,27 +50,20 @@
 /ggml/src/ggml-blas/                    @slaren
 /ggml/src/ggml-common.h                 @ggerganov @slaren
 /ggml/src/ggml-cpu/                     @ggerganov @slaren
-/ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/common.cuh          @slaren
 /ggml/src/ggml-cuda/fattn*              @JohannesGaessler
 /ggml/src/ggml-cuda/ggml-cuda.cu        @slaren
-/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
+/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler
 /ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
 /ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
 /ggml/src/ggml-cuda/mmvq.*              @JohannesGaessler
-/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
-/ggml/src/ggml-hip/                     @IMbackK
-/ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
 /ggml/src/ggml-impl.h                   @ggerganov @slaren
 /ggml/src/ggml-metal/                   @ggerganov
-/ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
-/ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov
 /ggml/src/ggml-threading.*              @ggerganov @slaren
 /ggml/src/ggml-vulkan/                  @0cc4m
-/ggml/src/ggml-webgpu/                  @reeselevine
 /ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
 /ggml/src/ggml.c                        @ggerganov @slaren
 /ggml/src/ggml.cpp                      @ggerganov @slaren

README.md

@@ -84,7 +84,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
-- [x] [Jamba](https://huggingface.co/ai21labs)
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
@@ -139,7 +138,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
-- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 
 #### Multimodal
 
@@ -189,7 +187,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
 - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 - Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
-- Go (no CGo needed): [hybridgroup/yzma](https://github.com/hybridgroup/yzma)
 
 </details>
 
@@ -281,7 +278,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
-| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
 
 ## Obtaining and quantizing models
 

ci/run.sh

@@ -22,9 +22,6 @@
 # # with MUSA support
 # GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with KLEIDIAI support
-# GG_BUILD_KLEIDIAI=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -37,9 +34,9 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -f $OUT/*.log
-rm -f $OUT/*.exit
-rm -f $OUT/*.md
+rm -f "$OUT/*.log"
+rm -f "$OUT/*.exit"
+rm -f "$OUT/*.md"
 
 sd=`dirname $0`
 cd $sd/../
@@ -75,7 +72,7 @@ if [ ! -z ${GG_BUILD_ROCM} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DAMDGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -117,35 +114,6 @@ if [ ! -z ${GG_BUILD_NO_SVE} ]; then
     # arm 9 and newer enables sve by default, adjust these flags depending on the cpu used
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm"
 fi
-
-if [ -n "${GG_BUILD_KLEIDIAI}" ]; then
-    echo ">>===== Enabling KleidiAI support"
-
-    CANDIDATES=("armv9-a+dotprod+i8mm" "armv8.6-a+dotprod+i8mm" "armv8.2-a+dotprod")
-    CPU=""
-
-    for cpu in "${CANDIDATES[@]}"; do
-        if echo 'int main(){}' | ${CXX:-c++} -march="$cpu" -x c++ - -c -o /dev/null >/dev/null 2>&1; then
-            CPU="$cpu"
-            break
-        fi
-    done
-
-    if [ -z "$CPU" ]; then
-        echo "ERROR: None of the required ARM baselines (armv9/armv8.6/armv8.2 + dotprod) are supported by this compiler."
-        exit 1
-    fi
-
-    echo ">>===== Using ARM baseline: ${CPU}"
-
-    CMAKE_EXTRA="${CMAKE_EXTRA:+$CMAKE_EXTRA } \
-        -DGGML_NATIVE=OFF \
-        -DGGML_CPU_KLEIDIAI=ON \
-        -DGGML_CPU_AARCH64=ON \
-        -DGGML_CPU_ARM_ARCH=${CPU} \
-        -DBUILD_SHARED_LIBS=OFF"
-fi
-
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -543,7 +511,12 @@ function gg_run_rerank_tiny {
     gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
     gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
     gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+
+    gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
 
     path_models="../models-mnt/rerank-tiny"
 
@@ -633,7 +606,6 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
@@ -651,6 +623,4 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run ctest_with_model_release
 fi
 
-cat $OUT/README.md
-
 exit $ret

cmake/riscv64-spacemit-linux-gnu-gcc.cmake

@@ -1,29 +0,0 @@
-set(CMAKE_SYSTEM_NAME Linux)
-set(CMAKE_SYSTEM_PROCESSOR riscv64)
-set(CMAKE_SYSTEM_VERSION 1)
-
-if (CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "^(riscv)")
-    message(STATUS "HOST SYSTEM ${CMAKE_HOST_SYSTEM_PROCESSOR}")
-else()
-    set(GNU_MACHINE riscv64-unknown-linux-gnu CACHE STRING "GNU compiler triple")
-    if (DEFINED ENV{RISCV_ROOT_PATH})
-        file(TO_CMAKE_PATH $ENV{RISCV_ROOT_PATH} RISCV_ROOT_PATH)
-    else()
-        message(FATAL_ERROR "RISCV_ROOT_PATH env must be defined")
-    endif()
-
-    set(RISCV_ROOT_PATH ${RISCV_ROOT_PATH} CACHE STRING "root path to riscv toolchain")
-    set(CMAKE_C_COMPILER ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-gcc)
-    set(CMAKE_CXX_COMPILER ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-g++)
-    set(CMAKE_STRIP ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-strip)
-    set(CMAKE_FIND_ROOT_PATH "${RISCV_ROOT_PATH}/riscv64-unknown-linux-gnu")
-    set(CMAKE_SYSROOT "${RISCV_ROOT_PATH}/sysroot")
-endif()
-
-set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
-set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
-set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
-set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
-set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CMAKE_C_FLAGS}")
-set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CXX_FLAGS}")
-set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -latomic")

common/CMakeLists.txt

@@ -56,7 +56,6 @@ add_library(${TARGET} STATIC
     common.h
     console.cpp
     console.h
-    http.h
     json-partial.cpp
     json-partial.h
     json-schema-to-grammar.cpp

common/arg.cpp

@@ -32,11 +32,13 @@
 #include <thread>
 #include <vector>
 
+//#define LLAMA_USE_CURL
+
 #if defined(LLAMA_USE_CURL)
 #include <curl/curl.h>
 #include <curl/easy.h>
 #else
-#include "http.h"
+#include <cpp-httplib/httplib.h>
 #endif
 
 #ifdef __linux__
@@ -52,13 +54,6 @@
 #endif
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
-// isatty
-#if defined(_WIN32)
-#include <io.h>
-#else
-#include <unistd.h>
-#endif
-
 using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
@@ -105,14 +100,6 @@ static void write_file(const std::string & fname, const std::string & content) {
     }
 }
 
-static bool is_output_a_tty() {
-#if defined(_WIN32)
-    return _isatty(_fileno(stdout));
-#else
-    return isatty(1);
-#endif
-}
-
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
     this->examples = std::move(examples);
     return *this;
@@ -230,55 +217,12 @@ struct common_hf_file_res {
     std::string mmprojFile;
 };
 
-static void write_etag(const std::string & path, const std::string & etag) {
-    const std::string etag_path = path + ".etag";
-    write_file(etag_path, etag);
-    LOG_DBG("%s: file etag saved: %s\n", __func__, etag_path.c_str());
-}
-
-static std::string read_etag(const std::string & path) {
-    std::string none;
-    const std::string etag_path = path + ".etag";
-
-    if (std::filesystem::exists(etag_path)) {
-        std::ifstream etag_in(etag_path);
-        if (!etag_in) {
-            LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
-            return none;
-        }
-        std::string etag;
-        std::getline(etag_in, etag);
-        return etag;
-    }
-
-    // no etag file, but maybe there is an old .json
-    // remove this code later
-    const std::string metadata_path = path + ".json";
-
-    if (std::filesystem::exists(metadata_path)) {
-        std::ifstream metadata_in(metadata_path);
-        try {
-            nlohmann::json metadata_json;
-            metadata_in >> metadata_json;
-            LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(),
-                    metadata_json.dump().c_str());
-            if (metadata_json.contains("etag") && metadata_json.at("etag").is_string()) {
-                std::string etag = metadata_json.at("etag");
-                write_etag(path, etag);
-                if (!std::filesystem::remove(metadata_path)) {
-                    LOG_WRN("%s: failed to delete old .json metadata file: %s\n", __func__, metadata_path.c_str());
-                }
-                return etag;
-            }
-        } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-        }
-    }
-    return none;
-}
-
 #ifdef LLAMA_USE_CURL
 
+bool common_has_curl() {
+    return true;
+}
+
 //
 // CURL utils
 //
@@ -429,15 +373,36 @@ static bool common_download_head(CURL *              curl,
 static bool common_download_file_single_online(const std::string & url,
                                                const std::string & path,
                                                const std::string & bearer_token) {
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
     for (int i = 0; i < max_attempts; ++i) {
-        std::string etag;
+        nlohmann::json metadata;  // TODO @ngxson : get rid of this json, use regex instead
+        std::string    etag;
+        std::string    last_modified;
 
         // Check if the file already exists locally
         const auto file_exists = std::filesystem::exists(path);
         if (file_exists) {
-            etag = read_etag(path);
+            // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
+            std::ifstream metadata_in(metadata_path);
+            if (metadata_in.good()) {
+                try {
+                    metadata_in >> metadata;
+                    LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(),
+                            metadata.dump().c_str());
+                    if (metadata.contains("etag") && metadata.at("etag").is_string()) {
+                        etag = metadata.at("etag");
+                    }
+                    if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
+                        last_modified = metadata.at("lastModified");
+                    }
+                } catch (const nlohmann::json::exception & e) {
+                    LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+                }
+            }
+            // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
         } else {
             LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
         }
@@ -475,6 +440,11 @@ static bool common_download_file_single_online(const std::string & url,
                         headers.etag.c_str());
                 should_download              = true;
                 should_download_from_scratch = true;
+            } else if (!last_modified.empty() && last_modified != headers.last_modified) {
+                LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__,
+                        last_modified.c_str(), headers.last_modified.c_str());
+                should_download              = true;
+                should_download_from_scratch = true;
             }
         }
 
@@ -505,9 +475,15 @@ static bool common_download_file_single_online(const std::string & url,
                     }
                 }
             }
-            if (head_request_ok) {
-                write_etag(path, headers.etag);
-            }
+
+            // Write the updated JSON metadata file.
+            metadata.update({
+                { "url",          url                   },
+                { "etag",         headers.etag          },
+                { "lastModified", headers.last_modified }
+            });
+            write_file(metadata_path, metadata.dump(4));
+            LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
 
             // start the download
             LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n",
@@ -594,11 +570,82 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 
 #else
 
-static void print_progress(size_t current, size_t total) {
-    if (!is_output_a_tty()) {
-        return;
+bool common_has_curl() {
+    return false;
+}
+
+struct common_url {
+    std::string scheme;
+    std::string user;
+    std::string password;
+    std::string host;
+    std::string path;
+};
+
+static common_url parse_url(const std::string & url) {
+    common_url parts;
+    auto scheme_end = url.find("://");
+
+    if (scheme_end == std::string::npos) {
+        throw std::runtime_error("invalid URL: no scheme");
+    }
+    parts.scheme = url.substr(0, scheme_end);
+
+    if (parts.scheme != "http" && parts.scheme != "https") {
+        throw std::runtime_error("unsupported URL scheme: " + parts.scheme);
     }
 
+    auto rest = url.substr(scheme_end + 3);
+    auto at_pos = rest.find('@');
+
+    if (at_pos != std::string::npos) {
+        auto auth = rest.substr(0, at_pos);
+        auto colon_pos = auth.find(':');
+        if (colon_pos != std::string::npos) {
+            parts.user = auth.substr(0, colon_pos);
+            parts.password = auth.substr(colon_pos + 1);
+        } else {
+            parts.user = auth;
+        }
+        rest = rest.substr(at_pos + 1);
+    }
+
+    auto slash_pos = rest.find('/');
+
+    if (slash_pos != std::string::npos) {
+        parts.host = rest.substr(0, slash_pos);
+        parts.path = rest.substr(slash_pos);
+    } else {
+        parts.host = rest;
+        parts.path = "/";
+    }
+    return parts;
+}
+
+static std::pair<httplib::Client, common_url> http_client(const std::string & url) {
+    common_url parts = parse_url(url);
+
+    if (parts.host.empty()) {
+        throw std::runtime_error("error: invalid URL format");
+    }
+
+    if (!parts.user.empty()) {
+        throw std::runtime_error("error: user:password@ not supported yet"); // TODO
+    }
+
+    httplib::Client cli(parts.scheme + "://" + parts.host);
+    cli.set_follow_location(true);
+
+    // TODO cert
+
+    return { std::move(cli), std::move(parts) };
+}
+
+static std::string show_masked_url(const common_url & parts) {
+    return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path;
+}
+
+static void print_progress(size_t current, size_t total) { // TODO isatty
     if (!total) {
         return;
     }
@@ -617,6 +664,51 @@ static void print_progress(size_t current, size_t total) {
     std::cout.flush();
 }
 
+struct common_file_metadata {
+    std::string etag;
+    std::string last_modified;
+};
+
+static std::optional<common_file_metadata> read_metadata(const std::string & path) {
+    if (!std::filesystem::exists(path)) {
+        return std::nullopt;
+    }
+
+    nlohmann::json metadata_json;
+    common_file_metadata metadata;
+
+    std::ifstream metadata_in(path);
+    try {
+        metadata_in >> metadata_json;
+        LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, path.c_str(),
+                metadata_json.dump().c_str());
+        if (metadata_json.contains("etag") && metadata_json.at("etag").is_string()) {
+            metadata.etag = metadata_json.at("etag");
+        }
+        if (metadata_json.contains("lastModified") && metadata_json.at("lastModified").is_string()) {
+            metadata.last_modified = metadata_json.at("lastModified");
+        }
+    } catch (const nlohmann::json::exception & e) {
+        LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, path.c_str(), e.what());
+        return std::nullopt;
+    }
+
+    return metadata;
+}
+
+static void write_metadata(const std::string & path,
+                           const std::string & url,
+                           const common_file_metadata & metadata) {
+    nlohmann::json metadata_json = {
+        { "url",          url                    },
+        { "etag",         metadata.etag          },
+        { "lastModified", metadata.last_modified }
+    };
+
+    write_file(path, metadata_json.dump(4));
+    LOG_DBG("%s: file metadata saved: %s\n", __func__, path.c_str());
+}
+
 static bool common_pull_file(httplib::Client & cli,
                              const std::string & resolve_path,
                              const std::string & path_tmp,
@@ -683,10 +775,12 @@ static bool common_pull_file(httplib::Client & cli,
 static bool common_download_file_single_online(const std::string & url,
                                                const std::string & path,
                                                const std::string & bearer_token) {
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
 
-    auto [cli, parts] = common_http_client(url);
+    auto [cli, parts] = http_client(url);
 
     httplib::Headers default_headers = {{"User-Agent", "llama-cpp"}};
     if (!bearer_token.empty()) {
@@ -694,11 +788,12 @@ static bool common_download_file_single_online(const std::string & url,
     }
     cli.set_default_headers(default_headers);
 
+    common_file_metadata last;
     const bool file_exists = std::filesystem::exists(path);
-
-    std::string last_etag;
     if (file_exists) {
-        last_etag = read_etag(path);
+        if (auto opt = read_metadata(metadata_path)) {
+            last = *opt;
+        }
     } else {
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
@@ -714,9 +809,14 @@ static bool common_download_file_single_online(const std::string & url,
             }
         }
 
-        std::string etag;
-        if (head_ok && head->has_header("ETag")) {
-            etag = head->get_header_value("ETag");
+        common_file_metadata current;
+        if (head_ok) {
+            if (head->has_header("ETag")) {
+                current.etag = head->get_header_value("ETag");
+            }
+            if (head->has_header("Last-Modified")) {
+                current.last_modified = head->get_header_value("Last-Modified");
+            }
         }
 
         size_t total_size = 0;
@@ -734,10 +834,16 @@ static bool common_download_file_single_online(const std::string & url,
         }
 
         bool should_download_from_scratch = false;
-        if (!last_etag.empty() && !etag.empty() && last_etag != etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__,
-                    last_etag.c_str(), etag.c_str());
-            should_download_from_scratch = true;
+        if (head_ok) {
+            if (!last.etag.empty() && last.etag != current.etag) {
+                LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__,
+                        last.etag.c_str(), current.etag.c_str());
+                should_download_from_scratch = true;
+            } else if (!last.last_modified.empty() && last.last_modified != current.last_modified) {
+                LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__,
+                        last.last_modified.c_str(), current.last_modified.c_str());
+                should_download_from_scratch = true;
+            }
         }
 
         if (file_exists) {
@@ -765,8 +871,9 @@ static bool common_download_file_single_online(const std::string & url,
         }
 
         // start the download
-        LOG_INF("%s: trying to download model from %s to %s (etag:%s)...\n",
-                __func__, common_http_show_masked_url(parts).c_str(), path_temporary.c_str(), etag.c_str());
+        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n",
+                __func__, show_masked_url(parts).c_str(), path_temporary.c_str(),
+                current.etag.c_str(), current.last_modified.c_str());
         const bool was_pull_successful = common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size);
         if (!was_pull_successful) {
             if (i + 1 < max_attempts) {
@@ -776,6 +883,7 @@ static bool common_download_file_single_online(const std::string & url,
             } else {
                 LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
             }
+
             continue;
         }
 
@@ -783,9 +891,7 @@ static bool common_download_file_single_online(const std::string & url,
             LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
             return false;
         }
-        if (!etag.empty()) {
-            write_etag(path, etag);
-        }
+        write_metadata(metadata_path, url, current);
         break;
     }
 
@@ -794,7 +900,7 @@ static bool common_download_file_single_online(const std::string & url,
 
 std::pair<long, std::vector<char>> common_remote_get_content(const std::string          & url,
                                                              const common_remote_params & params) {
-    auto [cli, parts] = common_http_client(url);
+    auto [cli, parts] = http_client(url);
 
     httplib::Headers headers = {{"User-Agent", "llama-cpp"}};
     for (const auto & header : params.headers) {
@@ -1615,14 +1721,18 @@ static void add_rpc_devices(const std::string & servers) {
     if (!rpc_reg) {
         throw std::invalid_argument("failed to find RPC backend");
     }
-    typedef ggml_backend_reg_t (*ggml_backend_rpc_add_server_t)(const char * endpoint);
-    ggml_backend_rpc_add_server_t ggml_backend_rpc_add_server_fn = (ggml_backend_rpc_add_server_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_server");
-    if (!ggml_backend_rpc_add_server_fn) {
-        throw std::invalid_argument("failed to find RPC add server function");
+    typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
+    ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
+    if (!ggml_backend_rpc_add_device_fn) {
+        throw std::invalid_argument("failed to find RPC device add function");
     }
     for (const auto & server : rpc_servers) {
-        auto reg = ggml_backend_rpc_add_server_fn(server.c_str());
-        ggml_backend_register(reg);
+        ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
+        if (dev) {
+            ggml_backend_device_register(dev);
+        } else {
+            throw std::invalid_argument("failed to register RPC device");
+        }
     }
 }
 
@@ -1760,7 +1870,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
     add_opt(common_arg(
         {"-t", "--threads"}, "N",
-        string_format("number of CPU threads to use during generation (default: %d)", params.cpuparams.n_threads),
+        string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads),
         [](common_params & params, int value) {
             params.cpuparams.n_threads = value;
             if (params.cpuparams.n_threads <= 0) {
@@ -1928,21 +2038,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
-        {"--ctx-checkpoints", "--swa-checkpoints"}, "N",
-        string_format("max number of context checkpoints to create per slot (default: %d)\n"
-            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_ctx_checkpoints),
+        {"--swa-checkpoints"}, "N",
+        string_format("max number of SWA checkpoints per slot to create (default: %d)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_swa_checkpoints),
         [](common_params & params, int value) {
-            params.n_ctx_checkpoints = value;
+            params.n_swa_checkpoints = value;
         }
-    ).set_env("LLAMA_ARG_CTX_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--cache-ram", "-cram"}, "N",
-        string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)\n"
-            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)", params.cache_ram_mib),
-        [](common_params & params, int value) {
-            params.cache_ram_mib = value;
-        }
-    ).set_env("LLAMA_ARG_CACHE_RAM").set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_env("LLAMA_ARG_SWA_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--kv-unified", "-kvu"},
         string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
@@ -2592,13 +2694,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.no_extra_bufts = true;
         }
     ).set_env("LLAMA_ARG_NO_REPACK"));
-    add_opt(common_arg(
-        {"--no-host"},
-        "bypass host buffer allowing extra buffers to be used",
-        [](common_params & params) {
-            params.no_host = true;
-        }
-    ).set_env("LLAMA_ARG_NO_HOST"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
         string_format(
@@ -3203,7 +3298,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--parse-special"},
-        string_format("parse special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
+        string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
         [](common_params & params) {
             params.parse_special = true;
         }
@@ -3248,7 +3343,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--embd-output-format"}, "FORMAT",
-        "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix, \"raw\" = plain whitespace-delimited output (one embedding per line)",
+        "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix",
         [](common_params & params, const std::string & value) {
             params.embd_out = value;
         }
@@ -3358,7 +3453,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     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) {
+        [](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();
@@ -3435,13 +3530,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.use_jinja = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_JINJA"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
         "- none: leaves thoughts unparsed in `message.content`\n"
-        "- deepseek: puts thoughts in `message.reasoning_content`\n"
-        "- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`\n"
+        "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
         "(default: auto)",
         [](common_params & params, const std::string & value) {
             params.reasoning_format = common_reasoning_format_from_name(value);
@@ -3570,23 +3664,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_file(common_log_main(), value.c_str());
         }
     ));
-    add_opt(common_arg(
-        {"--log-colors"}, "[on|off|auto]",
-        "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
-        "'auto' enables colors when output is to a terminal",
-        [](common_params &, const std::string & value) {
-            if (is_truthy(value)) {
-                common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
-            } else if (is_falsey(value)) {
-                common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
-            } else if (is_autoy(value)) {
-                common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
-            } else {
-                throw std::invalid_argument(
-                    string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
-            }
-        }
-    ).set_env("LLAMA_LOG_COLORS"));
+    add_opt(common_arg({ "--log-colors" }, "[on|off|auto]",
+                       "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
+                       "'auto' enables colors when output is to a terminal",
+                       [](common_params &, const std::string & value) {
+                           if (is_truthy(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
+                           } else if (is_falsey(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
+                           } else if (is_autoy(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
+                           } else {
+                               throw std::invalid_argument(
+                                   string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
+                           }
+                       }).set_env("LLAMA_LOG_COLORS"));
     add_opt(common_arg(
         {"-v", "--verbose", "--log-verbose"},
         "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",
@@ -3852,87 +3944,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
-    add_opt(common_arg(
-        {"--diffusion-steps"}, "N",
-        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
-        [](common_params & params, int value) { params.diffusion.steps = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-visual"},
-        string_format("enable visual diffusion mode (show progressive generation) (default: %s)", params.diffusion.visual_mode ? "true" : "false"),
-        [](common_params & params) { params.diffusion.visual_mode = true; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-eps"}, "F",
-        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
-        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-algorithm"}, "N",
-        string_format("diffusion algorithm: 0=ORIGIN, 1=ENTROPY_BASED, 2=MARGIN_BASED, 3=RANDOM, 4=LOW_CONFIDENCE (default: %d)", params.diffusion.algorithm),
-        [](common_params & params, int value) { params.diffusion.algorithm = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-alg-temp"}, "F",
-        string_format("dream algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
-        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-block-length"}, "N",
-        string_format("llada block length for generation (default: %d)", params.diffusion.block_length),
-        [](common_params & params, int value) { params.diffusion.block_length = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-cfg-scale"}, "F",
-        string_format("llada classifier-free guidance scale (default: %.3f)", (double) params.diffusion.cfg_scale),
-        [](common_params & params, const std::string & value) { params.diffusion.cfg_scale = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        {"--diffusion-add-gumbel-noise"}, "F",
-        string_format("add gumbel noise to the logits if temp > 0.0 (default: %s)", params.diffusion.add_gumbel_noise ? "true" : "false"),
-        [](common_params & params, const std::string & value) { params.diffusion.add_gumbel_noise = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "-lr", "--learning-rate" }, "ALPHA",
-        string_format("adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)", (double) params.lr.lr0),
-        [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
-        string_format("(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
-            (double) params.lr.lr_min),
-        [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg(
-        {"-decay-epochs", "--learning-rate-decay-epochs"}, "ALPHA",
-        string_format("(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)", (double) params.lr.decay_epochs),
-        [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg(
-        {"-wd", "--weight-decay"}, "WD",
-        string_format("adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).", (double) params.lr.wd),
-        [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg(
-        {"-val-split", "--val-split"}, "FRACTION",
-        string_format("fraction of data to use as validation set for training (default: %.2g).", (double) params.val_split),
-        [](common_params & params, const std::string & value) { params.val_split = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg(
-        {"-epochs", "--epochs"}, "N",
-        string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
-        [](common_params & params, int epochs) { params.lr.epochs = epochs; }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg(
-        {"-opt", "--optimizer"}, "sgd|adamw", "adamw or sgd",
-        [](common_params & params, const std::string & name) {
-            params.optimizer = common_opt_get_optimizer(name.c_str());
-            if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
-                throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
-            }
-        }
-    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-
-    // presets
+    // model-specific
     add_opt(common_arg(
         {"--tts-oute-default"},
         string_format("use default OuteTTS models (note: can download weights from the internet)"),
@@ -3945,16 +3957,42 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
     add_opt(common_arg(
-        {"--embd-gemma-default"},
-        string_format("use default EmbeddingGemma model (note: can download weights from the internet)"),
+        {"--embd-bge-small-en-default"},
+        string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.model.hf_repo = "ggml-org/embeddinggemma-300M-qat-q4_0-GGUF";
-            params.model.hf_file = "embeddinggemma-300M-qat-Q4_0.gguf";
-            params.port = 8011;
-            params.n_ubatch = 2048;
-            params.n_batch = 2048;
-            params.n_parallel = 32;
-            params.n_ctx = 2048*params.n_parallel;
+            params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
+            params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf";
+            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+            params.embd_normalize = 2;
+            params.n_ctx = 512;
+            params.verbose_prompt = true;
+            params.embedding = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--embd-e5-small-en-default"},
+        string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
+            params.model.hf_file = "e5-small-v2-q8_0.gguf";
+            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+            params.embd_normalize = 2;
+            params.n_ctx = 512;
+            params.verbose_prompt = true;
+            params.embedding = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--embd-gte-small-default"},
+        string_format("use default gte-small model (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
+            params.model.hf_file = "gte-small-q8_0.gguf";
+            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+            params.embd_normalize = 2;
+            params.n_ctx = 512;
             params.verbose_prompt = true;
             params.embedding = true;
         }
@@ -4049,65 +4087,96 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
     add_opt(common_arg(
-        {"--gpt-oss-20b-default"},
-        string_format("use gpt-oss-20b (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "ggml-org/gpt-oss-20b-GGUF";
-            params.model.hf_file = "gpt-oss-20b-mxfp4.gguf";
-            params.port = 8013;
-            params.n_ubatch = 2048;
-            params.n_batch = 32768;
-            params.n_parallel = 2;
-            params.n_ctx = 131072*params.n_parallel;
-            params.sampling.temp = 1.0f;
-            params.sampling.top_p = 1.0f;
-            params.sampling.top_k = 0;
-            params.sampling.min_p = 0.01f;
-            params.use_jinja = true;
-            //params.default_template_kwargs["reasoning_effort"] = "\"high\"";
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+        { "--diffusion-steps" }, "N",
+        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
+        [](common_params & params, int value) { params.diffusion.steps = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-visual" },
+        string_format("enable visual diffusion mode (show progressive generation) (default: %s)",
+                      params.diffusion.visual_mode ? "true" : "false"),
+        [](common_params & params) { params.diffusion.visual_mode = true; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
 
     add_opt(common_arg(
-        {"--gpt-oss-120b-default"},
-        string_format("use gpt-oss-120b (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "ggml-org/gpt-oss-120b-GGUF";
-            params.port = 8013;
-            params.n_ubatch = 2048;
-            params.n_batch = 32768;
-            params.n_parallel = 2;
-            params.n_ctx = 131072*params.n_parallel;
-            params.sampling.temp = 1.0f;
-            params.sampling.top_p = 1.0f;
-            params.sampling.top_k = 0;
-            params.sampling.min_p = 0.01f;
-            params.use_jinja = true;
-            //params.default_template_kwargs["reasoning_effort"] = "\"high\"";
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+        { "--diffusion-eps" }, "F",
+        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
+        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-algorithm" }, "N",
+        string_format("diffusion algorithm: 0=ORIGIN, 1=ENTROPY_BASED, 2=MARGIN_BASED, 3=RANDOM, 4=LOW_CONFIDENCE (default: %d)",
+                      params.diffusion.algorithm),
+        [](common_params & params, int value) { params.diffusion.algorithm = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-alg-temp" }, "F",
+        string_format("dream algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
+        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
 
     add_opt(common_arg(
-        {"--vision-gemma-4b-default"},
-        string_format("use Gemma 3 4B QAT (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "ggml-org/gemma-3-4b-it-qat-GGUF";
-            params.port = 8014;
-            params.n_ctx = 0;
-            params.use_jinja = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
-
+        { "--diffusion-block-length" }, "N",
+        string_format("llada block length for generation (default: %d)", params.diffusion.block_length),
+        [](common_params & params, int value) { params.diffusion.block_length = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
     add_opt(common_arg(
-        {"--vision-gemma-12b-default"},
-        string_format("use Gemma 3 12B QAT (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "ggml-org/gemma-3-12b-it-qat-GGUF";
-            params.port = 8014;
-            params.n_ctx = 0;
-            params.use_jinja = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+        { "--diffusion-cfg-scale" }, "F",
+        string_format("llada classifier-free guidance scale (default: %.3f)", (double) params.diffusion.cfg_scale),
+        [](common_params & params, const std::string & value) { params.diffusion.cfg_scale = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-add-gumbel-noise" }, "F",
+        string_format("add gumbel noise to the logits if temp > 0.0 (default: %s)", params.diffusion.add_gumbel_noise ? "true" : "false"),
+        [](common_params & params, const std::string & value) { params.diffusion.add_gumbel_noise = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+
+
+    add_opt(
+        common_arg({ "-lr", "--learning-rate" }, "ALPHA",
+                   string_format(
+                       "adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
+                       (double) params.lr.lr0),
+                   [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(
+        common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
+                   string_format(
+                       "(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
+                       (double) params.lr.lr_min),
+                   [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(
+        common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
+                   string_format(
+                       "(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
+                       (double) params.lr.decay_epochs),
+                   [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+                { "-wd", "--weight-decay" }, "WD",
+                string_format(
+                    "adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
+                    (double) params.lr.wd),
+                [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
+                       string_format("fraction of data to use as validation set for training (default: %.2g).",
+                                     (double) params.val_split),
+                       [](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-epochs", "--epochs" }, "N",
+                       string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
+                       [](common_params & params, int epochs) { params.lr.epochs = epochs; })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
+                       [](common_params & params, const std::string & name) {
+                           params.optimizer = common_opt_get_optimizer(name.c_str());
+                           if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
+                               throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
+                           }
+                       })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
 
     return ctx_arg;
 }

common/arg.h

@@ -78,6 +78,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 
 // function to be used by test-arg-parser
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_has_curl();
 
 struct common_remote_params {
     std::vector<std::string> headers;

common/chat-parser.cpp

@@ -3,12 +3,9 @@
 #include "log.h"
 #include "regex-partial.h"
 
-#include <algorithm>
-#include <cctype>
 #include <optional>
 #include <stdexcept>
 #include <string>
-#include <string_view>
 #include <vector>
 
 using json = nlohmann::ordered_json;
@@ -78,35 +75,6 @@ bool common_chat_msg_parser::add_tool_calls(const json & arr) {
     }
     return true;
 }
-
-bool common_chat_msg_parser::add_tool_call_short_form(const json & tool_call) {
-    if (!tool_call.is_object() || tool_call.size() != 1) {
-        return false;
-    }
-
-    // Get the tool name (the single key in the object)
-    auto it = tool_call.begin();
-    std::string name = it.key();
-
-    if (name.empty()) {
-        return false;
-    }
-
-    // Get the arguments (the nested object)
-    const json & args_json = it.value();
-    std::string arguments = "";
-
-    if (args_json.is_object()) {
-        arguments = args_json.dump();
-    } else if (args_json.is_string()) {
-        arguments = args_json;
-    } else if (!args_json.is_null()) {
-        // For other types, convert to string representation
-        arguments = args_json.dump();
-    }
-
-    return add_tool_call(name, "", arguments);
-}
 void common_chat_msg_parser::finish() {
     if (!is_partial_ && pos_ != input_.size()) {
         throw std::runtime_error("Unexpected content at end of input");// + input_.substr(pos_));
@@ -169,27 +137,6 @@ void common_chat_msg_parser::consume_literal(const std::string & literal) {
 }
 
 bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think, const std::string & end_think) {
-    std::string pending_reasoning_prefix;
-
-    if (syntax_.reasoning_format == COMMON_REASONING_FORMAT_NONE) {
-        return false;
-    }
-
-    auto set_reasoning_prefix = [&](size_t prefix_pos) {
-        if (!syntax_.thinking_forced_open || syntax_.reasoning_in_content) {
-            return;
-        }
-        if (prefix_pos + start_think.size() > input_.size()) {
-            pending_reasoning_prefix.clear();
-            return;
-        }
-        // Capture the exact literal that opened the reasoning section so we can
-        // surface it back to callers. This ensures formats that force the
-        // reasoning tag open (e.g. DeepSeek R1) retain their original prefix
-        // instead of dropping it during parsing.
-        pending_reasoning_prefix = input_.substr(prefix_pos, start_think.size());
-    };
-
     auto handle_reasoning = [&](const std::string & reasoning, bool closed) {
         auto stripped_reasoning = string_strip(reasoning);
         if (stripped_reasoning.empty()) {
@@ -202,116 +149,28 @@ bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think
                 add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "</think>" : end_think);
             }
         } else {
-            if (!pending_reasoning_prefix.empty()) {
-                add_reasoning_content(pending_reasoning_prefix);
-                pending_reasoning_prefix.clear();
-            }
             add_reasoning_content(stripped_reasoning);
         }
     };
-
-    const size_t saved_pos = pos_;
-    const size_t saved_content_size = result_.content.size();
-    const size_t saved_reasoning_size = result_.reasoning_content.size();
-
-    auto restore_state = [&]() {
-        move_to(saved_pos);
-        result_.content.resize(saved_content_size);
-        result_.reasoning_content.resize(saved_reasoning_size);
-    };
-
-    // Allow leading whitespace to be preserved as content when reasoning is present at the start
-    size_t cursor = pos_;
-    size_t whitespace_end = cursor;
-    while (whitespace_end < input_.size() && std::isspace(static_cast<unsigned char>(input_[whitespace_end]))) {
-        ++whitespace_end;
-    }
-
-    if (whitespace_end >= input_.size()) {
-        restore_state();
-        if (syntax_.thinking_forced_open) {
-            auto rest = input_.substr(saved_pos);
+    if (syntax_.reasoning_format != COMMON_REASONING_FORMAT_NONE) {
+        if (syntax_.thinking_forced_open || try_consume_literal(start_think)) {
+            if (auto res = try_find_literal(end_think)) {
+                handle_reasoning(res->prelude, /* closed */ true);
+                consume_spaces();
+                return true;
+            }
+            auto rest = consume_rest();
             if (!rest.empty()) {
                 handle_reasoning(rest, /* closed */ !is_partial());
             }
-            move_to(input_.size());
+            // Allow unclosed thinking tags, for now (https://github.com/ggml-org/llama.cpp/issues/13812, https://github.com/ggml-org/llama.cpp/issues/13877)
+            // if (!syntax_.thinking_forced_open) {
+            //     throw common_chat_msg_partial_exception(end_think);
+            // }
             return true;
         }
-        return false;
-    }
-
-    cursor = whitespace_end;
-    const size_t remaining = input_.size() - cursor;
-    const size_t start_prefix = std::min(start_think.size(), remaining);
-    const bool has_start_tag = input_.compare(cursor, start_prefix, start_think, 0, start_prefix) == 0;
-
-    if (has_start_tag && start_prefix < start_think.size()) {
-        move_to(input_.size());
-        return true;
-    }
-
-    if (has_start_tag) {
-        if (whitespace_end > pos_) {
-            add_content(input_.substr(pos_, whitespace_end - pos_));
-        }
-        set_reasoning_prefix(cursor);
-        cursor += start_think.size();
-    } else if (syntax_.thinking_forced_open) {
-        cursor = whitespace_end;
-    } else {
-        restore_state();
-        return false;
-    }
-    while (true) {
-        if (cursor >= input_.size()) {
-            move_to(input_.size());
-            return true;
-        }
-
-        size_t end_pos = input_.find(end_think, cursor);
-        if (end_pos == std::string::npos) {
-            std::string_view remaining_view(input_.data() + cursor, input_.size() - cursor);
-            size_t partial_off = string_find_partial_stop(remaining_view, end_think);
-            size_t reasoning_end = partial_off == std::string::npos ? input_.size() : cursor + partial_off;
-            if (reasoning_end > cursor) {
-                handle_reasoning(input_.substr(cursor, reasoning_end - cursor), /* closed */ partial_off == std::string::npos && !is_partial());
-            }
-            move_to(input_.size());
-            return true;
-        }
-
-        if (end_pos > cursor) {
-            handle_reasoning(input_.substr(cursor, end_pos - cursor), /* closed */ true);
-        } else {
-            handle_reasoning("", /* closed */ true);
-        }
-
-        cursor = end_pos + end_think.size();
-
-        while (cursor < input_.size() && std::isspace(static_cast<unsigned char>(input_[cursor]))) {
-            ++cursor;
-        }
-
-        const size_t next_remaining = input_.size() - cursor;
-        if (next_remaining == 0) {
-            move_to(cursor);
-            return true;
-        }
-
-        const size_t next_prefix = std::min(start_think.size(), next_remaining);
-        if (input_.compare(cursor, next_prefix, start_think, 0, next_prefix) == 0) {
-            if (next_prefix < start_think.size()) {
-                move_to(input_.size());
-                return true;
-            }
-            set_reasoning_prefix(cursor);
-            cursor += start_think.size();
-            continue;
-        }
-
-        move_to(cursor);
-        return true;
     }
+    return false;
 }
 
 std::string common_chat_msg_parser::consume_rest() {
@@ -432,7 +291,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         if (is_arguments_path({})) {
             // Entire JSON is the arguments and was parsed fully.
             return consume_json_result {
-                partial->json.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true),
+                partial->json.dump(),
                 /* .is_partial = */ false,
             };
         }
@@ -444,7 +303,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
     std::vector<std::string> path;
     std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
         if (is_arguments_path(path)) {
-            auto arguments = j.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true);
+            auto arguments = j.dump();
             if (is_partial() && !partial->healing_marker.marker.empty()) {
                 auto idx = arguments.find(partial->healing_marker.json_dump_marker);
                 if (idx != std::string::npos) {

common/chat-parser.h

@@ -64,9 +64,6 @@ class common_chat_msg_parser {
     // Adds an array of tool calls using their "name", "id" and "arguments" fields.
     bool add_tool_calls(const nlohmann::ordered_json & arr);
 
-    // Adds a tool call using the short form: { "tool_name": { "arg1": val, "arg2": val } }
-    bool add_tool_call_short_form(const nlohmann::ordered_json & tool_call);
-
     void finish();
 
     bool consume_spaces();

common/chat.cpp

@@ -9,11 +9,8 @@
 #include <minja/chat-template.hpp>
 #include <minja/minja.hpp>
 
-#include <algorithm>
 #include <cstdio>
-#include <cctype>
 #include <exception>
-#include <functional>
 #include <iostream>
 #include <optional>
 #include <stdexcept>
@@ -628,7 +625,6 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_CONTENT_ONLY: return "Content-only";
         case COMMON_CHAT_FORMAT_GENERIC: return "Generic";
         case COMMON_CHAT_FORMAT_MISTRAL_NEMO: return "Mistral Nemo";
-        case COMMON_CHAT_FORMAT_MAGISTRAL: return "Magistral";
         case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x";
         case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools";
         case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1";
@@ -642,8 +638,6 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_GPT_OSS: return "GPT-OSS";
         case COMMON_CHAT_FORMAT_SEED_OSS: return "Seed-OSS";
         case COMMON_CHAT_FORMAT_NEMOTRON_V2: return "Nemotron V2";
-        case COMMON_CHAT_FORMAT_APERTUS: return "Apertus";
-        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS: return "LFM2 with JSON tools";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -807,7 +801,6 @@ static std::string apply(
     }
     tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
     tmpl_inputs.extra_context = inputs.extra_context;
-    tmpl_inputs.extra_context["enable_thinking"] = inputs.enable_thinking;
     if (additional_context) {
         tmpl_inputs.extra_context.merge_patch(*additional_context);
     }
@@ -989,185 +982,6 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
 }
-
-
-// Case-insensitive find
-static size_t ifind_string(const std::string & haystack, const std::string & needle, size_t pos = 0) {
-    auto it = std::search(
-        haystack.begin() + pos, haystack.end(),
-        needle.begin(), needle.end(),
-        [](char a, char b) { return std::tolower(a) == std::tolower(b); }
-    );
-    return (it == haystack.end()) ? std::string::npos : std::distance(haystack.begin(), it);
-}
-
-static common_chat_params common_chat_params_init_lfm2(const common_chat_template & tmpl, const struct templates_params & inputs) {
-    common_chat_params data;
-    const auto is_json_schema_provided = !inputs.json_schema.is_null();
-    const auto is_grammar_provided = !inputs.grammar.empty();
-    const auto are_tools_provided = inputs.tools.is_array() && !inputs.tools.empty();
-
-    // the logic requires potentially modifying the messages
-    auto tweaked_messages = inputs.messages;
-
-    auto replace_json_schema_marker = [](json & messages) -> bool {
-        static std::string marker1 = "force json schema.\n";
-        static std::string marker2 = "force json schema.";
-
-        if (messages.empty() || messages.at(0).at("role") != "system") {
-            return false;
-        }
-
-        std::string content = messages.at(0).at("content");
-
-        for (const auto & marker : {marker1, marker2}) {
-            const auto pos = ifind_string(content, marker);
-            if (pos != std::string::npos) {
-                content.replace(pos, marker.length(), "");
-                // inject modified content back into the messages
-                messages.at(0).at("content") = content;
-                return true;
-            }
-        }
-
-        return false;
-    };
-
-    // Lfm2 model does not natively work with json, but can generally understand the tools structure
-    //
-    // Example of the pytorch dialog structure:
-    //     <|startoftext|><|im_start|>system
-    //     List of tools: <|tool_list_start|>[{"name": "get_candidate_status", "description": "Retrieves the current status of a candidate in the recruitment process", "parameters": {"type": "object", "properties": {"candidate_id": {"type": "string", "description": "Unique identifier for the candidate"}}, "required": ["candidate_id"]}}]<|tool_list_end|><|im_end|>
-    //     <|im_start|>user
-    //     What is the current status of candidate ID 12345?<|im_end|>
-    //     <|im_start|>assistant
-    //     <|tool_call_start|>[get_candidate_status(candidate_id="12345")]<|tool_call_end|>Checking the current status of candidate ID 12345.<|im_end|>
-    //     <|im_start|>tool
-    //     <|tool_response_start|>{"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"}<|tool_response_end|><|im_end|>
-    //     <|im_start|>assistant
-    //     The candidate with ID 12345 is currently in the "Interview Scheduled" stage for the position of Clinical Research Associate, with an interview date set for 2023-11-20.<|im_end|>
-    //
-    // For the llama server compatibility with json tools semantic,
-    // the client can add "Follow json schema." line into the system message prompt to force the json output.
-    //
-    if (are_tools_provided && (is_json_schema_provided || is_grammar_provided)) {
-        // server/utils.hpp prohibits that branch for the custom grammar anyways
-        throw std::runtime_error("Tools call must not use \"json_schema\" or \"grammar\", use non-tool invocation if you want to use custom grammar");
-    } else if (are_tools_provided && replace_json_schema_marker(tweaked_messages)) {
-        LOG_INF("%s: Using tools to build a grammar\n", __func__);
-
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            auto schemas = json::array();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                schemas.push_back({
-                    {"type", "object"},
-                    {"properties", {
-                        {"name", {
-                            {"type", "string"},
-                            {"const", function.at("name")},
-                        }},
-                        {"arguments", function.at("parameters")},
-                    }},
-                    {"required", json::array({"name", "arguments", "id"})},
-                });
-            });
-            auto schema = json {
-                {"type", "array"},
-                {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
-                {"minItems", 1},
-            };
-            if (!inputs.parallel_tool_calls) {
-                schema["maxItems"] = 1;
-            }
-
-            builder.add_rule("root", "\"<|tool_call_start|>\"" + builder.add_schema("tool_calls", schema) + "\"<|tool_call_end|>\"");
-        });
-        // model has no concept of tool selection mode choice,
-        // if the system prompt rendered correctly it will produce a tool call
-        // the grammar goes inside the tool call body
-        data.grammar_lazy = true;
-        data.grammar_triggers = {{COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL, "\\s*<\\|tool_call_start\\|>\\s*\\["}};
-        data.preserved_tokens = {"<|tool_call_start|>", "<|tool_call_end|>"};
-        data.format = COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS;
-    } else if (are_tools_provided && (!is_json_schema_provided && !is_grammar_provided)) {
-        LOG_INF("%s: Using tools without json schema or grammar\n", __func__);
-        // output those tokens
-        data.preserved_tokens = {"<|tool_call_start|>", "<|tool_call_end|>"};
-    } else if (is_json_schema_provided) {
-        LOG_INF("%s: Using provided json schema to build a grammar\n", __func__);
-        data.grammar = json_schema_to_grammar(inputs.json_schema);
-    } else if (is_grammar_provided) {
-        LOG_INF("%s: Using provided grammar\n", __func__);
-        data.grammar = inputs.grammar;
-    } else {
-        LOG_INF("%s: Using content relying on the template\n", __func__);
-    }
-
-    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
-    LOG_DBG("%s: Prompt: %s\n", __func__, data.prompt.c_str());
-
-    return data;
-}
-
-static common_chat_params common_chat_params_init_magistral(const common_chat_template & tmpl, const struct templates_params & inputs) {
-    common_chat_params data;
-    data.prompt = apply(tmpl, inputs);
-    data.format = COMMON_CHAT_FORMAT_MAGISTRAL;
-    data.preserved_tokens = {
-        "[THINK]",
-        "[/THINK]",
-    };
-
-    if (inputs.tools.is_array() && !inputs.tools.empty()) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            auto schemas = json::array();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                schemas.push_back({
-                    {"type", "object"},
-                    {"properties", {
-                        {"name", {
-                            {"type", "string"},
-                            {"const", function.at("name")},
-                        }},
-                        {"arguments", function.at("parameters")},
-                        {"id", {
-                            {"type", "string"},
-                            {"pattern", "^[a-zA-Z0-9]{9}$"},
-                        }},
-                    }},
-                    {"required", json::array({"name", "arguments", "id"})},
-                });
-            });
-            auto schema = json {
-                {"type", "array"},
-                {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
-                {"minItems", 1},
-            };
-            if (!inputs.parallel_tool_calls) {
-                schema["maxItems"] = 1;
-            }
-            builder.add_rule("root", "\"[TOOL_CALLS]\" " + builder.add_schema("tool_calls", schema));
-        });
-        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]"});
-        data.preserved_tokens.push_back("[TOOL_CALLS]");
-    } else {
-        data.grammar_lazy = false;
-        if (!inputs.json_schema.is_null()) {
-            if (!inputs.grammar.empty()) {
-                throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both");
-            }
-            data.grammar = json_schema_to_grammar(inputs.json_schema);
-        } else {
-            data.grammar = inputs.grammar;
-        }
-    }
-
-    return data;
-}
-
 static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
     if (!builder.syntax().parse_tool_calls) {
         builder.add_content(builder.consume_rest());
@@ -1178,18 +992,6 @@ static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
     parse_prefixed_json_tool_call_array(builder, prefix);
 }
 
-static void common_chat_parse_magistral(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("[THINK]", "[/THINK]");
-
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
-    parse_prefixed_json_tool_call_array(builder, prefix);
-}
-
 static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
@@ -1462,78 +1264,7 @@ static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_
     }
     return data;
 }
-
-static common_chat_params common_chat_params_init_apertus(const common_chat_template & tmpl, const struct templates_params & inputs) {
-    common_chat_params data;
-
-    // Generate the prompt using the apply() function with the template
-    data.prompt = apply(tmpl, inputs);
-    data.format = COMMON_CHAT_FORMAT_APERTUS;
-
-    // Handle thinking tags appropriately based on inputs.enable_thinking
-    if (string_ends_with(data.prompt, "<|inner_prefix|>")) {
-        if (!inputs.enable_thinking) {
-            data.prompt += "<|inner_suffix|>";
-        } else {
-            data.thinking_forced_open = true;
-        }
-    }
-
-    // When tools are present, build grammar for the <|tools_prefix|> format
-    if (!inputs.tools.is_null() && inputs.tools.is_array() && !inputs.tools.empty()) {
-        data.grammar_lazy = true;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            auto schemas = json::array();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                schemas.push_back({
-                    { "type",       "object"                                                   },
-                    { "properties",
-                        {
-                            { function.at("name"), function.at("parameters") }
-                        }                                                                        },
-                    { "required",   json::array({ function.at("name") }) },
-                });
-            });
-            auto schema = json{
-                        { "type",     "array"                                                         },
-                        { "items",    schemas.size() == 1 ? schemas[0] : json{ { "anyOf", schemas } } },
-                        { "minItems", 1                                                               },
-            };
-            if (!inputs.parallel_tool_calls) {
-                schema["maxItems"] = 1;
-            }
-            builder.add_rule("root",
-                                std::string(data.thinking_forced_open ? "( \"<|inner_suffix|>\" space )? " : "") +
-                                    "\"<|tools_prefix|>\"" + builder.add_schema("tool_calls", schema) + "\"<|tools_suffix|>\"");
-                            });
-        data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
-            // If thinking_forced_open, then we capture the <|inner_suffix|> tag in the grammar,
-            // (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
-            std::string(data.thinking_forced_open ?
-                            "[\\s\\S]*?(<\\|inner_suffix\\|>\\s*)" :
-                            "(?:<\\|inner_prefix\\|>[\\s\\S]*?<\\|inner_suffix\\|>\\s*)?") +
-                "(<\\|tools_prefix\\|>)[\\s\\S]*" });
-        data.preserved_tokens = {
-            "<|system_start|>",
-            "<|system_end|>",
-            "<|developer_start|>",
-            "<|developer_end|>",
-            "<|user_start|>",
-            "<|user_end|>",
-            "<|assistant_start|>",
-            "<|assistant_end|>",
-            "<|inner_prefix|>",
-            "<|inner_suffix|>",
-            "<|tools_prefix|>",
-            "<|tools_suffix|>",
-        };
-    }
-    return data;
-}
 static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
-    builder.try_parse_reasoning("<think>", "</think>");
-
     if (!builder.syntax().parse_tool_calls) {
         builder.add_content(builder.consume_rest());
         return;
@@ -1885,36 +1616,17 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
                 );
             });
 
+            auto recipient_in_role = builder.add_rule("recipient_in_role",
+                "\"<|start|>assistant\"? \" to=functions.\" ( " +
+                string_join(tool_rules_recipient_in_role, " | ") + " )"
+            );
+
             auto recipient_in_channel = builder.add_rule("recipient_in_channel",
                 channel + " \" to=functions.\" ( " +
                 string_join(tool_rules_recipient_in_channel, " | ") + " )"
             );
 
-            if (data.grammar_lazy) {
-                auto recipient_in_role = builder.add_rule("recipient_in_role",
-                    "\"<|start|>assistant\"? \" to=functions.\" ( " +
-                    string_join(tool_rules_recipient_in_role, " | ") + " )"
-                );
-
-                builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
-            } else {
-                auto not_end = builder.add_rule("not-end",
-                    "[^<] | \"<\" [^|] | \"<|\" [^e] | \"<|e\" [^n] | \"<|en\" [^d] | \"<|end\" [^|] | \"<|end|\" [^>]");
-                auto analysis = builder.add_rule("analysis",
-                    "\"<|channel|>analysis<|message|>\" ( " + not_end + " )* \"<|end|>\"");
-                auto commentary = builder.add_rule("commentary",
-                    "\"<|channel|>commentary<|message|>\" ( " + not_end + " )* \"<|end|>\"");
-
-                auto recipient_in_role = builder.add_rule("recipient_in_role",
-                    "\" to=functions.\" ( " + string_join(tool_rules_recipient_in_role, " | ") + " )"
-                );
-
-                builder.add_rule("root",
-                    "( " + analysis + " \"<|start|>assistant\" )? " +
-                    "( " + commentary + " \"<|start|>assistant\" )? " +
-                    "( " + recipient_in_role + " | " + recipient_in_channel + " )"
-                );
-            }
+            builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
 
             // Trigger on tool calls that appear in the commentary channel
             data.grammar_triggers.push_back({
@@ -2592,102 +2304,6 @@ static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) {
     builder.add_content(builder.consume_rest());
 }
 
-static void common_chat_parse_apertus(common_chat_msg_parser & builder) {
-    // Parse thinking tags
-    builder.try_parse_reasoning("<|inner_prefix|>", "<|inner_suffix|>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // Look for tool calls
-    static const common_regex tool_call_regex(regex_escape("<|tools_prefix|>"));
-    if (auto res = builder.try_find_regex(tool_call_regex)) {
-        builder.move_to(res->groups[0].end);
-
-        auto tool_calls_data = builder.consume_json();
-        if (tool_calls_data.json.is_array()) {
-            builder.consume_spaces();
-            if (!builder.try_consume_literal("<|tools_suffix|>")) {
-                throw common_chat_msg_partial_exception("Incomplete tool call");
-            }
-            for (const auto & value : tool_calls_data.json) {
-                if (value.is_object()) {
-                    builder.add_tool_call_short_form(value);
-                }
-            }
-        } else {
-            throw common_chat_msg_partial_exception("Incomplete tool call");
-        }
-    }
-    builder.add_content(builder.consume_rest());
-}
-
-
-static void common_chat_parse_lfm2(common_chat_msg_parser & builder) {
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // LFM2 format: <|tool_call_start|>[{"name": "get_current_time", "arguments": {"location": "Paris"}}]<|tool_call_end|>
-    static const common_regex tool_call_start_regex(regex_escape("<|tool_call_start|>"));
-    static const common_regex tool_call_end_regex(regex_escape("<|tool_call_end|>"));
-
-    // Loop through all tool calls
-    while (auto res = builder.try_find_regex(tool_call_start_regex, std::string::npos, /* add_prelude_to_content= */ true)) {
-        builder.move_to(res->groups[0].end);
-
-        // Parse JSON array format: [{"name": "...", "arguments": {...}}]
-        auto tool_calls_data = builder.consume_json();
-
-        // Consume end marker
-        builder.consume_spaces();
-        if (!builder.try_consume_regex(tool_call_end_regex)) {
-            throw common_chat_msg_partial_exception("Expected <|tool_call_end|>");
-        }
-
-        // Process each tool call in the array
-        if (tool_calls_data.json.is_array()) {
-            for (const auto & tool_call : tool_calls_data.json) {
-                if (!tool_call.is_object()) {
-                    throw common_chat_msg_partial_exception("Tool call must be an object");
-                }
-
-                if (!tool_call.contains("name")) {
-                    throw common_chat_msg_partial_exception("Tool call missing 'name' field");
-                }
-
-                std::string function_name = tool_call.at("name");
-                std::string arguments = "{}";
-
-                if (tool_call.contains("arguments")) {
-                    if (tool_call.at("arguments").is_object()) {
-                        arguments = tool_call.at("arguments").dump();
-                    } else if (tool_call.at("arguments").is_string()) {
-                        arguments = tool_call.at("arguments");
-                    }
-                }
-
-                if (!builder.add_tool_call(function_name, "", arguments)) {
-                    throw common_chat_msg_partial_exception("Incomplete tool call");
-                }
-            }
-        } else {
-            throw common_chat_msg_partial_exception("Expected JSON array for tool calls");
-        }
-
-        // Consume any trailing whitespace after this tool call
-        builder.consume_spaces();
-    }
-
-    // Consume any remaining content after all tool calls
-    auto remaining = builder.consume_rest();
-    if (!string_strip(remaining).empty()) {
-        builder.add_content(remaining);
-    }
-}
-
 static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) {
     // Parse thinking tags first - this handles the main reasoning content
     builder.try_parse_reasoning("<seed:think>", "</seed:think>");
@@ -2932,17 +2548,6 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_nemotron_v2(tmpl, params);
     }
 
-    // Apertus format detection
-    if (src.find("<|system_start|>") != std::string::npos && src.find("<|tools_prefix|>") != std::string::npos) {
-        return common_chat_params_init_apertus(tmpl, params);
-    }
-
-    // LFM2 (w/ tools)
-    if (src.find("List of tools: <|tool_list_start|>[") != std::string::npos &&
-        src.find("]<|tool_list_end|>") != std::string::npos) {
-        return common_chat_params_init_lfm2(tmpl, params);
-    }
-
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
     if ((params.tools.is_array() && params.json_schema.is_object())) {
@@ -2971,10 +2576,6 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
     }
 
-    if (src.find("[THINK]") != std::string::npos && src.find("[/THINK]") != std::string::npos) {
-        return common_chat_params_init_magistral(tmpl, params);
-    }
-
     // Plain handler (no tools)
     if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
         return common_chat_params_init_without_tools(tmpl, params);
@@ -3059,7 +2660,6 @@ common_chat_params common_chat_templates_apply(
 }
 
 static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("<think>", "</think>");
     builder.add_content(builder.consume_rest());
 }
 
@@ -3076,9 +2676,6 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_MISTRAL_NEMO:
             common_chat_parse_mistral_nemo(builder);
             break;
-        case COMMON_CHAT_FORMAT_MAGISTRAL:
-            common_chat_parse_magistral(builder);
-            break;
         case COMMON_CHAT_FORMAT_LLAMA_3_X:
             common_chat_parse_llama_3_1(builder);
             break;
@@ -3118,12 +2715,6 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_NEMOTRON_V2:
             common_chat_parse_nemotron_v2(builder);
             break;
-        case COMMON_CHAT_FORMAT_APERTUS:
-            common_chat_parse_apertus(builder);
-            break;
-        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS:
-            common_chat_parse_lfm2(builder);
-            break;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }

common/chat.h

@@ -33,8 +33,8 @@ struct common_chat_msg_content_part {
 struct common_chat_msg {
     std::string role;
     std::string content;
-    std::vector<common_chat_msg_content_part> content_parts;
-    std::vector<common_chat_tool_call> tool_calls;
+    std::vector<common_chat_msg_content_part> content_parts = {};
+    std::vector<common_chat_tool_call> tool_calls = {};
     std::string reasoning_content;
     std::string tool_name;
     std::string tool_call_id;
@@ -44,7 +44,7 @@ struct common_chat_msg {
     bool empty() const {
         return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
     }
-    void set_tool_call_ids(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
         for (auto i = 0u; i < tool_calls.size(); i++) {
             if (ids_cache.size() <= i) {
                 auto id = tool_calls[i].id;
@@ -101,7 +101,6 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_CONTENT_ONLY,
     COMMON_CHAT_FORMAT_GENERIC,
     COMMON_CHAT_FORMAT_MISTRAL_NEMO,
-    COMMON_CHAT_FORMAT_MAGISTRAL,
     COMMON_CHAT_FORMAT_LLAMA_3_X,
     COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
     COMMON_CHAT_FORMAT_DEEPSEEK_R1,
@@ -115,8 +114,6 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_GPT_OSS,
     COMMON_CHAT_FORMAT_SEED_OSS,
     COMMON_CHAT_FORMAT_NEMOTRON_V2,
-    COMMON_CHAT_FORMAT_APERTUS,
-    COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };

common/common.cpp

@@ -1133,7 +1133,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;
-    mparams.no_host         = params.no_host;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;

common/common.h

@@ -378,7 +378,7 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = false; // context shift on infinite text generation
+    bool ctx_shift         = false;  // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 
@@ -392,7 +392,6 @@ struct common_params {
     bool check_tensors     = false; // validate tensor data
     bool no_op_offload     = false; // globally disable offload host tensor operations to device
     bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
-    bool no_host           = false; // bypass host buffer allowing extra buffers to be used
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -425,8 +424,7 @@ struct common_params {
     int32_t timeout_write     = timeout_read; // http write timeout in seconds
     int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
-    int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
-    int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
+    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -434,7 +432,7 @@ struct common_params {
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
     int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 

common/http.h

@@ -1,73 +0,0 @@
-#pragma once
-
-#include <cpp-httplib/httplib.h>
-
-struct common_http_url {
-    std::string scheme;
-    std::string user;
-    std::string password;
-    std::string host;
-    std::string path;
-};
-
-static common_http_url common_http_parse_url(const std::string & url) {
-    common_http_url parts;
-    auto scheme_end = url.find("://");
-
-    if (scheme_end == std::string::npos) {
-        throw std::runtime_error("invalid URL: no scheme");
-    }
-    parts.scheme = url.substr(0, scheme_end);
-
-    if (parts.scheme != "http" && parts.scheme != "https") {
-        throw std::runtime_error("unsupported URL scheme: " + parts.scheme);
-    }
-
-    auto rest = url.substr(scheme_end + 3);
-    auto at_pos = rest.find('@');
-
-    if (at_pos != std::string::npos) {
-        auto auth = rest.substr(0, at_pos);
-        auto colon_pos = auth.find(':');
-        if (colon_pos != std::string::npos) {
-            parts.user = auth.substr(0, colon_pos);
-            parts.password = auth.substr(colon_pos + 1);
-        } else {
-            parts.user = auth;
-        }
-        rest = rest.substr(at_pos + 1);
-    }
-
-    auto slash_pos = rest.find('/');
-
-    if (slash_pos != std::string::npos) {
-        parts.host = rest.substr(0, slash_pos);
-        parts.path = rest.substr(slash_pos);
-    } else {
-        parts.host = rest;
-        parts.path = "/";
-    }
-    return parts;
-}
-
-static std::pair<httplib::Client, common_http_url> common_http_client(const std::string & url) {
-    common_http_url parts = common_http_parse_url(url);
-
-    if (parts.host.empty()) {
-        throw std::runtime_error("error: invalid URL format");
-    }
-
-    httplib::Client cli(parts.scheme + "://" + parts.host);
-
-    if (!parts.user.empty()) {
-        cli.set_basic_auth(parts.user, parts.password);
-    }
-
-    cli.set_follow_location(true);
-
-    return { std::move(cli), std::move(parts) };
-}
-
-static std::string common_http_show_masked_url(const common_http_url & parts) {
-    return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path;
-}

common/json-partial.cpp

@@ -5,7 +5,6 @@
 #include <nlohmann/json.hpp>
 
 #include <string>
-#include <regex>
 
 using json = nlohmann::ordered_json;
 
@@ -169,47 +168,6 @@ bool common_json_parse(
                 }
             }
 
-            // Matches a potentially partial unicode escape sequence, e.g. \u, \uX, \uXX, \uXXX, \uXXXX
-            static const std::regex partial_unicode_regex(R"(\\u(?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F])?)?)?)?$)");
-
-            auto is_high_surrogate = [&](const std::string & s) {
-                // Check if a partial of a high surrogate (U+D800-U+DBFF)
-                return s.length() >= 4 &&
-                    s[0] == '\\' && s[1] == 'u' &&
-                    std::tolower(s[2]) == 'd' &&
-                    (s[3] == '8' || s[3] == '9' || std::tolower(s[3]) == 'a' || std::tolower(s[3]) == 'b');
-            };
-
-            // Initialize the unicode marker to a low surrogate to handle the edge case
-            // where a high surrogate (U+D800-U+DBFF) is immediately followed by a
-            // backslash (\)
-            std::string unicode_marker_padding = "udc00";
-            std::smatch last_unicode_seq;
-
-            if (std::regex_search(str, last_unicode_seq, partial_unicode_regex)) {
-                std::smatch second_last_seq;
-                std::string prelude = str.substr(0, last_unicode_seq.position());
-
-                // Pad the escape sequence with 0s until it forms a complete sequence of 6 characters
-                unicode_marker_padding = std::string(6 - last_unicode_seq.length(), '0');
-
-                if (is_high_surrogate(last_unicode_seq.str())) {
-                    // If the sequence is a partial match for a high surrogate, add a low surrogate (U+DC00-U+UDFF)
-                    unicode_marker_padding += "\\udc00";
-                } else if (std::regex_search(prelude, second_last_seq, partial_unicode_regex)) {
-                    if (is_high_surrogate(second_last_seq.str())) {
-                        // If this follows a high surrogate, pad it to be a low surrogate
-                        if (last_unicode_seq.length() == 2) {
-                            unicode_marker_padding = "dc00";
-                        } else if (last_unicode_seq.length() == 3) {
-                            unicode_marker_padding = "c00";
-                        } else {
-                            // The original unicode_marker_padding is already padded with 0s
-                        }
-                    }
-                }
-            }
-
             const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
 
             if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
@@ -228,9 +186,6 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
                     // Was inside an object value string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
-                    // Was inside an object value string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
                 } else {
                     // find last :
                     auto last_pos = str.find_last_of(':');
@@ -250,9 +205,6 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
                     // Was inside an array value string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
-                    // Was inside an array value string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
                 } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
                     // Had just finished a value
                     str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
@@ -278,9 +230,6 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
                     // Was inside an object key string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\": 1" + closing)) {
-                    // Was inside an object key string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\": 1" + closing;
                 } else {
                     auto last_pos = str.find_last_of(':');
                     if (last_pos == std::string::npos) {

common/json-schema-to-grammar.cpp

@@ -41,9 +41,9 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-static void _build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
-    auto has_min = min_value != std::numeric_limits<int64_t>::min();
-    auto has_max = max_value != std::numeric_limits<int64_t>::max();
+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();
 
     auto digit_range = [&](char from, char to) {
         out << "[";
@@ -159,7 +159,7 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
     if (has_min) {
         if (min_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
             out << ") | [0] | [1-9] ";
             more_digits(0, decimals_left - 1);
         } else if (min_value == 0) {
@@ -194,7 +194,7 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
             }
             digit_range(c, c);
             out << " (";
-            _build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
+            _build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
             out << ")";
             if (c < '9') {
                 out << " | ";
@@ -216,7 +216,7 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
             _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
         } else {
             out << "\"-\" (";
-            _build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
             out << ")";
         }
         return;
@@ -601,10 +601,7 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        auto it = ref.find('#');
-        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
-        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
-        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
+        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -777,24 +774,11 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_object() && target.contains(sel)) {
-                                target = target[sel];
-                            } else if (target.is_array()) {
-                                size_t sel_index;
-                                try {
-                                    sel_index = std::stoul(sel);
-                                } catch (const std::invalid_argument & e) {
-                                    sel_index = target.size();
-                                }
-                                if (sel_index >= target.size()) {
-                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
-                                    return;
-                                }
-                                target = target[sel_index];
-                            } else {
+                            if (target.is_null() || !target.contains(sel)) {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
+                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }
@@ -941,17 +925,17 @@ public:
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
         } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
-            int64_t min_value = std::numeric_limits<int64_t>::min();
-            int64_t max_value = std::numeric_limits<int64_t>::max();
+            int min_value = std::numeric_limits<int>::min();
+            int max_value = std::numeric_limits<int>::max();
             if (schema.contains("minimum")) {
-                min_value = schema["minimum"].get<int64_t>();
+                min_value = schema["minimum"].get<int>();
             } else if (schema.contains("exclusiveMinimum")) {
-                min_value = schema["exclusiveMinimum"].get<int64_t>() + 1;
+                min_value = schema["exclusiveMinimum"].get<int>() + 1;
             }
             if (schema.contains("maximum")) {
-                max_value = schema["maximum"].get<int64_t>();
+                max_value = schema["maximum"].get<int>();
             } else if (schema.contains("exclusiveMaximum")) {
-                max_value = schema["exclusiveMaximum"].get<int64_t>() - 1;
+                max_value = schema["exclusiveMaximum"].get<int>() - 1;
             }
             std::stringstream out;
             out << "(";

convert_hf_to_gguf_update.py

@@ -139,8 +139,7 @@ models = [
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
-    {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
-    {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
+    {"name": "llada-moe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/backend/SYCL.md

@@ -145,13 +145,12 @@ The docker build option is currently limited to *Intel GPU* targets.
 ```sh
 # Using FP16
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
-
-# Using FP32
-docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=OFF" --target light -f .devops/intel.Dockerfile .
 ```
 
 *Notes*:
 
+To build in default FP32 *(Slower than FP16 alternative)*, set `--build-arg="GGML_SYCL_F16=OFF"` in the previous command.
+
 You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
 Check the [documentation for Docker](../docker.md) to see the available images.
 
@@ -161,7 +160,7 @@ Check the [documentation for Docker](../docker.md) to see the available images.
 # First, find all the DRI cards
 ls -la /dev/dri
 # Then, pick the card that you want to use (here for e.g. /dev/dri/card1).
-docker run -it --rm -v "/path/to/models:/models" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card0:/dev/dri/card0 llama-cpp-sycl -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -c 4096 -s 0
+docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
 ```
 
 *Notes:*
@@ -216,19 +215,9 @@ To target AMD GPUs with SYCL, the ROCm stack must be installed first.
 
 2. **Install Intel® oneAPI Base toolkit**
 
-SYCL backend depends on:
-  - Intel® oneAPI DPC++/C++ compiler/running-time.
-  - Intel® oneAPI DPC++/C++ library (oneDPL).
-  - Intel® oneAPI Deep Neural Network Library (oneDNN).
-  - Intel® oneAPI Math Kernel Library (oneMKL).
-
 - **For Intel GPU**
 
-All above are included in both **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** packages.
-
-It's recommended to install **Intel® Deep Learning Essentials** which only provides the necessary libraries with less size.
-
-The **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
 
 Please follow the instructions for downloading and installing the Toolkit for Linux, and preferably keep the default installation values unchanged, notably the installation path *(`/opt/intel/oneapi` by default)*.
 
@@ -236,12 +225,6 @@ Following guidelines/code snippets assume the default installation values. Other
 
 Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs.
 
-|Verified release|
-|-|
-|2025.2.1|
-|2025.1|
-|2024.1|
-
 - **Adding support to Nvidia GPUs**
 
 **oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
@@ -272,11 +255,10 @@ sycl-ls
 When targeting an intel GPU, the user should expect one or more devices among the available SYCL devices. Please make sure that at least one GPU is present via `sycl-ls`, for instance `[level_zero:gpu]` in the sample output below:
 
 ```
-[level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Arc(TM) A770 Graphics 12.55.8 [1.3.29735+27]
-[level_zero:gpu][level_zero:1] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) UHD Graphics 730 12.2.0 [1.3.29735+27]
-[opencl:cpu][opencl:0] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i5-13400 OpenCL 3.0 (Build 0) [2025.20.8.0.06_160000]
-[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [24.39.31294]
-[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) UHD Graphics 730 OpenCL 3.0 NEO  [24.39.31294]
+[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu][opencl:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
 ```
 
 - **Nvidia GPU**
@@ -371,7 +353,7 @@ cmake --build build --config Release -j -v
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf?download=true) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -484,17 +466,7 @@ If you already have a recent version of Microsoft Visual Studio, you can skip th
 
 3. Install Intel® oneAPI Base toolkit
 
-SYCL backend depends on:
-  - Intel® oneAPI DPC++/C++ compiler/running-time.
-  - Intel® oneAPI DPC++/C++ library (oneDPL).
-  - Intel® oneAPI Deep Neural Network Library (oneDNN).
-  - Intel® oneAPI Math Kernel Library (oneMKL).
-
-All above are included in both **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** packages.
-
-It's recommended to install **Intel® Deep Learning Essentials** which only provides the necessary libraries with less size.
-
-The **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
 
 Please follow the instructions for downloading and installing the Toolkit for Windows, and preferably keep the default installation values unchanged, notably the installation path *(`C:\Program Files (x86)\Intel\oneAPI` by default)*.
 

docs/backend/hexagon/CMakeUserPresets.json

@@ -1,49 +0,0 @@
-{
-  "version": 4,
-  "configurePresets": [
-    {
-        "name": "arm64-android-snapdragon",
-        "hidden": true,
-        "architecture": { "value": "arm64",       "strategy": "external" },
-        "toolset":      { "value": "host=x86_64", "strategy": "external" },
-        "cacheVariables": {
-            "ANDROID_ABI":      "arm64-v8a",
-            "ANDROID_PLATFORM": "android-31",
-            "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
-            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
-            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
-            "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
-            "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
-            "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
-            "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
-            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
-            "PREBUILT_LIB_DIR": "android_aarch64",
-            "GGML_OPENMP":      "OFF",
-            "GGML_LLAMAFILE":   "OFF",
-            "GGML_OPENCL":      "ON",
-            "GGML_HEXAGON":     "ON",
-            "LLAMA_CURL":       "OFF"
-        }
-    },
-
-    {
-        "name": "arm64-windows-snapdragon",
-        "inherits": [ "base", "arm64-windows-llvm" ],
-        "cacheVariables": {
-            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
-            "PREBUILT_LIB_DIR": "windows_aarch64",
-            "GGML_OPENMP":      "OFF",
-            "GGML_LLAMAFILE":   "OFF",
-            "GGML_OPENCL":      "ON",
-            "GGML_HEXAGON":     "ON",
-            "LLAMA_CURL":       "OFF"
-        }
-    },
-
-    { "name": "arm64-android-snapdragon-debug"  , "inherits": [ "base", "arm64-android-snapdragon", "debug" ] },
-    { "name": "arm64-android-snapdragon-release", "inherits": [ "base", "arm64-android-snapdragon", "release" ] },
-
-    { "name": "arm64-windows-snapdragon-debug"  , "inherits": [ "base", "arm64-windows-snapdragon", "debug" ] },
-    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] }
-  ]
-}

docs/backend/hexagon/README.md

@@ -1,239 +0,0 @@
-# Snapdragon-based Android devices
-
-## How to Build
-
-The easiest way to build llama.cpp for a Snapdragon-based Android device is using the toolchain Docker image (see github.com/snapdragon-toolchain).
-This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
-
-This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
-
-```
-~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.3
-[d]/> cd /workspace
-```
-
-The rest of the Android build process assumes that you're running inside the toolchain container.
-Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
-
-```
-[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
-
-[d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
-Preset CMake variables:
-  ANDROID_ABI="arm64-v8a"
-  ...
-  CMAKE_TOOLCHAIN_FILE="/opt/android-ndk-r28b/build/cmake/android.toolchain.cmake"
-  GGML_HEXAGON="ON"
-  GGML_OPENCL="ON"
-  GGML_OPENMP="OFF"
-  HEXAGON_SDK_ROOT="/opt/hexagon/6.4.0.2"
-...
--- Including OpenCL backend
--- Including Hexagon backend
-...
--- Build files have been written to: /workspace/build-snapdragon
-
-[d]/workspace> cmake --build build-snapdragon
-...
-[144/356] Performing build step for 'htp-v73'
-[1/16] Generating htp_iface_skel.c, htp_iface_stub.c, htp_iface.h
-[2/16] Building C object CMakeFiles/ggml-htp-v73.dir/hvx-sigmoid.c.obj
-[3/16] Building C object CMakeFiles/ggml-htp-v73.dir/htp-dma.c.obj
-[4/16] Building C object CMakeFiles/ggml-htp-v73.dir/worker-pool.c.obj
-...
--- Installing: /workspace/build-snapdragon/ggml/src/ggml-hexagon/libggml-htp-v73.so
--- Installing: /workspace/build-snapdragon/ggml/src/ggml-hexagon/libggml-htp-v75.so
-...
-```
-
-To generate an installable "package" simply use cmake --install:
-
-```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-adb/llama.cpp
--- Install configuration: "Release"
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-cpu.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-opencl.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-hexagon.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v73.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v75.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v79.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v81.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml.so
-...
--- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-bench
--- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-cli
-...
-```
-
-## How to Install
-
-For this step, your device needs to be configured for on-device development.
-Please see https://developer.android.com/studio/debug/dev-options for details.
-
-Once ADB is enabled, use `adb push` to install `pkg-snapdragon` on the device.
-**Note that the toolchain Docker image doesn't have ADB and doesn't set up the ADB bridge. Please use native ADB on the host.**
-
-```
-~/src/llama.cpp$ adb push pkg-adb/llama.cpp /data/local/tmp/
-pkg-adb/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
-pkg-adb/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
-pkg-adb/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
-102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
-```
-
-At this point, you should also install some models:
-
-```
-~/src/llama.cpp$ wget https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf
-...
-2025-10-11 12:04:52 (10.7 MB/s) - ‘Llama-3.2-1B-Instruct-Q4_0.gguf’ saved [773025920/773025920]
-
-~/src/llama.cpp$ adb push Llama-3.2-1B-Instruct-Q4_0.gguf /data/local/tmp/gguf
-Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920 bytes in 19.250s)
-```
-
-## How to Run
-
-The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
-
-llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU (HTP0-4).
-You can select which backend to run the model on using the `D=` variable, which maps to the `--device` option.
-
-Hexagon NPU behaves as a "GPU" device when it comes to `-ngl` and other offload-related options.
-
-Here are some examples of running various llama.cpp tools via ADB.
-
-Simple question for Llama-3.2-1B
-
-```
-~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-cli.sh -no-cnv -p "what is the most popular cookie in the world?"
-...
-ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
-ggml-hex: Hexagon Arch version v79
-ggml-hex: allocating new session: HTP0
-ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v79.so?htp_iface_skel_handle_invoke&_modver=1.0&_dom=cdsp&_session=0 handle 0xb4000072c7955e50
-...
-load_tensors: offloading output layer to GPU
-load_tensors: offloaded 17/17 layers to GPU
-load_tensors:          CPU model buffer size =   225.49 MiB
-load_tensors:         HTP0 model buffer size =     0.26 MiB
-load_tensors:  HTP0-REPACK model buffer size =   504.00 MiB
-...
-I hope this helps you understand the world's most popular cookies! [end of text]
-...
-llama_perf_sampler_print:    sampling time =      30.08 ms /   487 runs   (    0.06 ms per token, 16191.77 tokens per second)
-llama_perf_context_print:        load time =     617.94 ms
-llama_perf_context_print: prompt eval time =      80.76 ms /    11 tokens (    7.34 ms per token,   136.21 tokens per second)
-llama_perf_context_print:        eval time =    9210.59 ms /   475 runs   (   19.39 ms per token,    51.57 tokens per second)
-llama_perf_context_print:       total time =    9454.92 ms /   486 tokens
-llama_perf_context_print:    graphs reused =        473
-llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
-llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - Host               |                  439 =   225 +     136 +      77                |
-llama_memory_breakdown_print: |   - HTP0-REPACK        |                  504 =   504 +       0 +       0                |
-```
-
-Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
-
-```
-~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-cli.sh -f surfing.txt -no-cnv
-...
-ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
-ggml-hex: Hexagon Arch version v81
-ggml-hex: allocating new session: HTP0
-ggml-hex: allocating new session: HTP1
-...
-load_tensors: offloading output layer to GPU
-load_tensors: offloaded 17/17 layers to GPU
-load_tensors:          CPU model buffer size =   143.86 MiB
-load_tensors:         HTP1 model buffer size =     0.23 MiB
-load_tensors:  HTP1-REPACK model buffer size =  1575.00 MiB
-load_tensors:         HTP0 model buffer size =     0.28 MiB
-load_tensors:  HTP0-REPACK model buffer size =  2025.00 MiB
-...
-llama_context:        CPU  output buffer size =     0.19 MiB
-llama_kv_cache:       HTP1 KV buffer size =   238.00 MiB
-llama_kv_cache:       HTP0 KV buffer size =   306.00 MiB
-llama_kv_cache: size =  544.00 MiB (  8192 cells,  16 layers,  1/1 seqs), K (q8_0):  272.00 MiB, V (q8_0):  272.00 MiB
-llama_context:       HTP0 compute buffer size =    15.00 MiB
-llama_context:       HTP1 compute buffer size =    15.00 MiB
-llama_context:        CPU compute buffer size =    24.56 MiB
-...
-llama_perf_context_print: prompt eval time =    1730.57 ms /   212 tokens (    8.16 ms per token,   122.50 tokens per second)
-llama_perf_context_print:        eval time =    5624.75 ms /   257 runs   (   21.89 ms per token,    45.69 tokens per second)
-llama_perf_context_print:       total time =    7377.33 ms /   469 tokens
-llama_perf_context_print:    graphs reused =        255
-llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
-llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - Host               |                  742 =   144 +     544 +      54                |
-llama_memory_breakdown_print: |   - HTP1-REPACK        |                 1575 =  1575 +       0 +       0                |
-llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2025 =  2025 +       0 +       0                |
-```
-
-Op test for MUL_MAT
-
-```
-~/src/llama.cpp$ HB=0 ./scripts/snapdragon/adb/run-tool.sh test-backend-ops -b HTP0 -o MUL_MAT
-...
-Backend 2/3: HTP0
-Device description: Hexagon
-Device memory: 2048 MB (2048 MB free)
-MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
-MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
-MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
-
-~/src/llama.cpp-hexagon$ M=Llama-3.2-1B-Instruct-Q4_0.gguf ./scripts/snapdragon/adb/run-bench.sh -p 128 -n 64
-...
-ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
-ggml-hex: Hexagon Arch version v79
-ggml-hex: allocating new session: HTP0
-ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v79.so?htp_iface_skel_handle_invoke&_modver=1.0&_dom=cdsp&_session=0 handle 0xb400007d4b231090
-| model          |       size | params | backend    | ngl | threads | n_batch | mmap |  test |           t/s |
-| ---------------| ---------: | -----: | ---------- | --: | ------: | ------: | ---: | ----: | ------------: |
-| llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 | pp128 | 169.42 ± 1.75 |
-| llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 |  tg64 |  51.54 ± 1.13 |
-
-build: 6a8cf8914 (6733)
-```
-
-## Environment variables
-
-- `GGML_HEXAGON_NDEV=1`
-  Controls the number of devices/sessions to allocate. The default is 1.
-  Most quantized models under 4B fit into a single session; an 8B model needs two, and a 20B model needs four.
-
-- `GGML_HEXAGON_NHVX=0`
-  Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
-
-- `GGML_HEXAGON_HOSTBUF=1`
-  Controls whether the Hexagon backend allocates host buffers. By default, all buffers except for REPACK are host buffers.
-  This option is required for testing Ops that require REPACK buffers (MUL_MAT and MUL_MAT_ID).
-
-- `GGML_HEXAGON_VERBOSE=1`
-  Enables verbose logging of Ops from the backend. Example output:
-
-  ```
-  ggml-hex: HTP0 graph-compute n_nodes 2
-  ggml-hex: HTP0 matmul : blk.27.ffn_up.weight x ffn_norm-27 -> ffn_up-27 : 3072:8192 x 3072:1 -> 8192:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x1
-  ggml-hex: HTP0 matmul : blk.27.ffn_gate.weight x ffn_norm-27 -> ffn_gate-27 : 3072:8192 x 3072:1 -> 8192:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x3
-  ggml-hex: HTP0 graph-compute n_nodes 1
-  ggml-hex: HTP0 matmul : blk.27.ffn_down.weight x ffn_gate_par-27 -> ffn_out-27 : 8192:3072 x 8192:1 -> 3072:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x0
-  ggml-hex: HTP0 get-tensor result_output : data 0x7592487000 offset 0 size 513024
-  ```
-
-- `GGML_HEXAGON_PROFILE=1`
-  Generates a host-side profile for the ggml-hexagon Ops.
-
-- `GGML_HEXAGON_OPMASK=0x0`
-  Allows enabling specific stages of the processing pipeline:
-
-  - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
-  - `0x2` Enable Dynamic Quantizer (if needed for the Op)
-  - `0x4` Enable Op Compute (MUL_MAT, etc.)
-
-  Examples:
-
-      `GGML_HEXAGON_OPMASK=0x1 llama-cli ...` - Ops are enqueued but NPU-side processing is stubbed out
-      `GGML_HEXAGON_OPMASK=0x3 llama-cli ...` - NPU performs dynamic quantization and skips the rest
-      `GGML_HEXAGON_OPMASK=0x7 llama-cli ...` - Full queuing and processing of Ops (default)

docs/backend/hexagon/developer.md

@@ -1,109 +0,0 @@
-# Hexagon backend developer details
-
-## Backend libraries
-
-The Hexagon backend consist of two parts:
-
-  - `libggml-hexagon`
-    This is the regular CPU-side GGML backend library, either shared or statically linked
-
-  - `libggml-htp-vNN`
-    This is the NPU-side (HTP stands for Hexagon Tensor Processor) shared library that contains the Op dispatcher and kernels.
-    The correct library is selected automatically at runtime based on the HW version.
-
-Here is an example of the build artifacts
-
-```
-~/src/llama.cpp$ ls -l pkg-adb/llama.cpp/lib/libggml*
-pkg-adb/llama.cpp/lib/libggml-base.so
-pkg-adb/llama.cpp/lib/libggml-cpu.so
-pkg-adb/llama.cpp/lib/libggml-hexagon.so      <<< CPU library
-pkg-adb/llama.cpp/lib/libggml-htp-v73.so      <<< HTP op/kernels for Hexagon v73
-pkg-adb/llama.cpp/lib/libggml-htp-v75.so
-pkg-adb/llama.cpp/lib/libggml-htp-v79.so
-pkg-adb/llama.cpp/lib/libggml-htp-v81.so
-```
-
-## Memory buffers
-
-Hexagon NPU backend takes advantage of the Snapdragon's unified memory model where all buffers are fully accessible by the CPU and GPU.
-The NPU does have a dedicated tightly-coupled memory called VTCM but that memory is used only for intermediate data (e.g. dynamically
-quantized tensors) or temporary data (chunks of the weight tensors fetched via DMA).
-
-Please note that currently the Hexagon backend does not implement SET/GET_ROWS Ops because there is no advantage in offloading those
-to the NPU at this point.
-
-The backend does allocates non-host buffers for the tensors with datatypes that require repacking: Q4_0, Q8_0, MXFP4.
-From the MMU perspective these buffers are still regular buffers (normal access by the CPU) they are marked as non-host simply to force
-the repacking.
-
-## Large model handling
-
-Hexagon NPU session (aka Process Domain (PD) in the Hexagon docs) is limited to a memory mapping of around 3.5GB.
-In llama.cpp/GGML the Hexagon session is mapped to a single GGML backend device (HTP0, HTP1, etc).
-
-In order to map models larger than 3.5GB we need to allocate multiple devices and split the model.
-For this we're taking advantage of the llama.cpp/GGML multi-GPU layer-splitting support.
-Each Hexagon device behaves like a GPU from the offload and model splitting perspective.
-
-Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
-
-```
-M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-cli.sh -no-cnv -f surfing.txt -n 32
-...
-LD_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
-ADSP_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
-GGML_HEXAGON_NDEV=4 ./bin/llama-cli --no-mmap -m /data/local/tmp/llama.cpp/../gguf/gpt-oss-20b-Q4_0.gguf
-      -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 --device HTP0,HTP1,HTP2,HTP3 -no-cnv -f surfing.txt
-...
-llama_model_loader: - type  f32:  289 tensors
-llama_model_loader: - type q4_0:   96 tensors
-llama_model_loader: - type q8_0:    2 tensors
-llama_model_loader: - type mxfp4:  72 tensors
-...
-load_tensors: offloaded 25/25 layers to GPU
-load_tensors:          CPU model buffer size =  1182.09 MiB
-load_tensors:         HTP1 model buffer size =     6.64 MiB
-load_tensors:  HTP1-REPACK model buffer size =  2505.94 MiB
-load_tensors:         HTP3 model buffer size =     5.55 MiB
-load_tensors:  HTP3-REPACK model buffer size =  2088.28 MiB
-load_tensors:         HTP0 model buffer size =     7.75 MiB
-load_tensors:  HTP0-REPACK model buffer size =  2923.59 MiB
-load_tensors:         HTP2 model buffer size =     6.64 MiB
-load_tensors:  HTP2-REPACK model buffer size =  2505.94 MiB
-...
-llama_context: n_ctx_per_seq (8192) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
-llama_context:        CPU  output buffer size =     0.77 MiB
-llama_kv_cache_iswa: creating non-SWA KV cache, size = 8192 cells
-llama_kv_cache:       HTP1 KV buffer size =    25.50 MiB
-llama_kv_cache:       HTP3 KV buffer size =    25.50 MiB
-llama_kv_cache:       HTP0 KV buffer size =    25.50 MiB
-llama_kv_cache:       HTP2 KV buffer size =    25.50 MiB
-llama_kv_cache: size =  102.00 MiB (  8192 cells,  12 layers,  1/1 seqs), K (q8_0):   51.00 MiB, V (q8_0):   51.00 MiB
-llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
-llama_kv_cache:       HTP1 KV buffer size =     0.80 MiB
-llama_kv_cache:       HTP3 KV buffer size =     0.53 MiB
-llama_kv_cache:       HTP0 KV buffer size =     1.06 MiB
-llama_kv_cache:       HTP2 KV buffer size =     0.80 MiB
-llama_kv_cache: size =    3.19 MiB (   256 cells,  12 layers,  1/1 seqs), K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
-llama_context:       HTP0 compute buffer size =    16.06 MiB
-llama_context:       HTP1 compute buffer size =    16.06 MiB
-llama_context:       HTP2 compute buffer size =    16.06 MiB
-llama_context:       HTP3 compute buffer size =    16.06 MiB
-llama_context:        CPU compute buffer size =    98.19 MiB
-...
-llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
-llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.39 ms per token, 18.39 tokens per second)
-llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
-llama_perf_context_print:    graphs reused =         30
-llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
-llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP2 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP3 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
-llama_memory_breakdown_print: |   - HTP1-REPACK        |                 2505 =  2505 +       0 +       0                |
-llama_memory_breakdown_print: |   - HTP3-REPACK        |                 2088 =  2088 +       0 +       0                |
-llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2923 =  2923 +       0 +       0                |
-llama_memory_breakdown_print: |   - HTP2-REPACK        |                 2505 =  2505 +       0 +       0                |
-```

docs/build-riscv64-spacemit.md

@@ -1,89 +0,0 @@
-> [!IMPORTANT]
-> This build documentation is specific only to RISC-V SpacemiT SOCs.
-
-## Build llama.cpp locally (for riscv64)
-
-1. Prepare Toolchain For RISCV
-~~~
-wget https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v1.1.2.tar.xz
-~~~
-
-2. Build
-Below is the build script: it requires utilizing RISC-V vector instructions for acceleration. Ensure the `GGML_CPU_RISCV64_SPACEMIT` compilation option is enabled. The currently supported optimization version is `RISCV64_SPACEMIT_IME1`, corresponding to the `RISCV64_SPACEMIT_IME_SPEC` compilation option. Compiler configurations are defined in the `riscv64-spacemit-linux-gnu-gcc.cmake` file. Please ensure you have installed the RISC-V compiler and set the environment variable via `export RISCV_ROOT_PATH={your_compiler_path}`.
-```bash
-
-cmake -B build \
-    -DCMAKE_BUILD_TYPE=Release \
-    -DGGML_CPU_RISCV64_SPACEMIT=ON \
-    -DLLAMA_CURL=OFF \
-    -DGGML_RVV=ON \
-    -DGGML_RV_ZFH=ON \
-    -DGGML_RV_ZICBOP=ON \
-    -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
-    -DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake \
-    -DCMAKE_INSTALL_PREFIX=build/installed
-
-cmake --build build --parallel $(nproc) --config Release
-
-pushd build
-make install
-popd
-```
-
-## Simulation
-You can use QEMU to perform emulation on non-RISC-V architectures.
-
-1. Download QEMU
-~~~
-wget https://archive.spacemit.com/spacemit-ai/qemu/jdsk-qemu-v0.0.14.tar.gz
-~~~
-
-2. Run Simulation
-After build your llama.cpp, you can run the executable file via QEMU for simulation, for example:
-~~~
-export QEMU_ROOT_PATH={your QEMU file path}
-export RISCV_ROOT_PATH_IME1={your RISC-V compiler path}
-
-${QEMU_ROOT_PATH}/bin/qemu-riscv64 -L ${RISCV_ROOT_PATH_IME1}/sysroot -cpu max,vlen=256,elen=64,vext_spec=v1.0 ${PWD}/build/bin/llama-cli -m ${PWD}/models/Qwen2.5-0.5B-Instruct-Q4_0.gguf -t 1
-~~~
-## Performance
-#### Quantization Support For Matrix
-~~~
-model name      : Spacemit(R) X60
-isa             : rv64imafdcv_zicbom_zicboz_zicntr_zicond_zicsr_zifencei_zihintpause_zihpm_zfh_zfhmin_zca_zcd_zba_zbb_zbc_zbs_zkt_zve32f_zve32x_zve64d_zve64f_zve64x_zvfh_zvfhmin_zvkt_sscofpmf_sstc_svinval_svnapot_svpbmt
-mmu             : sv39
-uarch           : spacemit,x60
-mvendorid       : 0x710
-marchid         : 0x8000000058000001
-~~~
-
-Q4_0
-|   Model    |   Size   | Params | backend | threads | test | t/s |
-| -----------| -------- | ------ | ------- | ------- | ---- |------|
-Qwen2.5 0.5B |403.20 MiB|630.17 M|   cpu   |    4    | pp512|64.12 ± 0.26|
-Qwen2.5 0.5B |403.20 MiB|630.17 M|   cpu   |    4    | tg128|10.03 ± 0.01|
-Qwen2.5 1.5B |1011.16 MiB| 1.78 B |   cpu   |    4    | pp512|24.16 ± 0.02|
-Qwen2.5 1.5B |1011.16 MiB| 1.78 B |   cpu   |    4    | tg128|3.83 ± 0.06|
-Qwen2.5 3B   | 1.86 GiB  | 3.40 B |   cpu   |    4    | pp512|12.08 ± 0.02|
-Qwen2.5 3B   | 1.86 GiB  | 3.40 B |   cpu   |    4    | tg128|2.23 ± 0.02|
-
-Q4_1
-|   Model    |   Size   | Params | backend | threads | test | t/s |
-| -----------| -------- | ------ | ------- | ------- | ---- |------|
-Qwen2.5 0.5B |351.50 MiB|494.03 M|   cpu   |    4    | pp512|62.07 ± 0.12|
-Qwen2.5 0.5B |351.50 MiB|494.03 M|   cpu   |    4    | tg128|9.91 ± 0.01|
-Qwen2.5 1.5B |964.06 MiB| 1.54 B |   cpu   |    4    | pp512|22.95 ± 0.25|
-Qwen2.5 1.5B |964.06 MiB| 1.54 B |   cpu   |    4    | tg128|4.01 ± 0.15|
-Qwen2.5 3B   | 1.85 GiB | 3.09 B |   cpu   |    4    | pp512|11.55 ± 0.16|
-Qwen2.5 3B   | 1.85 GiB | 3.09 B |   cpu   |    4    | tg128|2.25 ± 0.04|
-
-
-Q4_K
-|   Model    |   Size   | Params | backend | threads | test | t/s |
-| -----------| -------- | ------ | ------- | ------- | ---- |------|
-Qwen2.5 0.5B |462.96 MiB|630.17 M|   cpu   |    4    | pp512|9.29 ± 0.05|
-Qwen2.5 0.5B |462.96 MiB|630.17 M|   cpu   |    4    | tg128|5.67 ± 0.04|
-Qwen2.5 1.5B | 1.04 GiB | 1.78 B |   cpu   |    4    | pp512|10.38 ± 0.10|
-Qwen2.5 1.5B | 1.04 GiB | 1.78 B |   cpu   |    4    | tg128|3.17 ± 0.08|
-Qwen2.5 3B   | 1.95 GiB | 3.40 B |   cpu   |    4    | pp512|4.23 ± 0.04|
-Qwen2.5 3B   | 1.95 GiB | 3.40 B |   cpu   |    4    | tg128|1.73 ± 0.00|

docs/build.md

@@ -261,12 +261,10 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-      cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
 
-  Note: `GPU_TARGETS` is optional, omitting it will build the code for all GPUs in the current system.
-
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
   The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager.
@@ -284,17 +282,17 @@ You can download it from your Linux distro's package manager or from here: [ROCm
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
   HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-      cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build -- -j 16
   ```
 
 - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
   ```bash
   set PATH=%HIP_PATH%\bin;%PATH%
-  cmake -S . -B build -G Ninja -DGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
   cmake --build build
   ```
-  If necessary, adapt `GPU_TARGETS` to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+  Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
   Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
 
 

docs/ops.md

@@ -22,7 +22,6 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
@@ -32,7 +31,7 @@ Legend:
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -42,7 +41,6 @@ Legend:
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -53,7 +51,7 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
@@ -67,24 +65,23 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -95,22 +92,19 @@ Legend:
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
-|                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

docs/ops/CPU.csv

@@ -59,14 +59,6 @@
 "CPU","EXP","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
-"CPU","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
-"CPU","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
-"CPU","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
-"CPU","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","ABS","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
 "CPU","ABS","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","SGN","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
@@ -127,14 +119,6 @@
 "CPU","EXP","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
-"CPU","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
-"CPU","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
-"CPU","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
-"CPU","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
-"CPU","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","1","yes","CPU"

docs/ops/Vulkan.csv

@@ -3263,27 +3263,27 @@
 "Vulkan0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=0","support","1","yes","Vulkan"
 "Vulkan0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=1","support","1","yes","Vulkan"
 "Vulkan0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=1,n_seqs=1","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=1","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"

examples/embedding/README.md

@@ -38,7 +38,6 @@ The above command will output space-separated float values.
 |            | multiple embeddings          | $[[x_1,...,x_n],[x_1,...,x_n],...,[x_1,...,x_n]]$
 | 'json'     | openai style                 |
 | 'json+'    | add cosine similarity matrix |
-| 'raw'      | plain text output            |
 
 ### --embd-separator $"string"$
 | $"string"$   | |

examples/embedding/embedding.cpp

@@ -70,29 +70,6 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     }
 }
 
-// plain, pipe-friendly output: one embedding per line
-static void print_raw_embeddings(const float * emb,
-                                 int n_embd_count,
-                                 int n_embd,
-                                 const llama_model * model,
-                                 enum llama_pooling_type pooling_type,
-                                 int embd_normalize) {
-    const uint32_t n_cls_out = llama_model_n_cls_out(model);
-    const bool is_rank = (pooling_type == LLAMA_POOLING_TYPE_RANK);
-    const int cols = is_rank ? std::min<int>(n_embd, (int) n_cls_out) : n_embd;
-
-    for (int j = 0; j < n_embd_count; ++j) {
-        for (int i = 0; i < cols; ++i) {
-            if (embd_normalize == 0) {
-                LOG("%1.0f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
-            } else {
-                LOG("%1.7f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
-            }
-        }
-        LOG("\n");
-    }
-}
-
 int main(int argc, char ** argv) {
     common_params params;
 
@@ -395,8 +372,6 @@ int main(int argc, char ** argv) {
         }
 
         if (notArray) LOG("\n}\n");
-    } else if (params.embd_out == "raw") {
-        print_raw_embeddings(emb, n_embd_count, n_embd, model, pooling_type, params.embd_normalize);
     }
 
     LOG("\n");

examples/json_schema_to_grammar.py

@@ -371,17 +371,8 @@ class SchemaConverter:
                         raise ValueError(f'Unsupported ref {ref}')
 
                     for sel in ref.split('#')[-1].split('/')[1:]:
-                        assert target is not None, f'Error resolving ref {ref}: {sel} not in {target}'
-                        if isinstance(target, list):
-                            try:
-                                sel_index = int(sel)
-                            except ValueError:
-                                raise ValueError(f'Error resolving ref {ref}: {sel} not in {target}')
-                            assert 0 <= sel_index < len(target), f'Error resolving ref {ref}: {sel} not in {target}'
-                            target = target[sel_index]
-                        else:
-                            assert sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
-                            target = target[sel]
+                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                        target = target[sel]
 
                     self._refs[ref] = target
                 else:
@@ -556,8 +547,7 @@ class SchemaConverter:
 
 
     def _resolve_ref(self, ref):
-        ref_fragment = ref.split('#')[-1]
-        ref_name = 'ref' + re.sub(r'[^a-zA-Z0-9-]+', '-', ref_fragment)
+        ref_name = ref.split('/')[-1]
         if ref_name not in self._rules and ref not in self._refs_being_resolved:
             self._refs_being_resolved.add(ref)
             resolved = self._refs[ref]

examples/model-conversion/Makefile

@@ -116,36 +116,17 @@ embedding-convert-model:
 	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
 	./scripts/embedding/convert-model.sh
 
-embedding-convert-model-st:
-	$(call validate_embedding_model_path,embedding-convert-model-st)
-	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
-	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
-	./scripts/embedding/convert-model.sh -st
-
 embedding-run-original-model:
 	$(call validate_embedding_model_path,embedding-run-original-model)
 	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
-	USE_SENTENCE_TRANSFORMERS="$(USE_SENTENCE_TRANSFORMERS)" \
 	./scripts/embedding/run-original-model.py \
-	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)") \
-	$(if $(USE_SENTENCE_TRANSFORMERS),--use-sentence-transformers)
-
-embedding-run-original-model-st: USE_SENTENCE_TRANSFORMERS=1
-embedding-run-original-model-st: embedding-run-original-model
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
 
 embedding-run-converted-model:
 	@./scripts/embedding/run-converted-model.sh $(CONVERTED_EMBEDDING_MODEL) \
-	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)") \
-	$(if $(USE_POOLING),--pooling)
-
-embedding-run-converted-model-st: USE_POOLING=1
-embedding-run-converted-model-st: embedding-run-converted-model
-
-embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
-	@./scripts/embedding/compare-embeddings-logits.sh \
 	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
 
-embedding-verify-logits-st: embedding-run-original-model-st embedding-run-converted-model-st
+embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
 	@./scripts/embedding/compare-embeddings-logits.sh \
 	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
 

examples/model-conversion/README.md

@@ -189,23 +189,6 @@ This command will save two files to the `data` directory, one is a binary
 file containing logits which will be used for comparison with the converted
 model, and the other is a text file which allows for manual visual inspection.
 
-#### Using SentenceTransformer with numbered layers
-For models that have numbered SentenceTransformer layers (01_Pooling, 02_Dense,
-03_Dense, 04_Normalize), use the `-st` targets to apply all these layers:
-
-```console
-# Run original model with SentenceTransformer (applies all numbered layers)
-(venv) $ make embedding-run-original-model-st
-
-# Run converted model with pooling enabled
-(venv) $ make embedding-run-converted-model-st
-```
-
-This will use the SentenceTransformer library to load and run the model, which
-automatically applies all the numbered layers in the correct order. This is
-particularly useful when comparing with models that should include these
-additional transformation layers beyond just the base model output.
-
 ### Model conversion
 After updates have been made to [gguf-py](../../gguf-py) to add support for the
 new model the model can be converted to GGUF format using the following command:
@@ -225,13 +208,6 @@ was done manually in the previous steps) and compare the logits:
 (venv) $ make embedding-verify-logits
 ```
 
-For models with SentenceTransformer layers, use the `-st` verification target:
-```console
-(venv) $ make embedding-verify-logits-st
-```
-This convenience target automatically runs both the original model with SentenceTransformer
-and the converted model with pooling enabled, then compares the results.
-
 ### llama-server verification
 To verify that the converted model works with llama-server, the following
 command can be used:

examples/model-conversion/logits.cpp

@@ -1,7 +1,4 @@
 #include "llama.h"
-#include "common.h"
-
-
 #include <cstdio>
 #include <cstring>
 #include <string>
@@ -11,10 +8,7 @@
 
 static void print_usage(int, char ** argv) {
     printf("\nexample usage:\n");
-    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [-pooling] [-embd-norm <norm>] [prompt]\n", argv[0]);
-    printf("\n");
-    printf("  -embd-norm: normalization type for pooled embeddings (default: 2)\n");
-    printf("              -1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm\n");
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
     printf("\n");
 }
 
@@ -23,8 +17,6 @@ int main(int argc, char ** argv) {
     std::string prompt = "Hello, my name is";
     int ngl = 0;
     bool embedding_mode = false;
-    bool pooling_enabled = false;
-    int32_t embd_norm = 2;  // (-1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm)
 
     {
         int i = 1;
@@ -49,13 +41,9 @@ int main(int argc, char ** argv) {
                     return 1;
                 }
             } else if (strcmp(argv[i], "-embd-mode") == 0) {
-                embedding_mode = true;
-            } else if (strcmp(argv[i], "-pooling") == 0) {
-                pooling_enabled = true;
-            } else if (strcmp(argv[i], "-embd-norm") == 0) {
                 if (i + 1 < argc) {
                     try {
-                        embd_norm = std::stoi(argv[++i]);
+                        embedding_mode = true;
                     } catch (...) {
                         print_usage(argc, argv);
                         return 1;
@@ -124,7 +112,7 @@ int main(int argc, char ** argv) {
     ctx_params.no_perf = false;
     if (embedding_mode) {
         ctx_params.embeddings = true;
-        ctx_params.pooling_type = pooling_enabled ? LLAMA_POOLING_TYPE_MEAN : LLAMA_POOLING_TYPE_NONE;
+        ctx_params.pooling_type = LLAMA_POOLING_TYPE_NONE;
         ctx_params.n_ubatch = ctx_params.n_batch;
     }
 
@@ -155,27 +143,17 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    float * data_ptr;
-    int data_size;
+    float * logits;
+    int n_logits;
     const char * type;
-    std::vector<float> embd_out;
 
     if (embedding_mode) {
-        const int n_embd = llama_model_n_embd(model);
-        const int n_embd_count = pooling_enabled ? 1 : batch.n_tokens;
-        const int n_embeddings = n_embd * n_embd_count;
-        float * embeddings;
+        logits = llama_get_embeddings(ctx);
+        n_logits = llama_model_n_embd(model) * batch.n_tokens;
         type = "-embeddings";
 
-        if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_NONE) {
-            embeddings = llama_get_embeddings_seq(ctx, 0);
-            embd_out.resize(n_embeddings);
-            printf("Normalizing embeddings using norm: %d\n", embd_norm);
-            common_embd_normalize(embeddings, embd_out.data(), n_embeddings, embd_norm);
-            embeddings = embd_out.data();
-        } else {
-            embeddings = llama_get_embeddings(ctx);
-        }
+        const int n_embd = llama_model_n_embd(model);
+        const int n_embd_count = batch.n_tokens;
 
         printf("Embedding dimension: %d\n", n_embd);
         printf("\n");
@@ -186,7 +164,7 @@ int main(int argc, char ** argv) {
 
             // Print first 3 values
             for (int i = 0; i < 3 && i < n_embd; i++) {
-                printf("%9.6f ", embeddings[j * n_embd + i]);
+                printf("%9.6f ", logits[j * n_embd + i]);
             }
 
             printf(" ... ");
@@ -194,7 +172,7 @@ int main(int argc, char ** argv) {
             // Print last 3 values
             for (int i = n_embd - 3; i < n_embd; i++) {
                 if (i >= 0) {
-                    printf("%9.6f ", embeddings[j * n_embd + i]);
+                    printf("%9.6f ", logits[j * n_embd + i]);
                 }
             }
 
@@ -202,33 +180,27 @@ int main(int argc, char ** argv) {
         }
         printf("\n");
 
-        printf("Embeddings size: %d\n", n_embeddings);
-
-        data_ptr = embeddings;
-        data_size = n_embeddings;
+        printf("Embeddings size: %d\n", n_logits);
     } else {
-        float * logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
-        const int n_logits = llama_vocab_n_tokens(vocab);
+        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        n_logits = llama_vocab_n_tokens(vocab);
         type = "";
         printf("Vocab size: %d\n", n_logits);
-
-        data_ptr = logits;
-        data_size = n_logits;
     }
 
     std::filesystem::create_directory("data");
 
-    // Save data to binary file
+    // Save logits to binary file
     char bin_filename[512];
     snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
-    printf("Saving data to %s\n", bin_filename);
+    printf("Saving logits to %s\n", bin_filename);
 
     FILE * f = fopen(bin_filename, "wb");
     if (f == NULL) {
         fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
         return 1;
     }
-    fwrite(data_ptr, sizeof(float), data_size, f);
+    fwrite(logits, sizeof(float), n_logits, f);
     fclose(f);
 
     // Also save as text for debugging
@@ -239,27 +211,27 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
         return 1;
     }
-    for (int i = 0; i < data_size; i++) {
-        fprintf(f, "%d: %.6f\n", i, data_ptr[i]);
+    for (int i = 0; i < n_logits; i++) {
+        fprintf(f, "%d: %.6f\n", i, logits[i]);
     }
     fclose(f);
 
     if (!embedding_mode) {
         printf("First 10 logits: ");
-        for (int i = 0; i < 10 && i < data_size; i++) {
-            printf("%.6f ", data_ptr[i]);
+        for (int i = 0; i < 10 && i < n_logits; i++) {
+            printf("%.6f ", logits[i]);
         }
         printf("\n");
 
         printf("Last 10 logits: ");
-        for (int i = data_size - 10; i < data_size; i++) {
-            if (i >= 0) printf("%.6f ", data_ptr[i]);
+        for (int i = n_logits - 10; i < n_logits; i++) {
+            if (i >= 0) printf("%.6f ", logits[i]);
         }
         printf("\n\n");
     }
 
-    printf("Data saved to %s\n", bin_filename);
-    printf("Data saved to %s\n", txt_filename);
+    printf("Logits saved to %s\n", bin_filename);
+    printf("Logits saved to %s\n", txt_filename);
 
     llama_free(ctx);
     llama_model_free(model);

examples/model-conversion/requirements.txt

@@ -4,4 +4,3 @@ torchvision
 transformers
 huggingface-hub
 accelerate
-sentence-transformers

examples/model-conversion/scripts/causal/run-org-model.py

@@ -138,7 +138,7 @@ if model_path is None:
         "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
     )
 
-config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+config = AutoConfig.from_pretrained(model_path)
 
 print("Model type:       ", config.model_type)
 print("Vocab size:       ", config.vocab_size)
@@ -148,8 +148,8 @@ print("BOS token id:     ", config.bos_token_id)
 print("EOS token id:     ", config.eos_token_id)
 
 print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+config = AutoConfig.from_pretrained(model_path)
 
 if unreleased_model_name:
     model_name_lower = unreleased_model_name.lower()
@@ -171,7 +171,7 @@ if unreleased_model_name:
         exit(1)
 else:
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True
+        model_path, device_map="auto", offload_folder="offload"
     )
 
 for name, module in model.named_modules():

examples/model-conversion/scripts/embedding/convert-model.sh

@@ -2,21 +2,6 @@
 
 set -e
 
-# Parse command line arguments
-SENTENCE_TRANSFORMERS=""
-while [[ $# -gt 0 ]]; do
-    case $1 in
-        -st|--sentence-transformers)
-            SENTENCE_TRANSFORMERS="--sentence-transformers-dense-modules"
-            shift
-            ;;
-        *)
-            echo "Unknown option: $1"
-            exit 1
-            ;;
-    esac
-done
-
 MODEL_NAME="${MODEL_NAME:-$(basename "$EMBEDDING_MODEL_PATH")}"
 OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
 TYPE="${OUTTYPE:-f16}"
@@ -30,8 +15,7 @@ echo "Converted model path:: ${CONVERTED_MODEL}"
 python ../../convert_hf_to_gguf.py --verbose \
     ${EMBEDDING_MODEL_PATH} \
     --outfile ${CONVERTED_MODEL} \
-    --outtype ${TYPE} \
-    ${SENTENCE_TRANSFORMERS}
+    --outtype ${TYPE}
 
 echo ""
 echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -5,7 +5,6 @@ set -e
 # Parse command line arguments
 CONVERTED_MODEL=""
 PROMPTS_FILE=""
-USE_POOLING=""
 
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -13,10 +12,6 @@ while [[ $# -gt 0 ]]; do
             PROMPTS_FILE="$2"
             shift 2
             ;;
-        --pooling)
-            USE_POOLING="1"
-            shift
-            ;;
         *)
             if [ -z "$CONVERTED_MODEL" ]; then
                 CONVERTED_MODEL="$1"
@@ -52,8 +47,4 @@ echo $CONVERTED_MODEL
 
 cmake --build ../../build --target llama-logits -j8
 # TODO: update logits.cpp to accept a --file/-f option for the prompt
-if [ -n "$USE_POOLING" ]; then
-    ../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode -pooling "$PROMPT"
-else
-    ../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"
-fi
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -14,8 +14,6 @@ unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
 parser = argparse.ArgumentParser(description='Process model with specified path')
 parser.add_argument('--model-path', '-m', help='Path to the model')
 parser.add_argument('--prompts-file', '-p', help='Path to file containing prompts (one per line)')
-parser.add_argument('--use-sentence-transformers', action='store_true',
-                    help='Use SentenceTransformer to apply all numbered layers (01_Pooling, 02_Dense, 03_Dense, 04_Normalize)')
 args = parser.parse_args()
 
 def read_prompt_from_file(file_path):
@@ -33,52 +31,41 @@ model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
 if model_path is None:
     parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
 
-# Determine if we should use SentenceTransformer
-use_sentence_transformers = args.use_sentence_transformers or os.environ.get('USE_SENTENCE_TRANSFORMERS', '').lower() in ('1', 'true', 'yes')
+tokenizer = AutoTokenizer.from_pretrained(model_path)
 
-if use_sentence_transformers:
-    from sentence_transformers import SentenceTransformer
-    print("Using SentenceTransformer to apply all numbered layers")
-    model = SentenceTransformer(model_path)
-    tokenizer = model.tokenizer
-    config = model[0].auto_model.config  # type: ignore
+config = AutoConfig.from_pretrained(model_path)
+
+# This can be used to override the sliding window size for manual testing. This
+# can be useful to verify the sliding window attention mask in the original model
+# and compare it with the converted .gguf model.
+if hasattr(config, 'sliding_window'):
+    original_sliding_window = config.sliding_window
+    #original_sliding_window = 6
+    print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
+
+print(f"Using unreleased model: {unreleased_model_name}")
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}Model"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path, config=config)
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
 else:
-    tokenizer = AutoTokenizer.from_pretrained(model_path)
-
-    config = AutoConfig.from_pretrained(model_path)
-
-    # This can be used to override the sliding window size for manual testing. This
-    # can be useful to verify the sliding window attention mask in the original model
-    # and compare it with the converted .gguf model.
-    if hasattr(config, 'sliding_window'):
-        original_sliding_window = config.sliding_window
-        #original_sliding_window = 6
-        print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
-
-    print(f"Using unreleased model: {unreleased_model_name}")
-    if unreleased_model_name:
-        model_name_lower = unreleased_model_name.lower()
-        unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
-        class_name = f"{unreleased_model_name}Model"
-        print(f"Importing unreleased model module: {unreleased_module_path}")
-
-        try:
-            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
-            model = model_class.from_pretrained(model_path, config=config)
-        except (ImportError, AttributeError) as e:
-            print(f"Failed to import or load model: {e}")
-            exit(1)
-    else:
-        model = AutoModel.from_pretrained(model_path, config=config)
-    print(f"Model class: {type(model)}")
-    print(f"Model file: {type(model).__module__}")
+    model = AutoModel.from_pretrained(model_path, config=config)
+print(f"Model class: {type(model)}")
+print(f"Model file: {type(model).__module__}")
 
 # Verify the model is using the correct sliding window
-if not use_sentence_transformers:
-    if hasattr(model.config, 'sliding_window'):  # type: ignore
-        print(f"Model's sliding_window: {model.config.sliding_window}")  # type: ignore
-    else:
-        print("Model config does not have sliding_window attribute")
+if hasattr(model.config, 'sliding_window'):
+    print(f"Model's sliding_window: {model.config.sliding_window}")
+else:
+    print("Model config does not have sliding_window attribute")
 
 model_name = os.path.basename(model_path)
 
@@ -88,56 +75,34 @@ if args.prompts_file:
 else:
     texts = ["Hello world today"]
 
+encoded = tokenizer(
+    texts,
+    padding=True,
+    truncation=True,
+    return_tensors="pt"
+)
+
+tokens = encoded['input_ids'][0]
+token_strings = tokenizer.convert_ids_to_tokens(tokens)
+for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+    print(f"{token_id:6d} -> '{token_str}'")
+
 with torch.no_grad():
-    if use_sentence_transformers:
-        embeddings = model.encode(texts, convert_to_numpy=True)
-        all_embeddings = embeddings  # Shape: [batch_size, hidden_size]
+    outputs = model(**encoded)
+    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
 
-        encoded = tokenizer(
-            texts,
-            padding=True,
-            truncation=True,
-            return_tensors="pt"
-        )
-        tokens = encoded['input_ids'][0]
-        token_strings = tokenizer.convert_ids_to_tokens(tokens)
-        for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
-            print(f"{token_id:6d} -> '{token_str}'")
+    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
+    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
 
-        print(f"Embeddings shape (after all SentenceTransformer layers): {all_embeddings.shape}")
-        print(f"Embedding dimension: {all_embeddings.shape[1] if len(all_embeddings.shape) > 1 else all_embeddings.shape[0]}")  # type: ignore
-    else:
-        # Standard approach: use base model output only
-        encoded = tokenizer(
-            texts,
-            padding=True,
-            truncation=True,
-            return_tensors="pt"
-        )
+    print(f"Hidden states shape: {hidden_states.shape}")
+    print(f"All embeddings shape: {all_embeddings.shape}")
+    print(f"Embedding dimension: {all_embeddings.shape[1]}")
 
-        tokens = encoded['input_ids'][0]
-        token_strings = tokenizer.convert_ids_to_tokens(tokens)
-        for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
-            print(f"{token_id:6d} -> '{token_str}'")
+    # Print embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
+    n_embd = all_embeddings.shape[1]
+    n_embd_count = all_embeddings.shape[0]
 
-        outputs = model(**encoded)
-        hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
-
-        all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
-
-        print(f"Hidden states shape: {hidden_states.shape}")
-        print(f"All embeddings shape: {all_embeddings.shape}")
-        print(f"Embedding dimension: {all_embeddings.shape[1]}")
-
-    if len(all_embeddings.shape) == 1:
-        n_embd = all_embeddings.shape[0]  # type: ignore
-        n_embd_count = 1
-        all_embeddings = all_embeddings.reshape(1, -1)
-    else:
-        n_embd = all_embeddings.shape[1]  # type: ignore
-        n_embd_count = all_embeddings.shape[0]  # type: ignore
-
-    print()
+    print()  # Empty line to match C++ output
 
     for j in range(n_embd_count):
         embedding = all_embeddings[j]
@@ -155,23 +120,29 @@ with torch.no_grad():
 
         print()  # New line
 
-    print()
+    print()  # Final empty line to match C++ output
 
     data_dir = Path("data")
     data_dir.mkdir(exist_ok=True)
     bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
     txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
 
+    # Save all embeddings flattened (matching what embedding.cpp would save if it did)
     flattened_embeddings = all_embeddings.flatten()
     flattened_embeddings.astype(np.float32).tofile(bin_filename)
 
     with open(txt_filename, "w") as f:
-        idx = 0
+        f.write(f"# Model class: {model_name}\n")
+        f.write(f"# Tokens: {token_strings}\n")
+        f.write(f"# Shape: {all_embeddings.shape}\n")
+        f.write(f"# n_embd_count: {n_embd_count}, n_embd: {n_embd}\n\n")
+
         for j in range(n_embd_count):
-            for value in all_embeddings[j]:
-                f.write(f"{idx}: {value:.6f}\n")
-                idx += 1
-    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} embeddings × {n_embd} dimensions)")
+            f.write(f"# Token {j} ({token_strings[j]}):\n")
+            for i, value in enumerate(all_embeddings[j]):
+                f.write(f"{j}_{i}: {value:.6f}\n")
+            f.write("\n")
+    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} tokens × {n_embd} dimensions)")
     print("")
     print(f"Saved bin embeddings to: {bin_filename}")
     print(f"Saved txt embeddings to: {txt_filename}")

examples/model-conversion/scripts/utils/semantic_check.py

@@ -35,11 +35,7 @@ def cosine_similarity(a, b=None):
 
 def load_embeddings_from_file(filename, n_tokens, n_embd):
     embeddings = np.fromfile(filename, dtype=np.float32)
-    # Check if this is pooled (single embedding) or per-token embeddings
-    if len(embeddings) == n_embd:
-        return embeddings.reshape(1, n_embd)
-    else:
-        return embeddings.reshape(n_tokens, n_embd)
+    return embeddings.reshape(n_tokens, n_embd)
 
 def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
     np.set_printoptions(suppress=True, precision=6)
@@ -52,83 +48,58 @@ def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
     print(f"Embeddings shape: Python {python_emb.shape}, llama.cpp {cpp_emb.shape}")
 
     n_tokens = len(tokens)
-    is_pooled = python_emb.shape[0] == 1
 
-    if is_pooled:
-        print(f"\n[Pooled Embeddings Mode - comparing single sentence embeddings]")
-
-        # 1. Direct embedding comparison for pooled embeddings
-        print(f"\n1. Raw Embedding Magnitude Comparison:")
-        py_mag = np.linalg.norm(python_emb[0])
-        cpp_mag = np.linalg.norm(cpp_emb[0])
+    # 1. Direct embedding comparison
+    print(f"\n1. Raw Embedding Magnitude Comparison:")
+    # Check if the distance of each token embedding from the origin and compare
+    # if the vectors are on the same "sphere". This does not tell us about
+    # direction (meaning of the token embedding), just magnitude.
+    for i in range(n_tokens):
+        py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
+        cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
         ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
-        print(f"   Pooled embedding: Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+        print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
 
-        # 2. Cross-model similarity for pooled embeddings
-        print(f"\n2. Cross-Model Pooled Embedding Similarity:")
-        sim = cosine_similarity([python_emb[0]], [cpp_emb[0]])[0][0]
-        print(f"   Cosine similarity: {sim:.6f}")
+    # 2. Cosine similarity between tokens within each model
+    # Here we check the direction of token embeddings to see if the have the
+    # same meaning (similarity). This is done by calculating cosine similarity
+    # of a pair of token embeddings within each model.
+    print(f"\n2. Within-Model Token Similarities:")
+    print("   Python model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
 
-        return {
-            'cross_model_similarities': [sim],
-            'similarity_matrix_diff': np.array([[0.0]]),
-            'max_diff': 0.0,
-            'mean_diff': 0.0,
-            'rms_diff': 0.0
-        }
-    else:
-        # Original per-token comparison logic
-        # 1. Direct embedding comparison
-        print(f"\n1. Raw Embedding Magnitude Comparison:")
-        # Check if the distance of each token embedding from the origin and compare
-        # if the vectors are on the same "sphere". This does not tell us about
-        # direction (meaning of the token embedding), just magnitude.
-        for i in range(n_tokens):
-            py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
-            cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
-            ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
-            print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+    print("   llama.cpp model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
 
-        # 2. Cosine similarity between tokens within each model
-        # Here we check the direction of token embeddings to see if the have the
-        # same meaning (similarity). This is done by calculating cosine similarity
-        # of a pair of token embeddings within each model.
-        print(f"\n2. Within-Model Token Similarities:")
-        print("   Python model:")
-        for i in range(n_tokens):
-            for j in range(i+1, n_tokens):
-                sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
-                print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+    # 3. Cross-model similarity (same token position)
+    print(f"\n3. Cross-Model Same-Token Similarities:")
+    for i in range(n_tokens):
+        sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
+        print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
 
-        print("   llama.cpp model:")
-        for i in range(n_tokens):
-            for j in range(i+1, n_tokens):
-                sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
-                print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+    # 4. Similarity matrix comparison
+    print(f"\n4. Similarity Matrix Differences:")
+    py_sim_matrix = cosine_similarity(python_emb)
+    cpp_sim_matrix = cosine_similarity(cpp_emb)
+    diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
 
-        # 3. Cross-model similarity (same token position)
-        print(f"\n3. Cross-Model Same-Token Similarities:")
-        for i in range(n_tokens):
-            sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
-            print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
+    print(f"   Max difference: {np.max(diff_matrix):.4f}")
+    print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
+    print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
 
-        # 4. Similarity matrix comparison
-        print(f"\n4. Similarity Matrix Differences:")
-        py_sim_matrix = cosine_similarity(python_emb)
-        cpp_sim_matrix = cosine_similarity(cpp_emb)
-        diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
-
-        print(f"   Max difference: {np.max(diff_matrix):.4f}")
-        print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
-        print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
-
-        return {
-            'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
-            'similarity_matrix_diff': diff_matrix,
-            'max_diff': np.max(diff_matrix),
-            'mean_diff': np.mean(diff_matrix),
-            'rms_diff': np.sqrt(np.mean(diff_matrix**2))
-        }
+    return {
+        'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
+        'similarity_matrix_diff': diff_matrix,
+        'max_diff': np.max(diff_matrix),
+        'mean_diff': np.mean(diff_matrix),
+        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
+    }
 
 def read_prompt_from_file(file_path):
     try:

ggml/CMakeLists.txt

@@ -4,7 +4,8 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 4)
+set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION_DEV "-dev")  # "-dev" for development, "" for releases
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -25,8 +26,8 @@ if(GIT_EXE)
     )
 endif()
 
-# Build the version string with optional dirty flag
-set(GGML_VERSION "${GGML_VERSION_BASE}")
+# Build the version string with optional -dev suffix and dirty flag
+set(GGML_VERSION "${GGML_VERSION_BASE}${GGML_VERSION_DEV}")
 if(GGML_GIT_DIRTY AND NOT GGML_GIT_DIRTY EQUAL 0)
     set(GGML_VERSION "${GGML_VERSION}-dirty")
 endif()
@@ -209,6 +210,7 @@ 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_HIP_MMQ_MFMA                    "ggml: enable MFMA MMA for CDNA in MMQ"           ON)
 option(GGML_HIP_EXPORT_METRICS              "ggml: enable kernel perf metrics output"         OFF)
 option(GGML_MUSA_GRAPHS                     "ggml: use MUSA graph, experimental, unstable"    OFF)
@@ -222,9 +224,6 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
 option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
-option(GGML_WEBGPU_CPU_PROFILE              "ggml: enable WebGPU profiling (CPU)"             OFF)
-option(GGML_WEBGPU_GPU_PROFILE              "ggml: enable WebGPU profiling (GPU)"             OFF)
-
 option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -251,8 +250,6 @@ option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adr
 set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
                                             "gmml: OpenCL API version to target")
 
-option(GGML_HEXAGON                         "ggml: enable Hexagon backend"                    OFF)
-
 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
 

ggml/include/ggml-backend.h

@@ -215,8 +215,6 @@ extern "C" {
     // Backend registry
     //
 
-    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
-
     GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
 
     // Backend (reg) enumeration

ggml/include/ggml-hexagon.h

@@ -1,19 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-// backend API
-GGML_BACKEND_API ggml_backend_t ggml_backend_hexagon_init(void);
-
-GGML_BACKEND_API bool ggml_backend_is_hexagon(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_hexagon_reg(void);
-
-#ifdef  __cplusplus
-}
-#endif

ggml/include/ggml-rpc.h

@@ -7,24 +7,26 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    3
+#define RPC_PROTO_MAJOR_VERSION    2
 #define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
 GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, uint32_t device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total);
+GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_BACKEND_API void ggml_backend_rpc_start_server(const char * endpoint, const char * cache_dir,
-                                                    size_t n_threads, size_t n_devices, ggml_backend_dev_t * devices);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
+                                                    const char * cache_dir,
+                                                    size_t free_mem, size_t total_mem);
 
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_add_server(const char * endpoint);
+
+GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml.h

@@ -237,12 +237,9 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-// TODO: convert to enum https://github.com/ggml-org/llama.cpp/pull/16187#discussion_r2388538726
-#define GGML_ROPE_TYPE_NORMAL 0
 #define GGML_ROPE_TYPE_NEOX   2
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
-#define GGML_ROPE_TYPE_IMROPE 40 // binary: 101000
 
 #define GGML_MROPE_SECTIONS   4
 
@@ -577,11 +574,6 @@ extern "C" {
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
         GGML_UNARY_OP_GELU_ERF,
-        GGML_UNARY_OP_XIELU,
-        GGML_UNARY_OP_FLOOR,
-        GGML_UNARY_OP_CEIL,
-        GGML_UNARY_OP_ROUND,
-        GGML_UNARY_OP_TRUNC,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -1156,58 +1148,6 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    GGML_API struct ggml_tensor * ggml_floor(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_floor_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_ceil(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_ceil_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_round(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_round_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-     /**
-     * Truncates the fractional part of each element in the tensor (towards zero).
-     * For example: trunc(3.7) = 3.0, trunc(-2.9) = -2.0
-     * Similar to std::trunc in C/C++.
-     */
-
-    GGML_API struct ggml_tensor * ggml_trunc(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_trunc_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-
-
-    // xIELU activation function
-    // x = x * (c_a(alpha_n) + c_b(alpha_p, beta) * sigmoid(beta * x)) + eps * (x > 0)
-    // where c_a = softplus and c_b(a, b) = softplus(a) + b are constraining functions
-    // that constrain the positive and negative source alpha values respectively
-    GGML_API struct ggml_tensor * ggml_xielu(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            float alpha_n,
-            float alpha_p,
-            float beta,
-            float eps);
-
     // gated linear unit ops
     // A: n columns, r rows,
     // result is n / 2 columns, r rows,
@@ -1675,13 +1615,6 @@ extern "C" {
             float                 scale,
             float                 max_bias);
 
-    GGML_API struct ggml_tensor * ggml_soft_max_ext_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * mask,
-            float                 scale,
-            float                 max_bias);
-
     GGML_API void ggml_soft_max_add_sinks(
             struct ggml_tensor * a,
             struct ggml_tensor * sinks);

ggml/src/CMakeLists.txt

@@ -145,9 +145,6 @@ endif()
 # which was introduced in POSIX.1-2008, forcing us to go higher
 if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
     add_compile_definitions(_XOPEN_SOURCE=700)
-elseif (CMAKE_SYSTEM_NAME MATCHES "AIX")
-    # Don't define _XOPEN_SOURCE.  We need _ALL_SOURCE, which is the default,
-    # in order to define _SC_PHYS_PAGES.
 else()
     add_compile_definitions(_XOPEN_SOURCE=600)
 endif()
@@ -307,10 +304,6 @@ function(ggml_add_cpu_backend_variant tag_name)
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
-        foreach (feat ${ARGN})
-            set(GGML_INTERNAL_${feat} ON)
-        endforeach()
     endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -375,14 +368,6 @@ if (GGML_CPU_ALL_VARIANTS)
         else()
             message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
-        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            ggml_add_cpu_backend_variant(s390x_z15  Z15 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z16  Z16 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z17  Z17 VXE)
-        else()
-            message(FATAL_ERROR "Unsupported s390x 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()
@@ -402,7 +387,6 @@ ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
 ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
-ggml_add_backend(Hexagon)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-alloc.c

@@ -226,23 +226,16 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
     }
 
     if (best_fit_block == -1) {
-        // no suitable block found, try the last block (this may grow a chunks size)
-        int64_t best_reuse = INT64_MIN;
+        // no suitable block found, try the last block (this will grow a chunks size)
         for (int c = 0; c < alloc->n_chunks; ++c) {
             struct tallocr_chunk * chunk = alloc->chunks[c];
             if (chunk->n_free_blocks > 0) {
                 struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1];
                 max_avail = MAX(max_avail, block->size);
-                int64_t reuse_factor = chunk->max_size - block->offset - size;
-                // reuse_factor < 0 : amount of extra memory that needs to be allocated
-                // reuse_factor = 0 : allocated free space exactly matches tensor size
-                // reuse_factor > 0 : superfluous memory that will remain unused
-                bool better_reuse = best_reuse < 0 && reuse_factor > best_reuse;
-                bool better_fit = reuse_factor >= 0 && reuse_factor < best_reuse;
-                if (block->size >= size && (better_reuse || better_fit)) {
+                if (block->size >= size) {
                     best_fit_chunk = c;
                     best_fit_block = chunk->n_free_blocks - 1;
-                    best_reuse = reuse_factor;
+                    break;
                 }
             }
         }
@@ -275,7 +268,7 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, addr, tensor);
     size_t cur_max = addr.offset + size;
-    if (cur_max > chunk->max_size) {
+    if (cur_max > alloc->max_size[addr.chunk]) {
         // sort allocated_tensors by chunk/offset
         for (int i = 0; i < 1024; i++) {
             for (int j = i + 1; j < 1024; j++) {
@@ -399,8 +392,12 @@ static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
     free(alloc);
 }
 
-static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc, int chunk) {
-    return chunk < alloc->n_chunks ? alloc->chunks[chunk]->max_size : 0;
+static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
+    size_t max_size = 0;
+    for (int i = 0; i < alloc->n_chunks; i++) {
+        max_size += alloc->chunks[i]->max_size;
+    }
+    return max_size;
 }
 
 
@@ -420,8 +417,10 @@ static void ggml_vbuffer_free(struct vbuffer * buf) {
     free(buf);
 }
 
-static size_t ggml_vbuffer_chunk_size(struct vbuffer * buf, int chunk) {
-    return buf->chunks[chunk] ? ggml_backend_buffer_get_size(buf->chunks[chunk]) : 0;
+static int ggml_vbuffer_n_chunks(struct vbuffer * buf) {
+    int n = 0;
+    while (n < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[n]) n++;
+    return n;
 }
 
 static size_t ggml_vbuffer_size(struct vbuffer * buf) {
@@ -605,26 +604,6 @@ static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
-// free the extra space at the end if the new tensor is smaller
-static void ggml_gallocr_free_extra_space(ggml_gallocr_t galloc, struct ggml_tensor * node, struct ggml_tensor * parent) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
-
-    size_t parent_size = ggml_backend_buft_get_alloc_size(galloc->bufts[p_hn->buffer_id], parent);
-    size_t node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
-
-    GGML_ASSERT(parent_size >= node_size);
-
-    if (parent_size > node_size) {
-        struct ggml_dyn_tallocr * p_alloc = galloc->buf_tallocs[p_hn->buffer_id];
-        struct buffer_address p_addr = p_hn->addr;
-        p_addr.offset += node_size;
-        size_t extra_size = parent_size - node_size;
-        AT_PRINTF("freeing extra %zu bytes from parent %s for %s\n", extra_size, parent->name, node->name);
-        ggml_dyn_tallocr_free_tensor(p_alloc, p_addr, extra_size, parent);
-    }
-}
-
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
     GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
@@ -670,7 +649,6 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                             hn->addr = p_hn->addr;
                             p_hn->allocated = false; // avoid freeing the parent
                             view_src_hn->allocated = false;
-                            ggml_gallocr_free_extra_space(galloc, node, view_src);
                             return;
                         }
                     } else {
@@ -678,7 +656,6 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                         hn->buffer_id = p_hn->buffer_id;
                         hn->addr = p_hn->addr;
                         p_hn->allocated = false; // avoid freeing the parent
-                        ggml_gallocr_free_extra_space(galloc, node, parent);
                         return;
                     }
                 }
@@ -908,20 +885,12 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
             }
         }
 
+        size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
+        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
+
         // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
-        bool realloc = galloc->buffers[i] == NULL;
-        size_t new_size = 0;
-        for (int c = 0; c < galloc->buf_tallocs[i]->n_chunks; c++) {
-            size_t cur_chunk_size = galloc->buffers[i] ? ggml_vbuffer_chunk_size(galloc->buffers[i], c) : 0;
-            size_t new_chunk_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i], c);
-            new_size += new_chunk_size;
-            if (new_chunk_size > cur_chunk_size) {
-                realloc = true;
-            }
-        }
-        if (realloc) {
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
 #ifndef NDEBUG
-            size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
             GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
 

ggml/src/ggml-backend-impl.h

@@ -209,6 +209,9 @@ extern "C" {
         void * context;
     };
 
+    // Internal backend registry API
+    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
+
     // Add backend dynamic loading support to the backend
 
     // Initialize the backend

ggml/src/ggml-backend-reg.cpp

@@ -57,10 +57,6 @@
 #include "ggml-opencl.h"
 #endif
 
-#ifdef GGML_USE_HEXAGON
-#include "ggml-hexagon.h"
-#endif
-
 #ifdef GGML_USE_BLAS
 #include "ggml-blas.h"
 #endif
@@ -139,10 +135,6 @@ static void * dl_get_sym(dl_handle * handle, const char * name) {
     return p;
 }
 
-static const char * dl_error() {
-    return "";
-}
-
 #else
 
 using dl_handle = void;
@@ -163,11 +155,6 @@ static void * dl_get_sym(dl_handle * handle, const char * name) {
     return dlsym(handle, name);
 }
 
-static const char * dl_error() {
-    const char *rslt = dlerror();
-    return rslt != nullptr ? rslt : "";
-}
-
 #endif
 
 using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
@@ -203,9 +190,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
-#ifdef GGML_USE_HEXAGON
-        register_backend(ggml_backend_hexagon_reg());
-#endif
 #ifdef GGML_USE_CANN
         register_backend(ggml_backend_cann_reg());
 #endif
@@ -256,7 +240,7 @@ struct ggml_backend_registry {
         dl_handle_ptr handle { dl_load_library(path) };
         if (!handle) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path_str(path).c_str(), dl_error());
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(path).c_str());
             }
             return nullptr;
         }
@@ -546,7 +530,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
                 if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
                     dl_handle_ptr handle { dl_load_library(entry) };
                     if (!handle && !silent) {
-                        GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path_str(entry.path()).c_str(), dl_error());
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(entry.path()).c_str());
                     }
                     if (handle) {
                         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
@@ -605,7 +589,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("sycl", silent, dir_path);
     ggml_backend_load_best("vulkan", silent, dir_path);
     ggml_backend_load_best("opencl", silent, dir_path);
-    ggml_backend_load_best("hexagon", silent, dir_path);
     ggml_backend_load_best("musa", silent, dir_path);
     ggml_backend_load_best("cpu", silent, dir_path);
     // check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend

ggml/src/ggml-blas/CMakeLists.txt

@@ -74,7 +74,7 @@ if (BLAS_FOUND)
 
     target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
 
-    if ("${BLAS_INCLUDE_DIRS}" MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+    if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
         add_compile_definitions(GGML_BLAS_USE_MKL)
     endif()
 

ggml/src/ggml-cann/acl_tensor.cpp

@@ -51,31 +51,28 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
     return ACL_DT_UNDEFINED;
 }
 
-aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
-                                    int64_t *           ne,
-                                    size_t *            nb,
-                                    int64_t             dims,
-                                    aclFormat           format,
-                                    size_t              offset) {
+aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
+                                   size_t* nb, int64_t dims, aclFormat format,
+                                   size_t offset) {
     // If tensor is bcasted, Up to GGML_MAX_DIMS additional dimensions will be
     // added.
     int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
 
     if (ne == nullptr) {
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
-            acl_ne[i]     = tensor->ne[i];
+            acl_ne[i] = tensor->ne[i];
             // The step size of acl is in elements.
             acl_stride[i] = tensor->nb[i] / ggml_element_size(tensor);
         }
     } else {
         // With bcast
         for (int i = 0; i < dims; i++) {
-            acl_ne[i]     = ne[i];
+            acl_ne[i] = ne[i];
             acl_stride[i] = nb[i] / ggml_element_size(tensor);
         }
     }
 
-    int64_t final_dims      = (dims == 0 ? GGML_MAX_DIMS : dims);
+    int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
     int64_t acl_storage_len = 1;
     for (int i = 0; i < final_dims; i++) {
         acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
@@ -87,13 +84,15 @@ aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
     std::reverse(acl_ne, acl_ne + final_dims);
     std::reverse(acl_stride, acl_stride + final_dims);
 
-    aclTensor * acl_tensor = aclCreateTensor(acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
-                                             elem_offset, format, &acl_storage_len, 1, tensor->data);
+    aclTensor* acl_tensor = aclCreateTensor(
+        acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
+        elem_offset, format, &acl_storage_len, 1,
+        tensor->data);
 
     return acl_tensor;
 }
 
-bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1) {
+bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (t1->ne[i] != t0->ne[i] && t1->ne[i] != 1) {
             return true;
@@ -102,16 +101,15 @@ bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1) {
     return false;
 }
 
-int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
-                                  const ggml_tensor * src1,
-                                  int64_t *           bcast_src0_ne,
-                                  int64_t *           bcast_src1_ne,
-                                  size_t *            bcast_src0_nb,
-                                  size_t *            bcast_src1_nb) {
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
+                                  const ggml_tensor* src1,
+                                  int64_t* bcast_src0_ne,
+                                  int64_t* bcast_src1_ne, size_t* bcast_src0_nb,
+                                  size_t* bcast_src1_nb) {
     GGML_ASSERT(ggml_can_repeat(src1, src0));
     int bcast_dim_cnt = 0;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        int64_t nr                   = src0->ne[i] / src1->ne[i];
+        int64_t nr = src0->ne[i] / src1->ne[i];
         bcast_src0_ne[bcast_dim_cnt] = src0->ne[i] / nr;
         bcast_src1_ne[bcast_dim_cnt] = src1->ne[i];
         bcast_src0_nb[bcast_dim_cnt] = src0->nb[i];
@@ -121,26 +119,21 @@ int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
             // Need to add an extra dim.
             bcast_src0_ne[bcast_dim_cnt] = nr;
             bcast_src1_ne[bcast_dim_cnt] = 1;
-            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] * bcast_src0_ne[bcast_dim_cnt - 1];
-            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] * bcast_src1_ne[bcast_dim_cnt - 1];
+            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] *
+                                           bcast_src0_ne[bcast_dim_cnt - 1];
+            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] *
+                                           bcast_src1_ne[bcast_dim_cnt - 1];
             bcast_dim_cnt++;
         }
     }
     return bcast_dim_cnt;
 }
 
-int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
-                                         const int64_t * weight_ne,
-                                         const int64_t * dst_ne,
-                                         const size_t *  input_nb,
-                                         const size_t *  weight_nb,
-                                         const size_t *  dst_nb,
-                                         int64_t *       bcast_input_ne,
-                                         int64_t *       bcast_weight_ne,
-                                         int64_t *       bcast_dst_ne,
-                                         size_t *        bcast_input_nb,
-                                         size_t *        bcast_weight_nb,
-                                         size_t *        bcast_dst_nb) {
+int64_t ggml_cann_get_mulmat_bcast_shape(
+    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
+    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
+    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
+    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb) {
     // input and dst shoule in same shape, except first two dims.
     GGML_ASSERT(input_ne[2] == dst_ne[2]);
     GGML_ASSERT(input_ne[3] == dst_ne[3]);
@@ -155,30 +148,34 @@ int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
         // Do not use bcast in the first two dimensions because we only support
         // the bcast batch dimension. Just copy them.
         if (i < 2 || nr == 1) {
-            bcast_input_ne[bcast_dim_cnt]  = input_ne[i];
+            bcast_input_ne[bcast_dim_cnt] = input_ne[i];
             bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
-            bcast_dst_ne[bcast_dim_cnt]    = dst_ne[i];
+            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i];
 
-            bcast_input_nb[bcast_dim_cnt]  = input_nb[i];
+            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
             bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
-            bcast_dst_nb[bcast_dim_cnt]    = dst_nb[i];
+            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
             bcast_dim_cnt++;
         } else {
             // Need to add an extra dim.
-            bcast_input_ne[bcast_dim_cnt]  = nr;
-            bcast_dst_ne[bcast_dim_cnt]    = nr;
+            bcast_input_ne[bcast_dim_cnt] = nr;
+            bcast_dst_ne[bcast_dim_cnt] = nr;
             bcast_weight_ne[bcast_dim_cnt] = 1;
-            bcast_input_nb[bcast_dim_cnt]  = input_nb[i];
-            bcast_dst_nb[bcast_dim_cnt]    = dst_nb[i];
+            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
+            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
             bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
             bcast_dim_cnt++;
 
-            bcast_input_ne[bcast_dim_cnt]  = input_ne[i] / nr;
-            bcast_dst_ne[bcast_dim_cnt]    = dst_ne[i] / nr;
+            bcast_input_ne[bcast_dim_cnt] = input_ne[i] / nr;
+            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i] / nr;
             bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
-            bcast_input_nb[bcast_dim_cnt]  = bcast_input_nb[bcast_dim_cnt - 1] * bcast_input_ne[bcast_dim_cnt - 1];
-            bcast_dst_nb[bcast_dim_cnt]    = bcast_dst_nb[bcast_dim_cnt - 1] * bcast_dst_ne[bcast_dim_cnt - 1];
-            bcast_weight_nb[bcast_dim_cnt] = bcast_weight_nb[bcast_dim_cnt - 1] * bcast_weight_ne[bcast_dim_cnt - 1];
+            bcast_input_nb[bcast_dim_cnt] = bcast_input_nb[bcast_dim_cnt - 1] *
+                                            bcast_input_ne[bcast_dim_cnt - 1];
+            bcast_dst_nb[bcast_dim_cnt] = bcast_dst_nb[bcast_dim_cnt - 1] *
+                                          bcast_dst_ne[bcast_dim_cnt - 1];
+            bcast_weight_nb[bcast_dim_cnt] =
+                bcast_weight_nb[bcast_dim_cnt - 1] *
+                bcast_weight_ne[bcast_dim_cnt - 1];
             bcast_dim_cnt++;
         }
     }

ggml/src/ggml-cann/acl_tensor.h

@@ -62,12 +62,10 @@ aclDataType ggml_cann_type_mapping(ggml_type type);
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
-aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
-                                    int64_t *           ne     = nullptr,
-                                    size_t *            nb     = nullptr,
-                                    int64_t             dims   = 0,
-                                    aclFormat           format = ACL_FORMAT_ND,
-                                    size_t              offset = 0);
+aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
+                             size_t* nb = nullptr, int64_t dims = 0,
+                             aclFormat format = ACL_FORMAT_ND,
+                             size_t offset = 0);
 
 /**
  * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
@@ -89,15 +87,12 @@ aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
-template <typename TYPE>
-aclTensor * ggml_cann_create_tensor(void *      data_ptr,
-                                    aclDataType dtype,
-                                    TYPE        type_size,
-                                    int64_t *   ne,
-                                    TYPE *      nb,
-                                    int64_t     dims,
-                                    aclFormat   format = ACL_FORMAT_ND,
-                                    size_t      offset = 0) {
+template<typename TYPE>
+aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
     int64_t tmp_ne[GGML_MAX_DIMS * 2];
     int64_t tmp_stride[GGML_MAX_DIMS * 2];
 
@@ -114,8 +109,9 @@ aclTensor * ggml_cann_create_tensor(void *      data_ptr,
     std::reverse(tmp_ne, tmp_ne + dims);
     std::reverse(tmp_stride, tmp_stride + dims);
 
-    aclTensor * acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size, format, &acl_storage_len, 1, data_ptr);
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
 
     return acl_tensor;
 }
@@ -136,7 +132,7 @@ aclTensor * ggml_cann_create_tensor(void *      data_ptr,
  *          to 1. If such a dimension is found, broadcasting is required to align t1
  *          with t0 for element-wise operations.
  */
-bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1);
+bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
 
 /**
  * @brief   Computes broadcast shapes and strides for two ggml_tensors.
@@ -191,21 +187,19 @@ bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1);
  *  dim1 in a inserted dim, should add nb for dim1,
  *  and all other nb moves to next in order.
  */
-int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
-                                  const ggml_tensor * src1,
-                                  int64_t *           bcast_ne_src0,
-                                  int64_t *           bcast_ne_src1,
-                                  size_t *            bcast_nb_src0,
-                                  size_t *            bcast_nb_src1);
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
+                        int64_t* bcast_ne_src0, int64_t* bcast_ne_src1,
+                        size_t* bcast_nb_src0, size_t* bcast_nb_src1);
 
 // Bcast macro to avoid duplicate code.
-#define BCAST_SHAPE(src0, src1)                                                                      \
-    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                                                    \
-    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                                                    \
-    size_t  bcast_##src0##_nb[GGML_MAX_DIMS * 2];                                                    \
-    size_t  bcast_##src1##_nb[GGML_MAX_DIMS * 2];                                                    \
-    int64_t bcast_dims = ggml_cann_get_bcast_shape(src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, \
-                                                   bcast_##src0##_nb, bcast_##src1##_nb);
+#define BCAST_SHAPE(src0, src1)                                              \
+    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                            \
+    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                            \
+    size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2];                             \
+    size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2];                             \
+    int64_t bcast_dims = ggml_cann_get_bcast_shape(                          \
+        src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, bcast_##src0##_nb, \
+        bcast_##src1##_nb);
 
 #define BCAST_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
 
@@ -239,31 +233,26 @@ int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
  *       before cast dim.
  * @sa ggml_cann_get_bcast_shape
  */
-int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
-                                         const int64_t * weight_ne,
-                                         const int64_t * dst_ne,
-                                         const size_t *  input_nb,
-                                         const size_t *  weight_nb,
-                                         const size_t *  dst_nb,
-                                         int64_t *       bcast_input_ne,
-                                         int64_t *       bcast_weight_ne,
-                                         int64_t *       bcast_dst_ne,
-                                         size_t *        bcast_input_nb,
-                                         size_t *        bcast_weight_nb,
-                                         size_t *        bcast_dst_nb);
+int64_t ggml_cann_get_mulmat_bcast_shape(
+    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
+    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
+    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
+    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb);
 
 // Bcast macro to avoid duplicate code.
-#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                                                                  \
-    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                                                               \
-    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                                                              \
-    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                                                                 \
-    size_t  bcast_##input##_nb[GGML_MAX_DIMS * 2];                                                               \
-    size_t  bcast_##weight##_nb[GGML_MAX_DIMS * 2];                                                              \
-    size_t  bcast_##dst##_nb[GGML_MAX_DIMS * 2];                                                                 \
-    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(                                                       \
-        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, bcast_##input##_ne, bcast_##weight##_ne, \
-        bcast_##dst##_ne, bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
+#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                         \
+    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                      \
+    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                     \
+    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                        \
+    size_t bcast_##input##_nb[GGML_MAX_DIMS * 2];                       \
+    size_t bcast_##weight##_nb[GGML_MAX_DIMS * 2];                      \
+    size_t bcast_##dst##_nb[GGML_MAX_DIMS * 2];                         \
+    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(              \
+        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, \
+        bcast_##input##_ne, bcast_##weight##_ne, bcast_##dst##_ne,      \
+        bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
 
-#define BCAST_MUL_MAT_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
+#define BCAST_MUL_MAT_PARAM(tensor) \
+    bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
 
 #endif  // CANN_ACL_TENSOR_H

ggml/src/ggml-cann/aclnn_ops.h

@@ -62,7 +62,7 @@
  * @param   dst The ggml tensor representing the destination, which op is
  *              GGML_OP_REPEAT and specifies the desired dimensions.
  */
-void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
@@ -82,7 +82,7 @@ void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the result of the Leaky ReLU
  *            activation is stored, which op is `GGML_OP_LEAKY_RELU`
  */
-void ggml_cann_leaky_relu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief    Concatenates multiple tensors along a specified dimension using the
@@ -97,7 +97,7 @@ void ggml_cann_leaky_relu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @attention tensorList length should be 2 and the dimension using for concat
  *            default to 1.
  */
-void ggml_cann_concat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Generates a sequence of evenly spaced values within a specified
@@ -113,7 +113,7 @@ void ggml_cann_concat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *            `start`, 'stop' and 'step' are in dst->op_params and dst->op is
  *            `GGML_OP_ARANGE`.
  */
-void ggml_cann_arange(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Applies a clamp operation to the elements of a ggml tensor using the
@@ -131,7 +131,7 @@ void ggml_cann_arange(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the clamped values will be stored.
  *            dst->op is `GGML_OP_CLAMP`, `min` and `max` value is in dst->params.
  */
-void ggml_cann_clamp(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Scales the elements of a ggml tensor by a constant factor using the
@@ -148,7 +148,7 @@ void ggml_cann_clamp(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the scaled values will be stored.
  *            dst->op is `GGML_OP_SCALE` and `scale` value is in dst->params.
  */
-void ggml_cann_scale(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Sorts the elements of a ggml tensor and returns the indices that
@@ -163,7 +163,7 @@ void ggml_cann_scale(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the sorted indices will be stored.
  *            dst->op is `GGML_OP_ARGSORT`.
  */
-void ggml_cann_argsort(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the Layer Normalization for a ggml tensor using the CANN
@@ -185,7 +185,7 @@ void ggml_cann_argsort(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the normalized values will be stored.
  * @attention `Var` defaults to dst->ne[0].
  */
-void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief  Computes the Group Normalization for a ggml tensor using the CANN
@@ -209,7 +209,7 @@ void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *
  * @attention eps defaults to 1e-6f.
  */
-void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the accumulation of tensors using the CANN backend.
@@ -228,7 +228,7 @@ void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the accumulated values will be stored.
  *            `inplace` is in dst->params, and dst->op is `GGML_OP_ACC`.
  */
-void ggml_cann_acc(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the sum of elements along the last dimension of a ggml tensor
@@ -244,7 +244,7 @@ void ggml_cann_acc(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *
  * @attention `reduce_dims` defaults to 3, which means the last dimension.
  */
-void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the sum of elements in a ggml tensor.
@@ -258,7 +258,7 @@ void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *
  */
 
-void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
@@ -274,7 +274,8 @@ void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the upsampled values will be stored.
  *            dst->op is `GGML_OP_UPSCALE`.
  */
-void ggml_cann_upsample_nearest2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
+                                  ggml_tensor* dst);
 
 /**
  * @brief   Pads a ggml tensor to match the dimensions of the destination tensor
@@ -289,7 +290,7 @@ void ggml_cann_upsample_nearest2d(ggml_backend_cann_context & ctx, ggml_tensor *
  * @param dst The destination tensor, which specifies the target dimensions for
  *            padding. dst->op is `GGML_OP_PAD`.
  */
-void ggml_cann_pad(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Executes a 2D pooling operation on a ggml tensor using the CANN
@@ -306,7 +307,7 @@ void ggml_cann_pad(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor on which the pooling operation is to be
  *            performed. dst->op is `GGML_OP_POOL_2D`.
  */
-void ggml_cann_pool2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Duplicates a ggml tensor using the CANN backend.
@@ -325,7 +326,7 @@ void ggml_cann_pool2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *            different shape and dst is no-contiguous.
  * @note:     This func need to simplify.
  */
-void ggml_cann_dup(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the Root Mean Square (RMS) normalization of a ggml tensor
@@ -347,7 +348,7 @@ void ggml_cann_dup(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the normalized values will be stored.
  *            dst->op is `GGML_OP_RMS_NORM`.
  */
-void ggml_cann_rms_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Applies a diagonal mask to the tensor with a specified value.
@@ -362,7 +363,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *            `GGML_OP_DIAG_MASK`
  * @param value The value to use for masking.
  */
-void ggml_cann_diag_mask(ggml_backend_cann_context & ctx, ggml_tensor * dst, float value);
+void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, float value);
 
 /**
  * @brief   Performs an image-to-column transformation on the input tensor.
@@ -377,7 +378,7 @@ void ggml_cann_diag_mask(ggml_backend_cann_context & ctx, ggml_tensor * dst, flo
  * @param dst The destination tensor that stores the result of the operation.
  *            dst->op is `GGML_OP_IM2COL`.
  */
-void ggml_cann_im2col(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes time step embeddings using sine and cosine functions.
@@ -391,10 +392,10 @@ void ggml_cann_im2col(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the result of the embedding operation
  *            will be stored. dst->op is `GGML_OP_TIMESTEP_EMBEDDING`.
  */
-void ggml_cann_timestep_embedding(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 // @see ggml_cann_dup.
-void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the softmax activation with optional masking.
@@ -416,7 +417,7 @@ void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the result will be stored. dst->op is
  *            `GGML_OP_SOFTMAX`.
  */
-void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Extracts specific rows from a tensor based on indices.
@@ -428,7 +429,7 @@ void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the extracted rows will be stored.
  */
-void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Writes specific rows into a tensor at positions specified by indices.
@@ -440,7 +441,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the specified rows will be updated.
  */
-void ggml_cann_set_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Executes matrix multiplication for the given tensor.
@@ -453,7 +454,7 @@ void ggml_cann_set_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor for storing the result of the matrix
  *            multiplication. dst->op is `GGML_OP_MUL_MAT`.
  */
-void ggml_cann_mul_mat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief Applies Rotary Positional Embedding (RoPE) to the input tensor.
@@ -476,7 +477,7 @@ void ggml_cann_mul_mat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @note The function currently does not support cases where the freq_scale is
  *       not equal 1.
  */
-void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the index of the maximum value along the specified dimension
@@ -491,7 +492,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the indices of the maximum values will
  *            be stored. dst->op is `GGML_OP_ARGMAX`.
  */
-void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief Adds two tensors element-wise and stores the result in a destination
@@ -508,10 +509,8 @@ void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param acl_src1 The second source tensor.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_add(ggml_backend_cann_context & ctx,
-               aclTensor *                 acl_src0,
-               aclTensor *                 acl_src1,
-               aclTensor *                 acl_dst = nullptr);
+void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
+    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
 
 /**
  * @brief Sub two tensors element-wise and stores the result in a destination
@@ -528,10 +527,8 @@ void aclnn_add(ggml_backend_cann_context & ctx,
  * @param acl_src1 The second source tensor.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_sub(ggml_backend_cann_context & ctx,
-               aclTensor *                 acl_src0,
-               aclTensor *                 acl_src1,
-               aclTensor *                 acl_dst = nullptr);
+void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
+    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
 
 /**
  * @brief Performs element-wise multiplication of two tensors and stores the
@@ -549,10 +546,8 @@ void aclnn_sub(ggml_backend_cann_context & ctx,
  * @param acl_other The second tensor for element-wise multiplication.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_mul(ggml_backend_cann_context & ctx,
-               aclTensor *                 acl_src,
-               aclTensor *                 acl_other,
-               aclTensor *                 acl_dst = nullptr);
+void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
 
 /**
  * @brief Matrix division, optionally in-place.
@@ -572,10 +567,8 @@ void aclnn_mul(ggml_backend_cann_context & ctx,
  * @param inplace Flag indicating whether to perform the operation in-place on
  * `acl_src`.
  */
-void aclnn_div(ggml_backend_cann_context & ctx,
-               aclTensor *                 acl_src,
-               aclTensor *                 acl_other,
-               aclTensor *                 acl_dst = nullptr);
+void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
 
 /**
  * @brief Applies element-wise cosine function to the elements of a tensor.
@@ -591,7 +584,8 @@ void aclnn_div(ggml_backend_cann_context & ctx,
  * @param acl_dst The destination tensor where the cosine results will be
  * stored.
  */
-void aclnn_cos(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst);
+void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_dst);
 
 /**
  * @brief Applies element-wise sine function to the elements of a tensor.
@@ -608,7 +602,8 @@ void aclnn_cos(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor *
  * @param acl_src The source tensor on which the sine function will be applied.
  * @param acl_dst The destination tensor where the sine results will be stored.
  */
-void aclnn_sin(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst);
+void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_dst);
 
 /**
  * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one
@@ -626,12 +621,8 @@ void aclnn_sin(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor *
  * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
  * @param acl_dst  Output pointer to the created ACL tensor corresponding to dst.
  */
-void bcast_shape(ggml_tensor * src0,
-                 ggml_tensor * src1,
-                 ggml_tensor * dst,
-                 aclTensor **  acl_src0,
-                 aclTensor **  acl_src1,
-                 aclTensor **  acl_dst);
+void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
+    aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst);
 
 /**
  * @brief   Computes the 1D transposed convolution (deconvolution) of a ggml
@@ -646,7 +637,7 @@ void bcast_shape(ggml_tensor * src0,
  * @param dst The destination tensor where the transposed convolution result
  * will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`.
  */
-void ggml_cann_conv_transpose_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Applies the ELU (Exponential Linear Unit) activation to a ggml tensor
@@ -671,7 +662,7 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context & ctx, ggml_tensor *
  * @param dst The destination tensor where the ELU-activated result will be stored.
  *            dst->op is expected to be `GGML_OP_ELU`.
  */
-void ggml_cann_elu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Computes the mean of a ggml tensor element-wise using the CANN backend.
@@ -686,7 +677,7 @@ void ggml_cann_elu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the mean result will be stored.
  *            dst->op is expected to be `GGML_OP_MEAN`.
  */
-void ggml_cann_mean(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Applies 1D reflect padding to a ggml tensor using the CANN backend.
@@ -701,7 +692,7 @@ void ggml_cann_mean(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the padded result will be stored.
  *            dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
  */
-void ggml_cann_pad_reflect_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Counts the number of equal elements in two ggml tensors using the CANN backend.
@@ -717,7 +708,7 @@ void ggml_cann_pad_reflect_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
  */
-void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Applies the Step activation function to a ggml tensor using the CANN backend.
@@ -732,7 +723,7 @@ void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_STEP`.
  */
-void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Performs the Flash Attention extended operator using the CANN backend.
@@ -747,46 +738,59 @@ void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_FLASH_ATTN_EXT`.
  */
-void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /*
  * @brief A generic wrapper for ACL resources with custom deleter support.
  */
-using any_acl_resource = std::unique_ptr<void, std::function<void(void *)>>;
+using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
 
 /**
  * @brief Trait structure used to define how to destroy a given ACL resource type.
  *
  * @tparam T ACL resource type.
  */
-template <typename T> struct acl_resource_traits;
+template<typename T>
+struct acl_resource_traits;
 
 /**
  * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
  */
-template <> struct acl_resource_traits<aclTensor> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyTensor(static_cast<aclTensor *>(p))); }
+template<>
+struct acl_resource_traits<aclTensor> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
+    }
 };
 
 /**
  * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
  */
-template <> struct acl_resource_traits<aclIntArray> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray *>(p))); }
+template<>
+struct acl_resource_traits<aclIntArray> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
+    }
 };
 
 /**
  * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
  */
-template <> struct acl_resource_traits<aclScalar> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyScalar(static_cast<aclScalar *>(p))); }
+template<>
+struct acl_resource_traits<aclScalar> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
+    }
 };
 
 /**
  * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
  */
-template <> struct acl_resource_traits<aclTensorList> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList *>(p))); }
+template<>
+struct acl_resource_traits<aclTensorList> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
+    }
 };
 
 /**
@@ -796,8 +800,14 @@ template <> struct acl_resource_traits<aclTensorList> {
  * @param ptr Raw pointer to ACL resource.
  * @return any_acl_resource Smart pointer that handles destruction.
  */
-template <typename T> any_acl_resource make_acl_resource(T * ptr) {
-    return any_acl_resource(static_cast<void *>(ptr), [](void * p) { acl_resource_traits<T>::destroy(p); });
+template<typename T>
+any_acl_resource make_acl_resource(T* ptr) {
+    return any_acl_resource(
+        static_cast<void*>(ptr),
+        [](void* p) {
+            acl_resource_traits<T>::destroy(p);
+        }
+    );
 }
 
 /**
@@ -807,7 +817,8 @@ template <typename T> any_acl_resource make_acl_resource(T * ptr) {
  * @param vec Target vector to hold ACL resources.
  * @param args Raw pointers to ACL resources.
  */
-template <typename... Args> void register_acl_resources(std::vector<any_acl_resource> & vec, Args *... args) {
+template<typename... Args>
+void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
     (vec.emplace_back(make_acl_resource(args)), ...);
 }
 
@@ -815,36 +826,39 @@ template <typename... Args> void register_acl_resources(std::vector<any_acl_reso
  * @brief Task class that wraps the execution of an aclnn function call.
  */
 class aclnn_task : public cann_task {
-  public:
-    aclnn_task(aclnn_func_t    aclnn_func,
-               void *          workspace_addr,
-               uint64_t        workspace_size,
-               aclOpExecutor * executor,
-               aclrtStream     stream) :
-        aclnn_func_(aclnn_func),
-        workspace_addr_(workspace_addr),
-        workspace_size_(workspace_size),
-        executor_(executor),
-        stream_(stream) {}
-
-    virtual void run_task() override { ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_)); }
-  private:
-    aclnn_func_t    aclnn_func_;
-    void *          workspace_addr_;
-    uint64_t        workspace_size_;
-    aclOpExecutor * executor_;
-    aclrtStream     stream_;
+    public:
+        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
+                   uint64_t workspace_size, aclOpExecutor * executor,
+                   aclrtStream stream) :
+            aclnn_func_(aclnn_func),
+            workspace_addr_(workspace_addr),
+            workspace_size_(workspace_size),
+            executor_(executor),
+            stream_(stream) {}
+        virtual void run_task() override {
+            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
+        }
+    private:
+        aclnn_func_t aclnn_func_;
+        void *          workspace_addr_;
+        uint64_t        workspace_size_;
+        aclOpExecutor * executor_;
+        aclrtStream     stream_;
 };
 
 /**
  * @brief Task class that releases ACL resources after usage.
  */
 class release_resource_task : public cann_task {
-  public:
-    release_resource_task(std::vector<any_acl_resource> && resources) { resource_ = std::move(resources); }
+public:
+    release_resource_task(std::vector<any_acl_resource>&& resources){
+        resource_ = std::move(resources);
+    }
 
-    virtual void run_task() override { resource_.clear(); }
-  private:
+    virtual void run_task() override {
+        resource_.clear();
+    }
+private:
     std::vector<any_acl_resource> resource_;
 };
 
@@ -852,40 +866,38 @@ class release_resource_task : public cann_task {
  * @brief Task class for performing asynchronous memory copy operations.
  */
 class async_memcpy_task : public cann_task {
-  public:
-    async_memcpy_task(void * dst, const void * src, size_t size, aclrtMemcpyKind kind, aclrtStream stream) :
-        dst_(dst),
-        src_(src),
-        size_(size),
-        kind_(kind),
-        stream_(stream) {}
+public:
+    async_memcpy_task(void* dst, const void* src, size_t size,
+                      aclrtMemcpyKind kind, aclrtStream stream)
+        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
 
-    virtual void run_task() override { ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_)); }
-  private:
-    void *          dst_;
-    const void *    src_;
-    size_t          size_;
+    virtual void run_task() override {
+        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
+    }
+private:
+    void* dst_;
+    const void* src_;
+    size_t size_;
     aclrtMemcpyKind kind_;
-    aclrtStream     stream_;
+    aclrtStream stream_;
 };
 
 /**
  * @brief Task class for performing asynchronous memory set operations.
  */
 class async_memset_task : public cann_task {
-  public:
-    async_memset_task(void * buffer, size_t size, int32_t value, aclrtStream stream) :
-        buffer_(buffer),
-        size_(size),
-        value_(value),
-        stream_(stream) {}
+    public:
+    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
+            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
 
-    virtual void run_task() override { ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_)); }
-  private:
-    void *      buffer_;
-    size_t      size_;
-    int32_t     value_;
-    aclrtStream stream_;
+        virtual void run_task() override {
+            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
+        }
+    private:
+        void* buffer_;
+        size_t size_;
+        int32_t value_;
+        aclrtStream stream_;
 };
 
 /**
@@ -906,24 +918,25 @@ class async_memset_task : public cann_task {
  * same stream are executed in queue order.
  */
 
-#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                                                  \
-    do {                                                                                                            \
-        uint64_t        workspaceSize = 0;                                                                          \
-        aclOpExecutor * executor;                                                                                   \
-        void *          workspaceAddr = nullptr;                                                                    \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));                        \
-        /* workspace should alloced in main thread to keep malloc order when using vmm. */                          \
-        if (workspaceSize > 0) {                                                                                    \
-            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);                                    \
-            workspaceAddr = workspace_allocator.get();                                                              \
-        }                                                                                                           \
-        if (CTX.async_mode) {                                                                                       \
-            auto task =                                                                                             \
-                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize, executor, CTX.stream()); \
-            CTX.task_queue.submit_task(std::move(task));                                                            \
-        } else {                                                                                                    \
-            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));                        \
-        }                                                                                                           \
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
+    do {                                                                                    \
+        uint64_t        workspaceSize = 0;                                                  \
+        aclOpExecutor * executor;                                                           \
+        void *          workspaceAddr = nullptr;                                            \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
+        if (workspaceSize > 0) {                                                            \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
+            workspaceAddr = workspace_allocator.get();                                      \
+        }                                                                                   \
+        if (CTX.async_mode) {                                                               \
+            auto task =                                                                     \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
+                    executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                    \
+        } else {                                                                            \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
+        }                                                                                   \
     } while (0)
 
 /**
@@ -934,10 +947,11 @@ class async_memset_task : public cann_task {
  * @param ctx Backend context which manages task submission and async mode.
  * @param args Pointers to ACL resources to be released.
  */
-template <typename... Args> void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+template <typename... Args>
+void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
     std::vector<any_acl_resource> resources;
     register_acl_resources(resources, std::forward<Args>(args)...);
-    if (ctx.async_mode) {
+    if(ctx.async_mode) {
         auto task = std::make_unique<release_resource_task>(std::move(resources));
         ctx.task_queue.submit_task(std::move(task));
     }
@@ -952,11 +966,8 @@ template <typename... Args> void ggml_cann_release_resources(ggml_backend_cann_c
  * @param len Size of memory to copy (in bytes).
  * @param kind Type of memory copy (host-to-device, device-to-host, etc).
  */
-inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx,
-                                   void *                      dst,
-                                   const void *                src,
-                                   size_t                      len,
-                                   aclrtMemcpyKind             kind) {
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
     if (ctx.async_mode) {
         auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
         ctx.task_queue.submit_task(std::move(task));
@@ -965,11 +976,8 @@ inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx,
     }
 }
 
-inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx,
-                                   void *                      dst,
-                                   const void *                src,
-                                   size_t                      len,
-                                   aclrtMemcpyKind             kind) {
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
     if (ctx->async_mode) {
         auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
         ctx->task_queue.submit_task(std::move(task));
@@ -986,7 +994,8 @@ inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx,
  * @param size Size of the memory buffer (in bytes).
  * @param value Value to set in the buffer.
  */
-inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer, size_t size, int value) {
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
+                                   size_t size, int value) {
     if (ctx.async_mode) {
         auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
         ctx.task_queue.submit_task(std::move(task));
@@ -1020,7 +1029,7 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  * @param dst The destination tensor where the expert-weighted token outputs are stored.
  *            Expected to be of shape [M, K, N, 1].
  */
-void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
@@ -1032,14 +1041,20 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *
  * @param tensor Pointer to the target ggml_tensor object (const-qualified).
  */
-static bool is_matmul_weight(const ggml_tensor * tensor) {
-    std::string                                  name = ggml_get_name(tensor);
-    static const std::unordered_set<std::string> weight_suffixes{ "output.weight",      "attn_q.weight",
-                                                                  "attn_k.weight",      "attn_v.weight",
-                                                                  "attn_output.weight", "ffn_gate.weight",
-                                                                  "ffn_up.weight",      "ffn_down.weight" };
+static bool is_matmul_weight(const ggml_tensor* tensor) {
+    std::string name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{
+        "output.weight",
+        "attn_q.weight",
+        "attn_k.weight",
+        "attn_v.weight",
+        "attn_output.weight",
+        "ffn_gate.weight",
+        "ffn_up.weight",
+        "ffn_down.weight"
+    };
 
-    for (const auto & suffix : weight_suffixes) {
+    for (const auto& suffix : weight_suffixes) {
         if (name.find(suffix) != std::string::npos) {
             return true;
         }
@@ -1063,13 +1078,14 @@ static bool is_matmul_weight(const ggml_tensor * tensor) {
  * @param ctx The CANN backend context used to manage execution and resources.
  * @param dst The destination tensor.
  */
-template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];
-    ggml_tensor * src1 = dst->src[1];
+template <auto binary_op>
+void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
 
-    aclTensor * acl_src0;
-    aclTensor * acl_src1;
-    aclTensor * acl_dst;
+    aclTensor* acl_src0;
+    aclTensor* acl_src1;
+    aclTensor* acl_dst;
 
     // Need bcast
     bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
@@ -1078,6 +1094,7 @@ template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & c
     ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
+
 /**
  * @brief Applies a unary operation to an input tensor using the CANN backend.
  *
@@ -1090,12 +1107,12 @@ template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & c
  * @param ctx The CANN backend context for managing resources and execution.
  * @param dst The destination tensor. Its src[0] is treated as the input tensor.
  */
-template <void unary_op(ggml_backend_cann_context &, aclTensor *, aclTensor *)>
-void ggml_cann_op_unary(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src = dst->src[0];
+template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
+    void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
 
-    aclTensor * acl_src = ggml_cann_create_tensor(src);
-    aclTensor * acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
     unary_op(ctx, acl_src, acl_dst);
     ggml_cann_release_resources(ctx, acl_src, acl_dst);
@@ -1121,9 +1138,9 @@ void ggml_cann_op_unary(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
  *
  * @see GGML_CANN_CALL_OP_UNARY
  */
-void ggml_cann_op_unary(std::function<void(ggml_backend_cann_context &, aclTensor *, aclTensor *)> unary_op,
-                        ggml_backend_cann_context &                                                ctx,
-                        ggml_tensor *                                                              dst);
+void ggml_cann_op_unary(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief Applies a gated (GLU-style) unary operation using the CANN backend.
@@ -1155,9 +1172,9 @@ void ggml_cann_op_unary(std::function<void(ggml_backend_cann_context &, aclTenso
  *
  * @see GGML_CANN_CALL_OP_UNARY_GATED
  */
-void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, aclTensor *, aclTensor *)> unary_op,
-                              ggml_backend_cann_context &                                                ctx,
-                              ggml_tensor *                                                              dst);
+void ggml_cann_op_unary_gated(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
  * @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
@@ -1180,13 +1197,16 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
  * @see ggml_cann_op_unary
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                                                              \
-    do {                                                                                              \
-        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
-        };                                                                                            \
-        ggml_cann_op_unary(lambda, ctx, dst);                                                         \
-    } while (0)
+#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                              \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_op_unary(lambda, ctx, dst);                         \
+    }                                                                 \
+    while (0)
 
 /**
  * @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
@@ -1209,12 +1229,15 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
  * @see ggml_cann_op_unary_gated
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                                                        \
-    do {                                                                                              \
-        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
-        };                                                                                            \
-        ggml_cann_op_unary_gated(lambda, ctx, dst);                                                   \
-    } while (0)
+#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                        \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_op_unary_gated(lambda, ctx, dst);                   \
+    }                                                                 \
+    while (0)
 
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -44,7 +44,7 @@
 #include "../include/ggml.h"
 #include "../ggml-impl.h"
 
-#define MATRIX_ROW_PADDING    512
+#define MATRIX_ROW_PADDING 512
 #define GGML_CANN_MAX_STREAMS 8
 
 /**
@@ -56,7 +56,8 @@
  * @param line The line number at which the error occurred.
  * @param msg The error message.
  */
-[[noreturn]] void ggml_cann_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
+[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
+                                  const char* file, int line, const char* msg);
 
 /**
  * @brief Checks the result of a CANN function call and invokes the error
@@ -88,24 +89,25 @@ struct ggml_cann_device_info {
      * @brief Information about a single CANN device.
      */
     struct cann_device_info {
-        int    cc;              /**< Compute capability.                   */
+        int cc;                 /**< Compute capability.                   */
         size_t smpb;            /**< Maximum shared memory per block.      */
-        bool   vmm;             /**< Virtual memory support.               */
+        bool vmm;               /**< Virtual memory support.               */
         size_t vmm_granularity; /**< Granularity of virtual memory.        */
         size_t total_vram;      /**< Total video RAM available on the device. */
     };
 
-    cann_device_info devices[GGML_CANN_MAX_DEVICES] = {}; /**< Array of CANN device information. */
+    cann_device_info devices[GGML_CANN_MAX_DEVICES] =
+        {}; /**< Array of CANN device information. */
 };
 
-const ggml_cann_device_info & ggml_cann_info();
+const ggml_cann_device_info& ggml_cann_info();
 
-void    ggml_cann_set_device(int32_t device);
+void ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
-std::optional<std::string> get_env(const std::string & name);
-bool                       parse_bool(const std::string & value);
-int                        parse_integer(const std::string & value);
+std::optional<std::string> get_env(const std::string& name);
+bool parse_bool(const std::string& value);
+int parse_integer(const std::string& value);
 
 /**
  * @brief Abstract base class for memory pools used by CANN.
@@ -124,7 +126,7 @@ struct ggml_cann_pool {
      *                     will be stored.
      * @return             Pointer to the allocated memory block.
      */
-    virtual void * alloc(size_t size, size_t * actual_size) = 0;
+    virtual void* alloc(size_t size, size_t* actual_size) = 0;
 
     /**
      * @brief Frees a previously allocated memory block.
@@ -134,16 +136,16 @@ struct ggml_cann_pool {
      * @note Note that all CANN opertors are running async. Make sure memory is
      *       still avaiable before this operator finished.
      */
-    virtual void free(void * ptr, size_t size) = 0;
+    virtual void free(void* ptr, size_t size) = 0;
 };
 
 /**
  * @brief RAII wrapper for managing memory allocations from a CANN memory pool.
  */
 struct ggml_cann_pool_alloc {
-    ggml_cann_pool * pool        = nullptr; /**< Pointer to the memory pool. */
-    void *           ptr         = nullptr; /**< Pointer to the allocated memory block. */
-    size_t           actual_size = 0;       /**< Actual size of the allocated memory block. */
+    ggml_cann_pool* pool = nullptr; /**< Pointer to the memory pool. */
+    void* ptr = nullptr;    /**< Pointer to the allocated memory block. */
+    size_t actual_size = 0; /**< Actual size of the allocated memory block. */
 
     /**
      * @brief Default constructor.
@@ -154,14 +156,16 @@ struct ggml_cann_pool_alloc {
      * @brief Constructor that initializes the memory pool.
      * @param pool Reference to the memory pool.
      */
-    explicit ggml_cann_pool_alloc(ggml_cann_pool & pool) : pool(&pool) {}
+    explicit ggml_cann_pool_alloc(ggml_cann_pool& pool) : pool(&pool) {}
 
     /**
      * @brief Constructor that initializes the memory pool and allocates memory.
      * @param pool Reference to the memory pool.
      * @param size Size of the memory block to allocate.
      */
-    ggml_cann_pool_alloc(ggml_cann_pool & pool, size_t size) : pool(&pool) { alloc(size); }
+    ggml_cann_pool_alloc(ggml_cann_pool& pool, size_t size) : pool(&pool) {
+        alloc(size);
+    }
 
     /**
      * @brief Destructor that frees the allocated memory block.
@@ -177,7 +181,7 @@ struct ggml_cann_pool_alloc {
      * @param size Size of the memory block to allocate.
      * @return Pointer to the allocated memory block.
      */
-    void * alloc(size_t size) {
+    void* alloc(size_t size) {
         GGML_ASSERT(pool != nullptr);
         GGML_ASSERT(ptr == nullptr);
         ptr = pool->alloc(size, &this->actual_size);
@@ -190,7 +194,7 @@ struct ggml_cann_pool_alloc {
      * @param size Size of the memory block to allocate.
      * @return Pointer to the allocated memory block.
      */
-    void * alloc(ggml_cann_pool & pool, size_t size) {
+    void* alloc(ggml_cann_pool& pool, size_t size) {
         this->pool = &pool;
         return alloc(size);
     }
@@ -199,25 +203,25 @@ struct ggml_cann_pool_alloc {
      * @brief Gets the pointer to the allocated memory block.
      * @return Pointer to the allocated memory block.
      */
-    void * get() { return ptr; }
+    void* get() { return ptr; }
 
     // Deleted copy constructor
-    ggml_cann_pool_alloc(const ggml_cann_pool_alloc &) = delete;
+    ggml_cann_pool_alloc(const ggml_cann_pool_alloc&) = delete;
 
     // Deleted move constructor
-    ggml_cann_pool_alloc(ggml_cann_pool_alloc &&) = delete;
+    ggml_cann_pool_alloc(ggml_cann_pool_alloc&&) = delete;
 
     // Deleted copy assignment operator
-    ggml_cann_pool_alloc & operator=(const ggml_cann_pool_alloc &) = delete;
+    ggml_cann_pool_alloc& operator=(const ggml_cann_pool_alloc&) = delete;
 
     // Deleted move assignment operator
-    ggml_cann_pool_alloc & operator=(ggml_cann_pool_alloc &&) = delete;
+    ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
 };
 
 /**
  * @brief Function pointer type for ACLNN operator calls.
  */
-using aclnn_func_t = aclnnStatus (*)(void *, uint64_t, aclOpExecutor *, aclrtStream);
+using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
 
 /**
  * @brief Base class for all CANN tasks to be submitted to the task queue.
@@ -225,7 +229,7 @@ using aclnn_func_t = aclnnStatus (*)(void *, uint64_t, aclOpExecutor *, aclrtStr
  * Users should override the run_task() method with actual task logic.
  */
 class cann_task {
-  public:
+public:
     virtual void run_task() {}
 };
 
@@ -233,20 +237,16 @@ class cann_task {
  * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
  */
 class cann_task_queue {
-  public:
+public:
     /**
      * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
      *
      * @param capacity Queue capacity. Must be a power of 2.
      * @param device Target device ID (used for context setting).
      */
-    explicit cann_task_queue(size_t capacity, int32_t device) :
-        buffer_(capacity),
-        capacity_(capacity),
-        head_(0),
-        tail_(0),
-        running_(false),
-        device_(device) {
+    explicit cann_task_queue(size_t capacity, int32_t device)
+        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
+          running_(false), device_(device) {
         GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
         mask_ = capacity_ - 1;
     }
@@ -257,7 +257,7 @@ class cann_task_queue {
      * @param item Unique pointer to the task.
      * @return true if the task was successfully enqueued, false if the queue was full.
      */
-    bool enqueue(std::unique_ptr<cann_task> && item) {
+    bool enqueue(std::unique_ptr<cann_task>&& item) {
         size_t next_tail = (tail_ + 1) & mask_;
 
         if (next_tail == head_) {
@@ -276,16 +276,17 @@ class cann_task_queue {
      *
      * @param task Task to be submitted.
      */
-    void submit_task(std::unique_ptr<cann_task> && task) {
-        while (!enqueue(std::move(task))) {
+    void submit_task(std::unique_ptr<cann_task>&& task) {
+        while(!enqueue(std::move(task))) {
             std::this_thread::yield();
             continue;
         }
 
         if (!running_) {
             running_ = true;
-            thread_  = std::thread(&cann_task_queue::execute, this);
+            thread_ = std::thread(&cann_task_queue::execute, this);
         }
+
     }
 
     /**
@@ -308,7 +309,7 @@ class cann_task_queue {
         }
     }
 
-  private:
+private:
     /**
      * @brief Worker thread function that continuously dequeues and executes tasks.
      */
@@ -316,7 +317,7 @@ class cann_task_queue {
         ggml_cann_set_device(device_);
 
         while (running_) {
-            if (head_ == tail_) {
+            if(head_ == tail_) {
                 std::this_thread::yield();
                 continue;
             }
@@ -329,29 +330,22 @@ class cann_task_queue {
     }
 
     std::vector<std::unique_ptr<cann_task>> buffer_;
-    const size_t                            capacity_;
-    size_t                                  mask_;
-    size_t                                  head_;
-    size_t                                  tail_;
-    bool                                    running_;
-    std::thread                             thread_;
-    int32_t                                 device_;
+    const size_t capacity_;
+    size_t mask_;
+    size_t head_;
+    size_t tail_;
+    bool running_;
+    std::thread thread_;
+    int32_t device_;
 };
 
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
-    // dst tensor
-    void *  node_address;
-    int64_t ne[GGML_MAX_DIMS];
-    size_t  nb[GGML_MAX_DIMS];
-
-    // src tensor
-    void *  src_address[GGML_MAX_SRC];
-    int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
-    size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
-
-    // op
+    void * node_address;
     ggml_op node_op;
+    int64_t ne[GGML_MAX_DIMS];
+    size_t nb[GGML_MAX_DIMS];
+    void * src_address[GGML_MAX_SRC];
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 
@@ -375,11 +369,13 @@ struct ggml_cann_graph {
  * move existing graphs to the front (most recently used), and clear the cache.
  */
 struct ggml_cann_graph_lru_cache {
-    size_t capacity;                         /**< Maximum number of graphs in the cache. */
+    size_t capacity;  /**< Maximum number of graphs in the cache. */
 
-    std::list<ggml_cann_graph *> cache_list; /**< List storing cached graphs as raw pointers. */
+    std::list<ggml_cann_graph*> cache_list; /**< List storing cached graphs as raw pointers. */
 
-    ggml_cann_graph_lru_cache() { capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12")); }
+    ggml_cann_graph_lru_cache() {
+        capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12"));
+    }
 
     /**
      * @brief Push a new graph to the front of the cache.
@@ -387,11 +383,11 @@ struct ggml_cann_graph_lru_cache {
      * @param new_node Pointer to the new ggml_cann_graph to cache.
      *        Ownership is transferred to the cache (cache will delete it).
      */
-    void push(ggml_cann_graph * new_node) {
+    void push(ggml_cann_graph* new_node) {
         if (cache_list.size() >= capacity) {
-            ggml_cann_graph * old = cache_list.back();
+            ggml_cann_graph* old = cache_list.back();
             cache_list.pop_back();
-            delete old;  // free the old graph
+            delete old; // free the old graph
         }
         cache_list.push_front(new_node);
     }
@@ -400,7 +396,7 @@ struct ggml_cann_graph_lru_cache {
      * @brief Move an existing graph to the front of the cache.
      * @param node Pointer to the ggml_cann_graph to move.
      */
-    void move_to_front(ggml_cann_graph * node) {
+    void move_to_front(ggml_cann_graph* node) {
         cache_list.remove(node);
         cache_list.push_front(node);
     }
@@ -418,89 +414,92 @@ struct ggml_cann_graph_lru_cache {
     /**
      * @brief Destructor that clears the cache and frees all cached graphs.
      */
-    ~ggml_cann_graph_lru_cache() { clear(); }
+    ~ggml_cann_graph_lru_cache() {
+        clear();
+    }
 };
 #endif  // USE_ACL_GRAPH
 
 struct ggml_cann_rope_cache {
     ~ggml_cann_rope_cache() {
-        if (theta_scale_cache != nullptr) {
+        if(theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
-        if (sin_cache != nullptr) {
+        if(sin_cache != nullptr) {
             ACL_CHECK(aclrtFree(sin_cache));
         }
-        if (cos_cache != nullptr) {
+        if(cos_cache != nullptr) {
             ACL_CHECK(aclrtFree(cos_cache));
         }
     }
 
-    void *  theta_scale_cache  = nullptr;
+    void* theta_scale_cache = nullptr;
     int64_t theta_scale_length = 0;
     // sin/cos cache, used only to accelerate first layer on each device
-    void *  sin_cache          = nullptr;
-    void *  cos_cache          = nullptr;
-    int64_t position_length    = 0;
+    void* sin_cache = nullptr;
+    void* cos_cache = nullptr;
+    int64_t position_length = 0;
     // Properties to check before reusing the sincos cache
-    bool    cached             = false;
-    float   ext_factor         = 0.0f;
-    float   theta_scale        = 0.0f;
-    float   freq_scale         = 0.0f;
-    float   attn_factor        = 0.0f;
-    bool    is_neox            = false;
+    bool cached = false;
+    float ext_factor = 0.0f;
+    float theta_scale = 0.0f;
+    float freq_scale = 0.0f;
+    float attn_factor = 0.0f;
+    bool is_neox = false;
 };
 
 struct ggml_cann_tensor_cache {
     ~ggml_cann_tensor_cache() {
-        if (cache != nullptr) {
+        if(cache != nullptr) {
             ACL_CHECK(aclrtFree(cache));
         }
     }
 
-    void *  cache = nullptr;
-    int64_t size  = 0;
+    void* cache = nullptr;
+    int64_t size = 0;
 };
 
 /**
  * @brief Context for managing CANN backend operations.
  */
 struct ggml_backend_cann_context {
-    int32_t     device;               /**< Device ID. */
-    std::string name;                 /**< Name of the device. */
-    std::string description;          /**< Description of the device. */
-    aclrtEvent  copy_event = nullptr; /**< Event for managing copy operations. */
+    int32_t device;                  /**< Device ID. */
+    std::string name;                /**< Name of the device. */
+    std::string description;         /**< Description of the device. */
+    aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
     ggml_cann_graph_lru_cache graph_lru_cache;
-    bool                      acl_graph_mode = true;
+    bool acl_graph_mode = true;
 #endif
-    cann_task_queue        task_queue;
-    bool                   async_mode;
+    cann_task_queue task_queue;
+    bool async_mode;
     // Rope Cache
-    ggml_cann_rope_cache   rope_cache;
+    ggml_cann_rope_cache rope_cache;
     // Constant Pool
     ggml_cann_tensor_cache rms_norm_one_tensor_cache;
     ggml_cann_tensor_cache rms_norm_zero_tensor_cache;
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = { nullptr }; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
-    explicit ggml_backend_cann_context(int device) :
-        device(device),
-        name("CANN" + std::to_string(device)),
-        task_queue(1024, device) {
+    explicit ggml_backend_cann_context(int device)
+        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
 
         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");
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
+            device, async_mode ? "ON" : "OFF");
 #ifdef USE_ACL_GRAPH
         acl_graph_mode = parse_bool(get_env("GGML_CANN_ACL_GRAPH").value_or("on"));
-        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n", __func__, device, acl_graph_mode ? "GRAPH" : "EAGER",
-                      acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
+        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n",
+              __func__, device,
+              acl_graph_mode ? "GRAPH" : "EAGER",
+              acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
 #endif
     }
 
@@ -543,7 +542,8 @@ struct ggml_backend_cann_context {
     aclrtStream stream() { return stream(0); }
 
     // TODO: each stream should have a memory pool.
-    std::unique_ptr<ggml_cann_pool> mem_pool; /**< Memory pool for the device. */
+    std::unique_ptr<ggml_cann_pool>
+        mem_pool; /**< Memory pool for the device. */
 
     /**
      * @brief Create a new memory pool for a given device.
@@ -556,7 +556,7 @@ struct ggml_backend_cann_context {
      * @brief Get or create the memory pool for the context.
      * @return Reference to the memory pool.
      */
-    ggml_cann_pool & pool() {
+    ggml_cann_pool& pool() {
         if (mem_pool == nullptr) {
             mem_pool = new_pool_for_device(device);
         }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -439,15 +439,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ggml-cpu/arch/riscv/quants.c
             ggml-cpu/arch/riscv/repack.cpp
             )
-        if (GGML_CPU_RISCV64_SPACEMIT)
-            target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_RISCV64_SPACEMIT ${RISCV64_SPACEMIT_IME_SPEC})
-            list(APPEND GGML_CPU_SOURCES
-                ggml-cpu/spacemit/ime.cpp
-                ggml-cpu/spacemit/ime.h
-                ggml-cpu/spacemit/ime1_kernels.cpp
-                ggml-cpu/spacemit/ime_kernels.h
-            )
-        endif()
         set(MARCH_STR "rv64gc")
         if (GGML_RV_ZFH)
             string(APPEND MARCH_STR "_zfh")
@@ -466,45 +457,29 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
-        list(APPEND GGML_CPU_SOURCES
-            ggml-cpu/arch/s390/quants.c)
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
+        file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
+        string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
 
-        # for native compilation
-        if (GGML_NATIVE)
-            # check machine level to determine target
-            file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
-            string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
-
-            # TODO: Separation to determine activation of VX/VXE/VXE2
-            if (${S390X_M} MATCHES "8561|8562")
-                message(STATUS "z15 target")
-                list(APPEND ARCH_FLAGS -march=z15)
-            elseif (${S390X_M} MATCHES "3931")
-                message(STATUS "z16 target")
-                list(APPEND ARCH_FLAGS -march=z16)
-            elseif (${S390X_M} MATCHES "9175|9176")
-                # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
-                #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
-                message(STATUS "z17 target")
-                list(APPEND ARCH_FLAGS -march=arch15)
-            else()
-                message(STATUS "Unknown target")
-                message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
-                list(APPEND ARCH_FLAGS -march=native -mtune=native)
-            endif()
-        # for cross-compilation
-        elseif(GGML_CPU_ALL_VARIANTS)
-            # range through IBM z15 to z17
-            # NOTE: update when a new hardware level is released
-            foreach (ZHW RANGE 15 17)
-                if(DEFINED GGML_INTERNAL_Z${ZHW})
-                    message(STATUS "z${ZHW} cross-compile target")
-                    list(APPEND ARCH_FLAGS -march=z${ZHW})
-                endif()
-            endforeach()
+        # TODO: Separation to determine activation of VX/VXE/VXE2
+        if (${S390X_M} MATCHES "8561|8562")
+            message(STATUS "z15 target")
+            list(APPEND ARCH_FLAGS -march=z15)
+        elseif (${S390X_M} MATCHES "3931")
+            message(STATUS "z16 target")
+            list(APPEND ARCH_FLAGS -march=z16)
+        elseif (${S390X_M} MATCHES "9175|9176")
+            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
+            message(STATUS "z17 target")
+            list(APPEND ARCH_FLAGS -march=arch15)
+        else()
+            message(STATUS "Unknown target")
+            message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
+            list(APPEND ARCH_FLAGS -march=native -mtune=native)
         endif()
 
-        if (GGML_VXE OR GGML_INTERNAL_VXE)
+        if (GGML_VXE)
             message(STATUS "VX/VXE/VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
             list(APPEND ARCH_DEFINITIONS GGML_VXE)
@@ -529,9 +504,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.14.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.13.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "45e110675d93f99f82c23a1afcca76bc")
+        set(KLEIDIAI_ARCHIVE_MD5  "d82a8de939d9814621a5ba23907bdac1")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -608,7 +583,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
-                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
                 ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)

ggml/src/ggml-cpu/amx/amx.cpp

@@ -149,7 +149,6 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
         if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) &&  // src0 must be contiguous
             is_contiguous_2d(op->src[1]) &&                               // src1 must be contiguous
             op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
-            op->src[0]->ne[0] % (TILE_K * 2 * 32) == 0 && // TODO: not sure if correct (https://github.com/ggml-org/llama.cpp/pull/16315)
             op->ne[0] % (TILE_N * 2) == 0 &&                              // out_features is 32x
             (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
             // src1 must be host buffer

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -68,7 +68,7 @@ struct ggml_compute_params {
 #endif  // __VXE2__
 #endif  // __s390x__ && __VEC__
 
-#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
+#if defined(__ARM_FEATURE_SVE)
 #include <sys/prctl.h>
 #endif
 

ggml/src/ggml-cpu/ggml-cpu.c

@@ -689,13 +689,8 @@ bool ggml_is_numa(void) {
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
-#if defined(__linux__)
+#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);
-#else
-    // TODO: add support of SVE for non-linux systems
-#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
-#endif
 #endif
 }
 
@@ -1613,8 +1608,13 @@ static void ggml_compute_forward_mul_mat_id(
             chunk_size = 64;
         }
 
+#if defined(__aarch64__)
+        // disable for ARM
+        const bool disable_chunking = true;
+#else
         // disable for NUMA
         const bool disable_chunking = ggml_is_numa();
+#endif // defined(__aarch64__)
 
         int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
         int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
@@ -2179,10 +2179,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
-                case GGML_UNARY_OP_FLOOR:
-                case GGML_UNARY_OP_CEIL:
-                case GGML_UNARY_OP_ROUND:
-                case GGML_UNARY_OP_TRUNC:
                     {
                         n_tasks = 1;
                     } break;
@@ -2191,7 +2187,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
-                case GGML_UNARY_OP_XIELU:
                     {
                         n_tasks = n_threads;
                     } break;
@@ -3562,17 +3557,13 @@ void ggml_cpu_init(void) {
 #ifdef GGML_USE_OPENMP
             //if (!getenv("OMP_WAIT_POLICY")) {
             //    // set the wait policy to active, so that OpenMP threads don't sleep
-            //    setenv("OMP_WAIT_POLICY", "active", 0)
+            //    putenv("OMP_WAIT_POLICY=active");
             //}
 
             if (!getenv("KMP_BLOCKTIME")) {
                 // set the time to wait before sleeping a thread
                 // this is less aggressive than setting the wait policy to active, but should achieve similar results in most cases
-#ifdef _WIN32
-                _putenv_s("KMP_BLOCKTIME", "200"); // 200ms
-#else
-                setenv("KMP_BLOCKTIME", "200", 0); // 200ms
-#endif
+                putenv("KMP_BLOCKTIME=200"); // 200ms
             }
 #endif
         }

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -18,10 +18,6 @@
 #    include "kleidiai/kleidiai.h"
 #endif
 
-#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
-#    include "spacemit/ime.h"
-#endif
-
 #if defined(_WIN32)
 #    define WIN32_LEAN_AND_MEAN
 #    ifndef NOMINMAX
@@ -49,12 +45,6 @@ std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
-        if (ggml_backend_cpu_riscv64_spacemit_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_riscv64_spacemit_buffer_type());
-        }
-#endif
-
 #ifdef GGML_USE_CPU_KLEIDIAI
         if (ggml_backend_cpu_kleidiai_buffer_type()) {
             bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -29,108 +29,6 @@
 
 #define NELEMS(x) sizeof(x) / sizeof(*x)
 
-template<size_t(*Fn)(size_t,size_t,size_t)>
-static inline size_t kernel_offs_fn3(size_t a, size_t b, size_t c) {
-    return Fn(a, b, c);
-}
-
-template<size_t(*Fn)(size_t,size_t)>
-static inline size_t kernel_offs_fn2(size_t a, size_t b, size_t) {
-    return Fn(a, b);
-}
-
-template<void(*Fn)(size_t,size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
-static inline void kernel_run_fn11(size_t m, size_t n, size_t k, size_t bl,
-                                     const void* lhs, const void* rhs, void* dst,
-                                     size_t dst_stride_row, size_t dst_stride_col,
-                                     float clamp_min, float clamp_max) {
-    Fn(m, n, k, bl, lhs, rhs, static_cast<float*>(dst), dst_stride_row, dst_stride_col, clamp_min, clamp_max);
-}
-
-template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,void*,size_t,size_t,float,float)>
-static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
-                                   const void* lhs, const void* rhs, void* dst,
-                                   size_t dst_stride_row, size_t dst_stride_col,
-                                   float clamp_min, float clamp_max) {
-    Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
-}
-
-template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
-static inline size_t lhs_ps_fn6(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
-    return Fn(m, k, bl, mr, kr, sr);
-}
-
-template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
-static inline size_t lhs_ps_fn5(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr) {
-    return Fn(m, k, mr, kr, sr);
-}
-
-template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
-static inline size_t lhs_offs_fn6(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
-    return Fn(m_idx, k, bl, mr, kr, sr);
-}
-
-template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
-static inline size_t lhs_offs_fn5(size_t m_idx, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr) {
-    return Fn(m_idx, k, mr, kr, sr);
-}
-
-template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const float*,size_t,void*)>
-static inline void lhs_pack_float_fn10(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
-                                            size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
-    Fn(m, k, bl, mr, kr, sr, m_idx_start, static_cast<const float*>(lhs), lhs_stride, lhs_packed);
-}
-
-template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,size_t,void*)>
-static inline void lhs_pack_void_fn10(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
-                                           size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
-    Fn(m, k, bl, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
-}
-
-template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const void*,size_t,void*)>
-static inline void lhs_pack_void_fn9(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr,
-                                             size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
-    Fn(m, k, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
-}
-
-template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
-static inline size_t rhs_ps_fn5(size_t n, size_t k, size_t nr, size_t kr, size_t bl) {
-    return Fn(n, k, nr, kr, bl);
-}
-
-template<size_t(*Fn)(size_t,size_t)>
-static inline size_t rhs_ps_fn2(size_t n, size_t k, size_t /*nr*/, size_t /*kr*/, size_t /*bl*/) {
-    return Fn(n, k);
-}
-
-template<size_t(*Fn)(size_t,size_t,size_t,size_t)>
-static inline size_t rhs_stride_fn4(size_t k, size_t nr, size_t kr, size_t bl) {
-    return Fn(k, nr, kr, bl);
-}
-
-template<size_t(*Fn)(size_t)>
-static inline size_t rhs_stride_fn1(size_t k, size_t /*nr*/, size_t /*kr*/, size_t /*bl*/) {
-    return Fn(k);
-}
-
-template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const uint8_t*,const float*,void*,size_t,const struct kai_rhs_pack_qs4cxs1s0_param*)>
-static inline void rhs_pack_fn12(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
-                                      size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* /*scale*/,
-                                      void* rhs_packed, size_t extra_bytes, const void* params) {
-    Fn(num_groups, n, k, nr, kr, sr, bl,
-       static_cast<const uint8_t*>(rhs),
-       static_cast<const float*>(bias),
-       rhs_packed, extra_bytes,
-       static_cast<const kai_rhs_pack_qs4cxs1s0_param*>(params));
-}
-
-template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,const void*,const void*,void*,size_t,const void*)>
-static inline void rhs_pack_fn13(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
-                                               size_t rhs_stride, const void* rhs, const void* bias, const void* scale,
-                                               void* rhs_packed, size_t extra_bytes, const void* params) {
-    Fn(num_groups, n, k, nr, kr, sr, rhs_stride, rhs, bias, scale, rhs_packed, extra_bytes, params);
-}
-
 static const size_t INT4_PER_BYTE = 2;
 static const size_t INT4_BITS     = 4;
 static const int Q4_0_ZERO_POINT  = 8;
@@ -224,18 +122,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
         },
-
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
         },
         /* SME GEMV */
         /* .kern_info = */ {
@@ -245,24 +142,23 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
         },
         /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
-            /* .to_float              = */ dequantize_row_qsi4c32ps1s0scalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .to_float      = */ dequantize_row_qsi4c32ps1s0scalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -278,17 +174,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
-            /* .run_kernel_ex         = */ &kernel_run_fn10<kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme>,
-            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
-            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
         },
         /* SME GEMV */
         /* .kern_info = */ {
@@ -298,24 +194,23 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ nullptr,
-            /* .get_rhs_packed_offset_ex = */ nullptr,
-            /* .run_kernel_ex         = */ nullptr,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme>,
-            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
-            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
         },
         /* .rhs_info = */ {
-            /* .packed_stride         = */ nullptr,
-            /* .to_float              = */ nullptr,
-            /* .packed_size_ex        = */ &rhs_ps_fn2<kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn1<kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
-            /* .pack_func_ex          = */ &rhs_pack_fn13<kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .packed_stride = */ NULL,
+            /* .pack_func     = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .to_float      = */ NULL,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -334,17 +229,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* DOTPROD GEMV */
         /* .kern_info = */ {
@@ -354,24 +249,23 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -389,17 +283,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
         },
         /* i8mm GEMV */
         /* .kern_info = */ {
@@ -409,24 +303,23 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -445,17 +338,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
         },
         /* i8mm GEMV */
         /* .kern_info = */ {
@@ -465,24 +358,23 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -500,17 +392,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* DOTPROD GEMV */
         /* .kern_info = */ {
@@ -520,24 +412,23 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
+            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -552,7 +443,6 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
     ggml_kleidiai_kernels * kernel = nullptr;
 
     if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
         for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
             if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
                 gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
@@ -562,7 +452,6 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
                 break;
             }
         }
-#endif
     }
 
     return kernel;
@@ -571,14 +460,12 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
         if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
             kernels = &gemm_gemv_kernels[i];
             break;
         }
     }
-#endif
 
     return kernels;
 }

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -4,6 +4,8 @@
 
 #pragma once
 
+#include <functional>
+#include <variant>
 #include "ggml.h"
 
 enum cpu_feature {
@@ -13,7 +15,6 @@ enum cpu_feature {
     CPU_FEATURE_SVE     = 4,
     CPU_FEATURE_SME     = 8
 };
-
 inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
     lhs = static_cast<cpu_feature>(lhs | rhs);
     return lhs;
@@ -29,52 +30,63 @@ struct kernel_info {
     size_t (*get_nr)(void);
     size_t (*get_kr)(void);
     size_t (*get_sr)(void);
-
+    std::variant<
+        std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
+        std::function<size_t(size_t m_idx, size_t k)>
+    > get_lhs_offset;
+    std::variant<
+        std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
+        std::function<size_t(size_t n_idx, size_t k)>
+    > get_rhs_packed_offset;
     size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
     size_t (*get_dst_size)(size_t m, size_t n);
-
-    size_t (*get_lhs_offset_ex)(size_t m_idx, size_t k, size_t bl);
-
-    size_t (*get_rhs_packed_offset_ex)(size_t n_idx, size_t k, size_t bl);
-
-    void (*run_kernel_ex)(
-        size_t m, size_t n, size_t k, size_t bl,
-        const void* lhs_packed, const void* rhs_packed,
-        void* dst, size_t dst_stride_row, size_t dst_stride_col,
-        float clamp_min, float clamp_max);
+    std::variant<
+        std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
+            float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
+        std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
+            size_t dst_stride_col, float clamp_min, float clamp_max)>
+    > run_kernel;
 };
 
 struct lhs_packing_info {
     size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
-
-    size_t (*get_packed_offset_ex)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-
-    size_t (*packed_size_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-
-    void (*pack_func_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
-        size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed);
+    std::variant<
+        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
+        std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
+    > get_packed_offset;
+    std::variant<
+        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
+        std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
+    > packed_size;
+    std::variant<
+        std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
+            size_t lhs_stride, void* lhs_packed)>,
+        std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
+        void* lhs_packed)>
+    > pack_func;
 };
 
 struct rhs_packing_info {
+    std::variant<
+        std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
+        std::function<size_t(size_t n, size_t k)>
+    > packed_size;
     size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
-
-    void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out,
-                     size_t nr_pack, size_t packed_row_stride, size_t kr, size_t bl,
-                     size_t num_bytes_multiplier);
-
-    size_t (*packed_size_ex)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
-
-    size_t (*packed_stride_ex)(size_t k, size_t nr, size_t kr, size_t bl);
-
-    void (*pack_func_ex)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
-        size_t rhs_stride, const void * rhs, const void * bias, const void * scale, void * rhs_packed, size_t extra_bytes, const void * params);
+    std::variant<
+        std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
+            const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
+        std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
+            const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
+    > pack_func;
+    void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out, size_t nr_pack, size_t packed_row_stride,
+          size_t kr, size_t bl, size_t num_bytes_multiplier);
 };
 
 struct ggml_kleidiai_kernels {
-    kernel_info      gemm;
+    kernel_info gemm;
     lhs_packing_info gemm_lhs_info;
 
-    kernel_info      gemv;
+    kernel_info gemv;
     lhs_packing_info gemv_lhs_info;
 
     rhs_packing_info rhs_info;

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -8,7 +8,6 @@
 #include <stdexcept>
 #include <stdint.h>
 #include <string.h>
-#include <string>
 #if defined(__linux__)
 #include <asm/hwcap.h>
 #include <sys/auxv.h>
@@ -88,6 +87,17 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
     return tensor->ne[dim];
 }
 
+template<typename Ret, typename Variant, typename... Args>
+static Ret variant_call(const Variant & var, Args&&... args) {
+    return std::visit([&](auto&& func) -> Ret {
+        if constexpr (std::is_invocable_r_v<Ret, decltype(func), Args...>) {
+            return func(std::forward<Args>(args)...);
+        } else {
+            throw std::runtime_error("Invalid function type in variant_call");
+        }
+    }, var);
+}
+
 namespace ggml::cpu::kleidiai {
 
 static size_t round_down(size_t x, size_t y) {
@@ -112,9 +122,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             return false;
         }
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
-        if (!kernels) {
-            return false;
-        }
+        GGML_ASSERT(kernels);
         bool is_gemv = op->src[1]->ne[1] == 1;
         kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
@@ -128,23 +136,19 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         size_t sr = kernel->get_sr();
 
         if (kernels->rhs_type == GGML_TYPE_Q4_0) {
-            if (!lhs_info->packed_size_ex) return false;
-            size = lhs_info->packed_size_ex(m, k, QK4_0, mr, kr, sr);
+            size = variant_call<size_t>(lhs_info->packed_size, m, k, QK4_0, mr, kr, sr);
         } else if (kernels->rhs_type == GGML_TYPE_F16) {
-            if (!lhs_info->packed_size_ex || !kernels->rhs_info.packed_size_ex) return false;
-            const int64_t lhs_batch_size0 = op->src[1]->ne[2];
-            const int64_t rhs_batch_size0 = op->src[0]->ne[2];
-            const int64_t r = lhs_batch_size0 / rhs_batch_size0;
-            size = lhs_info->packed_size_ex(m * r, k, 0, mr, kr, sr) +
-                   kernels->rhs_info.packed_size_ex(n, k, kernel->get_nr(), kernel->get_kr(), 0) +
+            size = variant_call<size_t>(lhs_info->packed_size, m, k, mr, kr, sr) +
+                   variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
                    k * n * sizeof(float) + n * sizeof(float);
         } else {
-            return false;
+            GGML_ASSERT(false);
         }
 
         return true;
     }
 
+
     bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
         if (dst->op == GGML_OP_MUL_MAT) {
             if (dst->src[0]->type == GGML_TYPE_Q4_0) {
@@ -161,52 +165,45 @@ class tensor_traits : public ggml::cpu::tensor_traits {
     }
 
     bool compute_forward_fp16(ggml_compute_params * params, struct ggml_tensor * dst) {
+        static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
+
         const ggml_tensor * src0 = dst->src[0];
         const ggml_tensor * src1 = dst->src[1];
 
         GGML_TENSOR_BINARY_OP_LOCALS
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
-        if (!kernels) {
-            return false;
-        }
+        GGML_ASSERT(kernels);
 
-        const bool is_gemv = src1->ne[1] == 1;
+        bool is_gemv = src1->ne[1] == 1;
         kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
         GGML_ASSERT(kernel);
-        if (!kernels->rhs_info.pack_func_ex ||
-            !kernel->get_lhs_offset_ex || !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex) {
-            return false;
-        }
 
         const int nth = params->nth;
         const int ith = params->ith;
 
         const int64_t lhs_batch_size0 = ne12;
         const int64_t rhs_batch_size0 = ne02;
-        const int64_t batch_size      = lhs_batch_size0;
+        const int64_t batch_size      = rhs_batch_size0;
 
-        GGML_ASSERT(rhs_batch_size0 > 0);
-        GGML_ASSERT(lhs_batch_size0 % rhs_batch_size0 == 0);
         const int64_t r = lhs_batch_size0 / rhs_batch_size0;
 
-        const int64_t m_group = ne11;
-        const int64_t m       = m_group;
-        const int64_t n       = ne01;
-        const int64_t k       = ne00;
+        const int64_t m = ne11 * r;
+        const int64_t n = ne01;
+        const int64_t k = ne00;
 
         const size_t lhs_stride = src1->nb[1];
         const size_t rhs_stride = src0->nb[1];
         const size_t dst_stride = dst->nb[1];
 
-        const int64_t mr = (int64_t) kernel->get_mr();
-        const int64_t nr = (int64_t) kernel->get_nr();
-        const int64_t kr = (int64_t) kernel->get_kr();
-        const int64_t sr = (int64_t) kernel->get_sr();
+        const int64_t mr = static_cast<int64_t>(kernel->get_mr());
+        const int64_t nr = static_cast<int64_t>(kernel->get_nr());
+        const int64_t kr = static_cast<int64_t>(kernel->get_kr());
+        const int64_t sr = static_cast<int64_t>(kernel->get_sr());
 
-        const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
-        const size_t rhs_packed_size = kernels->rhs_info.packed_size_ex(n, k, nr, kr, 0);
+        const size_t lhs_packed_size = variant_call<size_t>(lhs_info->packed_size, m, k, mr, kr, sr);
+        const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, n, k);
         const size_t kxn_size        = k * n * sizeof(float);
         const size_t bias_size       = n * sizeof(float);
 
@@ -219,91 +216,82 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         uint8_t * bias       = rhs_kxn + kxn_size;
 
         for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
-            const int64_t rhs_batch_idx = batch_idx / r;
-            const uint8_t * rhs_batch_base = static_cast<const uint8_t *>(src0->data) + rhs_batch_idx * src0->nb[2];
-            uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];
+            const uint8_t * lhs_batch = static_cast<const uint8_t *>(src1->data) + batch_idx * m * lhs_stride;
+            const uint8_t * rhs_batch = static_cast<const uint8_t *>(src0->data) + batch_idx * n * rhs_stride;
+            uint8_t * dst_batch       = static_cast<uint8_t *>(dst->data) + batch_idx * m * dst_stride;
 
-            // LHS packing (threaded over m, honoring mr alignment and KV groups)
+            // LHS packing
             {
                 const int64_t m_roundup_mr = kai_roundup(m, mr);
                 const int64_t num_threads  = KAI_MIN(m_roundup_mr / mr, nth);
 
                 if (ith < num_threads) {
-                    const int64_t num_m_per_thread0   = round_down((size_t)(m_roundup_mr / num_threads), (size_t)mr);
+                    const int64_t num_m_per_thread0   = round_down(m_roundup_mr / num_threads, mr);
                     const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
 
-                    const int64_t m_start = ith * num_m_per_thread0;
-                    const int64_t m_count = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
+                    const int64_t m_start          = ith * num_m_per_thread0;
+                    const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
 
-                    // Base packed offset (aligned) and per-row stride in bytes
-                    const size_t base_packed_off  = lhs_info->get_packed_offset_ex(m_start, k, 0, mr, kr, sr);
-                    const size_t next_block_off   = lhs_info->get_packed_offset_ex(m_start + mr, k, 0, mr, kr, sr);
-                    const size_t row_stride_bytes = (next_block_off - base_packed_off) / (size_t)mr;
+                    const size_t lhs_offset        = variant_call<size_t>(kernels->gemm.get_lhs_offset, m_start, lhs_stride);
+                    const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, mr, kr, sr);
 
-                    int64_t remaining = m_count;
-                    int64_t cur       = m_start;
+                    const void * src_ptr = static_cast<const uint8_t *>(lhs_batch) + lhs_offset;
+                    void * dst_ptr       = static_cast<uint8_t *>(lhs_packed) + lhs_packed_offset;
 
-                    while (remaining > 0) {
-                        const int64_t row_in_group = cur;
-                        const int64_t avail        = m_group - row_in_group;
-                        const int64_t take         = std::min(avail, remaining);
-
-                        const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
-                        const void * src_ptr = lhs_batch_base + (size_t)row_in_group * lhs_stride;
-                        const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
-                        void * dst_ptr       = lhs_packed + dst_off;
-
-                        lhs_info->pack_func_ex(take, k, 0, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
-
-                        cur       += take;
-                        remaining -= take;
-                    }
+                    variant_call<void>(lhs_info->pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
                 }
             }
 
-            // RHS packing (single thread), then synchronize
-            if (ith == 0) {
-                memset(bias, 0, (size_t)n * sizeof(float));
-                transpose_f32kxn_f16nxk((size_t)n, (size_t)k,
-                                        reinterpret_cast<float *>(rhs_kxn),
-                                        reinterpret_cast<const uint16_t *>(rhs_batch_base),
-                                        rhs_stride);
+            // RHS packing
+            if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) {
+                // First thread to reach this point handles RHS packing
+                memset(bias, 0, n * sizeof(float));
+                transpose_f32kxn_f16nxk(n, k, reinterpret_cast<float *>(rhs_kxn),
+                                        reinterpret_cast<const uint16_t *>(rhs_batch), rhs_stride);
 
-                kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, 0, n * sizeof(float),
+                variant_call<void>(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float),
                              rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
             }
 
             ggml_barrier(params->threadpool);
 
-            // Matmul (threaded over n)
+            first_to_arrive.clear(std::memory_order_release);
+
+            // Perform the matmul
             {
-                const int64_t n_step  = (int64_t) kernel->get_n_step();
-                int64_t num_threads_n = KAI_MIN(n / n_step, nth);
-                if (num_threads_n <= 0) {
-                    num_threads_n = 1;
+                const int64_t m_to_process = m;
+                const int64_t m_start      = 0;
+
+                const int64_t n_step      = static_cast<int64_t>(kernel->get_n_step());
+                int64_t num_threads       = KAI_MIN(n / n_step, nth);
+                if (num_threads <= 0) {
+                    num_threads = 1;
                 }
 
-                if (ith < num_threads_n) {
-                    const int64_t num_n_per_thread0   = round_down((size_t)(n / num_threads_n), (size_t)n_step);
-                    const int64_t num_n_per_threadN_1 = n - (num_threads_n - 1) * num_n_per_thread0;
+                if (ith < num_threads) {
+                    const int64_t num_n_per_thread0   = round_down(n / num_threads, n_step);
+                    const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0;
 
                     const int64_t n_start      = ith * num_n_per_thread0;
-                    const int64_t n_to_process = (ith == num_threads_n - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
+                    const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
 
-                    // LHS packed base at row 0 (consistent with packing above)
-                    const size_t lhs_packed_offset0 = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
-                    const size_t rhs_packed_offset  = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
-                    const size_t dst_offset         = kernel->get_dst_offset((size_t)0, (size_t)n_start, dst_stride);
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernel->get_lhs_offset, m_start, k);
+                    const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k);
+                    const size_t dst_offset        = kernel->get_dst_offset(m_start, n_start, dst_stride);
 
-                    const void * lhs_ptr = lhs_packed + lhs_packed_offset0;
+                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
                     const void * rhs_ptr = rhs_packed + rhs_packed_offset;
-                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);
+                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch + dst_offset);
 
-                    kernel->run_kernel_ex(m, n_to_process, k, 0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
+                    variant_call<void>(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
                 }
             }
 
             if (batch_idx != batch_size - 1) {
+                // This barrier is necessary when the batch size is larger than 1. While processing a batch,
+                // the work data buffer (params->wdata) is used as temporary storage which means that only
+                // a single batch can be processed at any given time. No barrier is needed for the last
+                // batch since GGML inserts a barrier between the execution of every operator.
                 ggml_barrier(params->threadpool);
             }
         }
@@ -320,19 +308,13 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_TENSOR_BINARY_OP_LOCALS
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
-        if (!kernels) {
-            return false;
-        }
+        GGML_ASSERT(kernels);
 
         bool is_gemv = src1->ne[1] == 1;
         kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
 
         GGML_ASSERT(kernel);
-        if (!lhs_info->get_packed_offset_ex || !lhs_info->pack_func_ex ||
-            !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
-            return false;
-        }
 
         const int ith = params->ith;
         const int nth_raw = params->nth;
@@ -374,26 +356,25 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             // Transform LHS
             const size_t src_stride        = src1->nb[1];
             const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
-            const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(m_start, k, QK4_0, mr, kr, sr);
+            const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, QK4_0, mr, kr, sr);
             void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
 
-            // Pack this thread's chunk with m_idx_start = 0 and per-thread output pointer
-            lhs_info->pack_func_ex(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+            variant_call<void>(lhs_info->pack_func, m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
         }
 
         ggml_barrier(params->threadpool);
 
         // Perform the operation
         const size_t dst_stride        = dst->nb[1];
-        const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, QK4_0, mr, kr, sr);
-        const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, QK4_0);
+        const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, 0, k, QK4_0, mr, kr, sr);
+        const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k, QK4_0);
         const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
         const void * rhs_ptr           = static_cast<const void *>(rhs_packed + rhs_packed_offset);
         const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
         float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
 
         if (n_to_process > 0) {
-            kernel->run_kernel_ex(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+            variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
                                sizeof(float), -FLT_MAX, FLT_MAX);
         }
 
@@ -402,9 +383,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
 
     bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
         GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
-        if (!ctx.kernels) {
-            return false;
-        }
+        GGML_ASSERT(ctx.kernels);
 
         const ggml_tensor * src0 = dst->src[0];
         const ggml_tensor * src1 = dst->src[1];
@@ -413,9 +392,6 @@ class tensor_traits : public ggml::cpu::tensor_traits {
 
         rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
         kernel_info * kernel        = &ctx.kernels->gemm;
-        if (!rhs_info->to_float || !kernel->get_nr) {
-            return false;
-        }
 
         const int64_t nc     = ne00;
         const int64_t nr     = ggml_nelements(src1);
@@ -458,7 +434,7 @@ public:
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
-        ctx.kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, QK4_0, 0, (const uint8_t*)data, nullptr, nullptr, tensor->data, 0, &params);
+        variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
 
         return 0;
         GGML_UNUSED(data_size);
@@ -526,7 +502,7 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_
     const size_t nr = ctx.kernels->gemm.get_nr();
     const size_t kr = ctx.kernels->gemm.get_kr();
 
-    return ctx.kernels->rhs_info.packed_size_ex(n, k, nr, kr, QK4_0);
+    return variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
 
     GGML_UNUSED(buft);
 }

ggml/src/ggml-cpu/ops.cpp

@@ -3467,27 +3467,31 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_ASSERT(eps >= 0.0f);
 
+    // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
                 const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
 
-                float sum = 0.0;
-                ggml_vec_sum_f32(ne00, &sum, x);
+                ggml_float sum = 0.0;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    sum += (ggml_float)x[i00];
+                }
+
                 float mean = sum/ne00;
 
                 float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-                float variance = 0;
 
-#ifdef GGML_USE_ACCELERATE
-                mean = -mean;
-                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
-                vDSP_measqv(y, 1, &variance, ne00);
-#else
-                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
-#endif //GGML_USE_ACCELERATE
+                ggml_float sum2 = 0.0;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    float v = x[i00] - mean;
+                    y[i00] = v;
+                    sum2 += (ggml_float)(v*v);
+                }
 
+                float variance = sum2/ne00;
                 const float scale = 1.0f/sqrtf(variance + eps);
+
                 ggml_vec_scale_f32(ne00, y, scale);
             }
         }
@@ -5474,7 +5478,7 @@ static void ggml_rope_cache_init(
 }
 
 static void ggml_mrope_cache_init(
-     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool is_imrope, bool indep_sects,
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
      float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
@@ -5509,26 +5513,14 @@ static void ggml_mrope_cache_init(
         }
 
         float theta = theta_t;
-        if (is_imrope) { // qwen3vl apply interleaved mrope
-            if (sector % 3 == 1 && sector < 3 * sections[1]) {
-                theta = theta_h;
-            } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
-                theta = theta_w;
-            } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
-                theta = theta_t;
-            } else {
-                theta = theta_e;
-            }
-        } else {
-            if (sector >= sections[0] && sector < sec_w) {
-                theta = theta_h;
-            }
-            else if (sector >= sec_w && sector < sec_w + sections[2]) {
-                theta = theta_w;
-            }
-            else if (sector >= sec_w + sections[2]) {
-                theta = theta_e;
-            }
+        if (sector >= sections[0] && sector < sec_w) {
+            theta = theta_h;
+        }
+        else if (sector >= sec_w && sector < sec_w + sections[2]) {
+            theta = theta_w;
+        }
+        else if (sector >= sec_w + sections[2]) {
+            theta = theta_e;
         }
 
         rope_yarn(
@@ -5601,7 +5593,6 @@ static void ggml_compute_forward_rope_f32(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
-    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5640,7 +5631,7 @@ static void ggml_compute_forward_rope_f32(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -5788,7 +5779,6 @@ static void ggml_compute_forward_rope_f16(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
-    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5827,7 +5817,7 @@ static void ggml_compute_forward_rope_f16(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -7533,8 +7523,8 @@ static void ggml_compute_forward_upscale_f32(
         float pixel_offset = 0.5f;
         if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
             pixel_offset = 0.0f;
-            sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
-            sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+            sf0 = (float)(ne0 - 1) / (src0->ne[0] - 1);
+            sf1 = (float)(ne1 - 1) / (src0->ne[1] - 1);
         }
 
         for (int64_t i3 = 0; i3 < ne3; i3++) {
@@ -7923,10 +7913,10 @@ void ggml_compute_forward_argsort(
 
 // ggml_compute_forward_flash_attn_ext
 
-static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
+static void ggml_compute_forward_flash_attn_ext_f16(
         const ggml_compute_params * params,
-        ggml_tensor * dst,
-        int ir0, int ir1) {
+        ggml_tensor * dst) {
+
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -7942,6 +7932,9 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
     GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
+    const int ith = params->ith;
+    const int nth = params->nth;
+
     const int64_t DK = nek0;
     const int64_t DV = nev0;
     const int64_t N  = neq1;
@@ -7975,6 +7968,16 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
     // parallelize by q rows using ggml_vec_dot_f32
 
+    // total rows in q
+    const int nr = neq1*neq2*neq3;
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
     float scale         = 1.0f;
     float max_bias      = 0.0f;
     float logit_softcap = 0.0f;
@@ -8001,8 +8004,6 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
     GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
     GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
-    int ith = params->ith;
-
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -8134,7 +8135,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         }
 
         // V /= S
-        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
+        const float S_inv = 1.0f/S;
         ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
         // dst indices
@@ -8150,91 +8151,6 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
     }
 }
 
-static void ggml_compute_forward_flash_attn_ext_f16(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
-    const ggml_tensor * q     = dst->src[0];
-    const ggml_tensor * k     = dst->src[1];
-    const ggml_tensor * v     = dst->src[2];
-
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
-
-    const int64_t DK = nek0;
-    const int64_t DV = nev0;
-    const int64_t N  = neq1;
-
-    GGML_ASSERT(ne0 == DV);
-    GGML_ASSERT(ne2 == N);
-
-    // input tensor rows must be contiguous
-    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
-    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
-    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
-
-    GGML_ASSERT(neq0 == DK);
-    GGML_ASSERT(nek0 == DK);
-    GGML_ASSERT(nev0 == DV);
-
-    GGML_ASSERT(neq1 == N);
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    // parallelize by q rows using ggml_vec_dot_f32
-
-    // total rows in q
-    const int64_t nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    // disable for NUMA
-    const bool disable_chunking = ggml_is_numa();
-
-    // 4x chunks per thread
-    int nth_scaled = nth * 4;
-    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
-    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
-
-    if (nth == 1 || nchunk < nth || disable_chunking) {
-        nchunk = nth;
-    }
-
-    if (ith == 0) {
-        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-        ggml_threadpool_chunk_set(params->threadpool, nth);
-    }
-
-    ggml_barrier(params->threadpool);
-
-    // The number of elements in each chunk
-    const int64_t dr = (nr + nchunk - 1) / nchunk;
-
-    // The first chunk comes from our thread_id, the rest will get auto-assigned.
-    int current_chunk = ith;
-
-    while (current_chunk < nchunk) {
-        const int64_t ir0 = dr * current_chunk;
-        const int64_t ir1 = MIN(ir0 + dr, nr);
-
-        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
-
-        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
-    }
-}
-
 void ggml_compute_forward_flash_attn_ext(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8721,7 +8637,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                 // n_head
                 for (int h = ih0; h < ih1; ++h) {
                     // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
-                    const float dt_soft_plus = ggml_softplus(dt[h]);
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
                     const float dA = expf(dt_soft_plus * A[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
@@ -8730,41 +8646,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                         const int ii = i1 + h*nr;
                         const float x_dt = x[ii] * dt_soft_plus;
                         float sumf = 0.0f;
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-                        const int ggml_f32_epr = svcntw();
-                        const int ggml_f32_step = 1 * ggml_f32_epr;
-
-                        const int np = (nc & ~(ggml_f32_step - 1));
-
-                        GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
-
-                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
-                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
-
-                        for (int i = 0; i < np; i += ggml_f32_step) {
-                            // TODO: maybe unroll more?
-                            for (int j = 0; j < 1; j++) {
-                                GGML_F32_VEC t0 = GGML_F32_VEC_LOAD(s0 + i + j*ggml_f32_epr + ii*nc);
-                                GGML_F32_VEC t1 = GGML_F32_VEC_LOAD(B + i + j*ggml_f32_epr + g*nc);
-                                GGML_F32_VEC t2 = GGML_F32_VEC_LOAD(C + i + j*ggml_f32_epr + g*nc);
-
-                                t0 = GGML_F32_VEC_MUL(t0, adA);
-                                t1 = GGML_F32_VEC_MUL(t1, axdt);
-
-                                t0 = GGML_F32_VEC_ADD(t0, t1);
-
-                                sum = GGML_F32_VEC_FMA(sum, t0, t2);
-
-                                GGML_F32_VEC_STORE(s + i + j*ggml_f32_epr + ii*nc, t0);
-                            }
-                        }
-
-                        sumf = GGML_F32xt_REDUCE_ONE(sum);
-    #elif defined(__riscv_v_intrinsic)
-                        // todo: RVV implementation
-                        const int np = 0;
-    #else
+#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
                         const int np = (nc & ~(GGML_F32_STEP - 1));
 
                         GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
@@ -8795,7 +8677,6 @@ static void ggml_compute_forward_ssm_scan_f32(
 
                         // reduce sum0..sum3 to sum0
                         GGML_F32_VEC_REDUCE(sumf, sum);
-    #endif
 #else
                         const int np = 0;
 #endif
@@ -8818,37 +8699,13 @@ static void ggml_compute_forward_ssm_scan_f32(
                 // n_head
                 for (int h = ih0; h < ih1; ++h) {
                     // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
-                    const float dt_soft_plus = ggml_softplus(dt[h]);
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
                     const int g = h / (nh / ng); // repeat_interleave
 
                     // dim
                     for (int i1 = 0; i1 < nr; ++i1) {
                         const int ii = i1 + h*nr;
                         const float x_dt = x[ii] * dt_soft_plus;
-#if defined(__ARM_FEATURE_SVE)
-                        svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
-                        svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
-                        svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
-
-                        // d_state
-                        // TODO: what happens when (d_state % svcntw()) != 0?
-                        for (int64_t k = 0; k < nc; k += svcntw()) {
-                            svfloat32_t vA = GGML_F32_VEC_LOAD(&A[h*nc + k]);
-                            svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k + g*nc]);
-                            svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k + g*nc]);
-                            svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[ii*nc + k]);
-
-                            svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
-                            t1 = exp_ps_sve(svptrue_b32(), t1);
-                            svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
-
-                            vs0 = GGML_F32_VEC_FMA(t2, vs0, t1);
-                            r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
-
-                            GGML_F32_VEC_STORE(&s[ii*nc + k], vs0);
-                        }
-                        y[ii] = GGML_F32xt_REDUCE_ONE(r1_vector);
-#else
                         float sumf = 0.0f;
                         // NOTE: can't really use GGML_SIMD here because d_state is usually 16
                         //       and also because expf is used within the loop.
@@ -8863,7 +8720,6 @@ static void ggml_compute_forward_ssm_scan_f32(
                             s[i] = state;
                         }
                         y[ii] = sumf;
-#endif
                     }
                 }
             }
@@ -9081,26 +8937,6 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_exp(params, dst);
             } break;
-        case GGML_UNARY_OP_FLOOR:
-            {
-                ggml_compute_forward_floor(params, dst);
-            } break;
-        case GGML_UNARY_OP_CEIL:
-            {
-                ggml_compute_forward_ceil(params, dst);
-            } break;
-        case GGML_UNARY_OP_ROUND:
-            {
-                ggml_compute_forward_round(params, dst);
-            } break;
-        case GGML_UNARY_OP_TRUNC:
-            {
-                ggml_compute_forward_trunc(params, dst);
-            } break;
-        case GGML_UNARY_OP_XIELU:
-            {
-                ggml_compute_forward_xielu(params, dst);
-            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -9335,14 +9171,6 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
         #define GGML_F32X_MUL GGML_F32x16_MUL
         #define GGML_F32X_FMA GGML_F32x16_FMA
         #define WKV_VECTOR_SIZE 16
-    #elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
-        #define GGML_F32X GGML_F32xt
-        #define GGML_F32X_SET1 GGML_F32xt_SET1
-        #define GGML_F32X_LOAD GGML_F32xt_LOAD
-        #define GGML_F32X_STORE GGML_F32xt_STORE
-        #define GGML_F32X_MUL GGML_F32xt_MUL
-        #define GGML_F32X_FMA GGML_F32xt_FMA
-        #define WKV_VECTOR_SIZE 8
     #elif defined(__ARM_NEON) && defined(__aarch64__)
         #define GGML_F32X GGML_F32x4
         #define GGML_F32X_SET1 GGML_F32x4_SET1
@@ -9355,11 +9183,7 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
 
     #ifdef WKV_VECTOR_SIZE
         int wkv_vector_size;
-        #if defined(__ARM_FEATURE_SVE)
-            wkv_vector_size = svcntw();
-        #else
-            wkv_vector_size = WKV_VECTOR_SIZE;
-        #endif
+        wkv_vector_size = WKV_VECTOR_SIZE;
         const int64_t vec_count = head_size / wkv_vector_size;
 
         for (int64_t t = 0; t < T; t++) {
@@ -9551,14 +9375,6 @@ static void ggml_compute_forward_gla_f32(
         #define GGML_F32X_MUL GGML_F32x16_MUL
         #define GGML_F32X_FMA GGML_F32x16_FMA
         #define GLA_VECTOR_SIZE 16
-    #elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
-        #define GGML_F32X GGML_F32xt
-        #define GGML_F32X_SET1 GGML_F32xt_SET1
-        #define GGML_F32X_LOAD GGML_F32xt_LOAD
-        #define GGML_F32X_STORE GGML_F32xt_STORE
-        #define GGML_F32X_MUL GGML_F32xt_MUL
-        #define GGML_F32X_FMA GGML_F32xt_FMA
-        #define GLA_VECTOR_SIZE 8
     #elif defined(__ARM_NEON) && defined(__aarch64__)
         #define GGML_F32X GGML_F32x4
         #define GGML_F32X_SET1 GGML_F32x4_SET1
@@ -9571,11 +9387,7 @@ static void ggml_compute_forward_gla_f32(
 
     #ifdef GLA_VECTOR_SIZE
         int gla_vector_size;
-        #if defined(__ARM_FEATURE_SVE)
-            gla_vector_size = svcntw();
-        #else
-            gla_vector_size = GLA_VECTOR_SIZE;
-        #endif
+        gla_vector_size = GLA_VECTOR_SIZE;
         const int64_t vec_count = head_size / gla_vector_size;
 
         for (int64_t t = 0; t < T; t++) {
@@ -9735,127 +9547,84 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
     GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS
     int64_t h_stride_2d = head_size * head_size;
 
-    #if defined(GGML_SIMD)
-        #if defined(__ARM_FEATURE_SVE) || defined(__riscv_v_intrinsic)
-            // scalar Route to scalar implementation       //TODO: Write SVE code and RVV code
-            for (int64_t t = 0; t < T; t++) {
-                int64_t t_offset = t * t_stride;
-                int64_t state_offset = head_size * C * (t / (T / n_seqs));
-                float * state_cur = state + state_offset;
-                float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+    #if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
+        for (int64_t t = 0; t < T; t++) {
+            int64_t t_offset = t * t_stride;
+            int64_t state_offset = head_size * C * (t / (T / n_seqs));
+            float * state_cur = state + state_offset;
+            float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
 
-                for (int64_t h = h_start; h < h_end; h++) {
-                    int64_t h_offset = h * h_stride;
-                    int64_t t_h_offset = t_offset + h_offset;
-                    int64_t h_2d_offset = h * h_stride_2d;
+            for (int64_t h = h_start; h < h_end; h++) {
+                int64_t h_offset = h * h_stride;
+                int64_t t_h_offset = t_offset + h_offset;
+                int64_t h_2d_offset = h * h_stride_2d;
 
-                    for (int64_t i = 0; i < head_size; i++) {
-                        int64_t t_h_i_offset = t_h_offset + i;
-                        int64_t h_2d_i_offset = h_2d_offset + i * h_stride;
+                for (int64_t ii = 0; ii < head_size; ii++) {
+                    int64_t t_h_i_offset = t_h_offset + ii;
+                    int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
 
-                        float v_val = v[t_h_i_offset];
+                    GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
 
-                        float sa = 0, result = 0;
-                        for (int64_t j = 0; j < head_size; j++) {
-                            sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j];
-                        }
-
-                        for (int64_t j = 0; j < head_size; j++) {
-                            int64_t t_h_j_offset = t_h_offset + j;
-                            int64_t h_2d_i_j_offset = h_2d_i_offset + j;
-
-                            float r_val = r[t_h_j_offset];
-                            float w_val = w[t_h_j_offset];
-                            float k_val = k[t_h_j_offset];
-                            float b_val = b[t_h_j_offset];
-                            float kv_val = v_val * k_val;
-                            float prev_state_val = state_prev[h_2d_i_j_offset];
-                            state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
-                            result += state_cur[h_2d_i_j_offset] * r_val;
-                        }
-                        dst_data[t_h_i_offset] = result;
-                    }
-                }
-            }
-        #else
-            for (int64_t t = 0; t < T; t++) {
-                int64_t t_offset = t * t_stride;
-                int64_t state_offset = head_size * C * (t / (T / n_seqs));
-                float * state_cur = state + state_offset;
-                float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
-
-                for (int64_t h = h_start; h < h_end; h++) {
-                    int64_t h_offset = h * h_stride;
-                    int64_t t_h_offset = t_offset + h_offset;
-                    int64_t h_2d_offset = h * h_stride_2d;
-
-                    for (int64_t ii = 0; ii < head_size; ii++) {
-                        int64_t t_h_i_offset = t_h_offset + ii;
-                        int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
-
-                        GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
-
-                        float sa = 0;
-                        {
-                            GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
-                            GGML_F32_VEC ax[GGML_F32_ARR];
-                            GGML_F32_VEC ay[GGML_F32_ARR];
-                            for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
-                                for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
-                                    ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
-                                    ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
-                                    sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
-                                }
-                            }
-                            GGML_F32_VEC_REDUCE(sa, sum);
-                        }
-
-                        GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
-
-                        int64_t j = 0;
-                        GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
-                        for (; j < head_size; j += GGML_F32_STEP) {
+                    float sa = 0;
+                    {
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
                             for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
-                                int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
-                                int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
-
-                                GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
-                                GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
-                                GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
-                                GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
-
-                                k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
-
-                                GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
-                                // kv + s * decay + sa * b
-                                state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
-                                state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
-                                GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
-
-                                result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
+                                ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
+                                ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
+                                sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
                             }
                         }
-                        GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
+                        GGML_F32_VEC_REDUCE(sa, sum);
+                    }
 
-                        // There shouldn't be left-overs though.
-                        for (; j < head_size; j++) {
-                            int64_t t_h_j_offset = t_h_offset + j;
-                            int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+                    GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
 
-                            float r_val = r[t_h_j_offset];
-                            float w_val = w[t_h_j_offset];
-                            float k_val = k[t_h_j_offset];
-                            float b_val = b[t_h_j_offset];
-                            float kv_val = v[t_h_i_offset] * k_val;
+                    int64_t j = 0;
+                    GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                    for (; j < head_size; j += GGML_F32_STEP) {
+                        for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                            int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
+                            int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
 
-                            float prev_state_val = state_prev[h_2d_i_j_offset];
-                            state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
-                            dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
+                            GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
+                            GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
+                            GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
+                            GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
+
+                            k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
+
+                            GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
+                            // kv + s * decay + sa * b
+                            state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
+                            state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
+                            GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
+
+                            result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
                         }
                     }
+                    GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
+
+                    // There shouldn't be left-overs though.
+                    for (; j < head_size; j++) {
+                        int64_t t_h_j_offset = t_h_offset + j;
+                        int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                        float r_val = r[t_h_j_offset];
+                        float w_val = w[t_h_j_offset];
+                        float k_val = k[t_h_j_offset];
+                        float b_val = b[t_h_j_offset];
+                        float kv_val = v[t_h_i_offset] * k_val;
+
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                        dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
+                    }
                 }
             }
-        #endif
+        }
     #else
         for (int64_t t = 0; t < T; t++) {
             int64_t t_offset = t * t_stride;

ggml/src/ggml-cpu/repack.cpp

@@ -1600,32 +1600,6 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         return false;
     }
 
-    void forward_mul_mat_one_chunk(ggml_compute_params * params, ggml_tensor * op, int64_t src0_start, int64_t src0_end) {
-        const ggml_tensor * src0 = op->src[0];
-        const ggml_tensor * src1 = op->src[1];
-        ggml_tensor *       dst  = op;
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        const void * src1_wdata      = params->wdata;
-        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
-
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
-            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                    src0_end - src0_start);
-        }
-    }
-
     void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
         const ggml_tensor * src0 = op->src[0];
         const ggml_tensor * src1 = op->src[1];
@@ -1669,41 +1643,31 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
         }
 
-        // disable for NUMA
-        const bool disable_chunking = ggml_is_numa();
-
-        // 4x chunks per thread
-        int64_t nr = ggml_nrows(op->src[0]);
-        int nth_scaled = nth * 4;
-        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
-        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
-
-        if (nth == 1 || nchunk < nth || disable_chunking) {
-            nchunk = nth;
-        }
-
-        if (ith == 0) {
-            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            ggml_threadpool_chunk_set(params->threadpool, nth);
-        }
-
         ggml_barrier(params->threadpool);
 
-        // The first chunk comes from our thread_id, the rest will get auto-assigned.
-        int current_chunk = ith;
+        const void * src1_wdata      = params->wdata;
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+        int64_t      src0_start      = (ith * ne01) / nth;
+        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
+        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
+        if (src0_start >= src0_end) {
+            return;
+        }
 
-        while (current_chunk < nchunk) {
-            int64_t src0_start = (current_chunk * ne01) / nchunk;
-            int64_t src0_end   = ((current_chunk + 1) * ne01) / nchunk;
-            src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
-            src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
-            if (src0_start >= src0_end) {
-                break;
-            }
-
-            forward_mul_mat_one_chunk(params, dst, src0_start, src0_end);
-
-            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                    src0_end - src0_start);
         }
     }
 

ggml/src/ggml-cpu/simd-mappings.h

@@ -2,10 +2,6 @@
 
 #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:
 //
@@ -149,164 +145,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 //   number of elements to fit in a single register
 //
 
-#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_FMA)
-
-#define GGML_SIMD
-
-// F32 SVE
-#define GGML_F32_EPR 8
-#define DEFAULT_PG svptrue_b32()
-
-#define GGML_F32xt                        svfloat32_t
-#define GGML_F32xt_ZERO                   svdup_n_f32(0.0f)
-#define GGML_F32xt_SET1(x)                svdup_n_f32(x)
-#define GGML_F32xt_LOAD_IMPL(pg, a, ...)  svld1_f32(pg, a)
-#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, 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__)
-#define GGML_F32xt_MUL_IMPL(pg, a, b)     svmul_f32_m(pg, a, b)
-#define GGML_F32xt_MUL(...)               GGML_F32xt_MUL_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
-#define GGML_F32xt_REDUCE_ONE(...)        GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
-{                                                      \
-    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2);        \
-    sum3 = svadd_f32_m(DEFAULT_PG, sum3, sum4);        \
-    sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum6);        \
-    sum7 = svadd_f32_m(DEFAULT_PG, sum7, sum8);        \
-    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum3);        \
-    sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum7);        \
-    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5);        \
-    (res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1);  \
-}
-#define GGML_F32xt_REDUCE(...) GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, __VA_ARGS__)
-
-#define GGML_F32_VEC        GGML_F32xt
-#define GGML_F32_VEC_ZERO   GGML_F32xt_ZERO
-#define GGML_F32_VEC_SET1   GGML_F32xt_SET1
-#define GGML_F32_VEC_LOAD   GGML_F32xt_LOAD
-#define GGML_F32_VEC_STORE  GGML_F32xt_STORE
-#define GGML_F32_VEC_FMA    GGML_F32xt_FMA
-#define GGML_F32_VEC_ADD    GGML_F32xt_ADD
-#define GGML_F32_VEC_MUL    GGML_F32xt_MUL
-#define GGML_F32_VEC_REDUCE GGML_F32xt_REDUCE
-
-// F16 SVE
-#define DEFAULT_PG32    svptrue_b32()
-#define DEFAULT_PG16    svptrue_b16()
-
-#define GGML_F32Cxt                         svfloat16_t
-#define GGML_F32Cxt_ZERO                    svdup_n_f16(0.0f)
-#define GGML_F32Cxt_SET1(x)                 svdup_n_f16(x)
-#define GGML_F32Cxt_LOAD(p)                 svld1_f16(DEFAULT_PG16, (const __fp16 *)(p))
-#define GGML_F32Cxt_STORE(dst_ptr, src_vec) svst1_f16(DEFAULT_PG16, (__fp16 *)(dst_ptr), (src_vec))
-
-#define GGML_F32Cxt_FMA_IMPL(pg, a, b, c)   svmad_f16_x(pg, b, c, a)
-#define GGML_F32Cxt_FMA(...)                GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, __VA_ARGS__)
-#define GGML_F32Cxt_ADD_IMPL(pg, a, b)      svadd_f16_x(pg, a, b)
-#define GGML_F32Cxt_ADD(...)                GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, __VA_ARGS__)
-#define GGML_F32Cxt_MUL_IMPL(pg, a, b)      svmul_f16_x(pg, a, b)
-#define GGML_F32Cxt_MUL(...)                GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, __VA_ARGS__)
-#define GGML_F32Cxt_REDUCE                  GGML_F16xt_REDUCE_MIXED
-
-#define GGML_F16x_VEC                GGML_F32Cxt
-#define GGML_F16x_VEC_ZERO           GGML_F32Cxt_ZERO
-#define GGML_F16x_VEC_SET1           GGML_F32Cxt_SET1
-#define GGML_F16x_VEC_LOAD(p, i)     GGML_F32Cxt_LOAD(p)
-#define GGML_F16x_VEC_STORE(p, r, i) GGML_F32Cxt_STORE((__fp16 *)(p), r)
-#define GGML_F16x_VEC_FMA            GGML_F32Cxt_FMA
-#define GGML_F16x_VEC_ADD            GGML_F32Cxt_ADD
-#define GGML_F16x_VEC_MUL            GGML_F32Cxt_MUL
-#define GGML_F16x_VEC_REDUCE         GGML_F32Cxt_REDUCE
-
-#define GGML_F16xt_REDUCE_ONE_IMPL(pg, a) svaddv_f16(pg, a)
-#define GGML_F16xt_REDUCE_ONE(...)        GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, __VA_ARGS__)
-
-#define GGML_F16xt_REDUCE_MIXED_IMPL(pg16, res, sum1, sum2, sum3, sum4)  \
-{                                                      \
-    sum1 = svadd_f16_x(pg16, sum1, sum2);              \
-    sum3 = svadd_f16_x(pg16, sum3, sum4);              \
-    sum1 = svadd_f16_x(pg16, sum1, sum3);              \
-    __fp16 sum_f16 = svaddv_f16(pg16, sum1);           \
-    (res) = (ggml_float) sum_f16;                      \
-}
-#define GGML_F16xt_REDUCE_MIXED(...) GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, __VA_ARGS__)
-
-// F16 NEON
-
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-    #define GGML_F16_STEP 32
-    #define GGML_F16_EPR  8
-
-    #define GGML_F16x8              float16x8_t
-    #define GGML_F16x8_ZERO         vdupq_n_f16(0.0f)
-    #define GGML_F16x8_SET1(x)      vdupq_n_f16(x)
-    #define GGML_F16x8_LOAD(x)      vld1q_f16((const __fp16 *)(x))
-    #define GGML_F16x8_STORE        vst1q_f16
-    #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
-    #define GGML_F16x8_ADD          vaddq_f16
-    #define GGML_F16x8_MUL          vmulq_f16
-    #define GGML_F16x8_REDUCE(res, x)                               \
-    do {                                                            \
-        int offset = GGML_F16_ARR >> 1;                             \
-        for (int i = 0; i < offset; ++i) {                          \
-            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
-        }                                                           \
-        offset >>= 1;                                               \
-        for (int i = 0; i < offset; ++i) {                          \
-            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
-        }                                                           \
-        offset >>= 1;                                               \
-        for (int i = 0; i < offset; ++i) {                          \
-            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
-        }                                                           \
-        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
-        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
-        (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
-    } while (0)
-
-    #define GGML_F16_VEC                GGML_F16x8
-    #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
-    #define GGML_F16_VEC_SET1           GGML_F16x8_SET1
-    #define GGML_F16_VEC_LOAD(p, i)     GGML_F16x8_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
-    #define GGML_F16_VEC_FMA            GGML_F16x8_FMA
-    #define GGML_F16_VEC_ADD            GGML_F16x8_ADD
-    #define GGML_F16_VEC_MUL            GGML_F16x8_MUL
-    #define GGML_F16_VEC_REDUCE         GGML_F16x8_REDUCE
-#else
-    // if FP16 vector arithmetic is not supported, we use FP32 instead
-    // and take advantage of the vcvt_ functions to convert to/from FP16
-
-    #define GGML_F16_STEP 16
-    #define GGML_F16_EPR  4
-
-    #define GGML_F32Cx4              float32x4_t
-    #define GGML_F32Cx4_ZERO         vdupq_n_f32(0.0f)
-    #define GGML_F32Cx4_SET1(x)      vdupq_n_f32(x)
-    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
-    #define GGML_F32Cx4_STORE(x, y)  vst1_f16(x, vcvt_f16_f32(y))
-    #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
-    #define GGML_F32Cx4_ADD          vaddq_f32
-    #define GGML_F32Cx4_MUL          vmulq_f32
-    #define GGML_F32Cx4_REDUCE       GGML_F32x4_REDUCE
-
-    #define GGML_F16_VEC                GGML_F32Cx4
-    #define GGML_F16_VEC_ZERO           GGML_F32Cx4_ZERO
-    #define GGML_F16_VEC_SET1           GGML_F32Cx4_SET1
-    #define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx4_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
-    #define GGML_F16_VEC_FMA            GGML_F32Cx4_FMA
-    #define GGML_F16_VEC_ADD            GGML_F32Cx4_ADD
-    #define GGML_F16_VEC_MUL            GGML_F32Cx4_MUL
-    #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
-#endif
-
-#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
 
 #define GGML_SIMD
 

ggml/src/ggml-cpu/spacemit/ime.h

@@ -1,13 +0,0 @@
-#pragma once
-
-#include "ggml-alloc.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-ggml_backend_buffer_type_t ggml_backend_cpu_riscv64_spacemit_buffer_type(void);
-
-#ifdef __cplusplus
-}
-#endif

ggml/src/ggml-cpu/spacemit/ime_kernels.h

@@ -1,26 +0,0 @@
-#pragma once
-
-#include <cstddef>
-
-namespace sqnbitgemm_spacemit_ime {
-namespace ime1 {
-size_t gemm_kernel_i8i4(size_t            blk_len,
-                        const std::byte * quant_a_ptr,
-                        const std::byte * quant_b_data,
-                        const float *     quant_b_scale,
-                        const std::byte * quant_b_zp,
-                        float *           c_ptr,
-                        size_t            count_m,
-                        size_t            count_n,
-                        size_t            count_k,
-                        size_t            block_count_k,
-                        size_t            ldc,
-                        const float *     bias,
-                        const size_t      scale_stride);
-
-void quantize_a_row_i8(size_t blk_len, const float * a_ptr, size_t count_k, std::byte * quant_a_ptr);
-
-void quantize_a_4row_i8(size_t blk_len, const float * a_ptr, size_t count_k, std::byte * quant_a_ptr);
-
-}  // namespace ime1
-}  // namespace sqnbitgemm_spacemit_ime

ggml/src/ggml-cpu/unary-ops.cpp

@@ -52,15 +52,6 @@ static inline float op_sqrt(float x) {
     return sqrtf(x);
 }
 
-static inline float op_xielu(float x, float alpha_n, float alpha_p, float beta, float eps) {
-    if (x > 0.0f) {
-        return alpha_p * x * x + beta * x;
-    } else {
-        const float min_x_eps = fminf(x, eps);
-        return (expm1f(min_x_eps) - x) * alpha_n + beta * x;
-    }
-}
-
 static inline float op_sin(float x) {
     return sinf(x);
 }
@@ -73,22 +64,6 @@ static inline float op_log(float x) {
     return logf(x);
 }
 
-static inline float op_floor(float x) {
-    return floorf(x);
-}
-
-static inline float op_ceil(float x) {
-    return ceilf(x);
-}
-
-static inline float op_round(float x) {
-    return roundf(x);
-}
-
-static inline float op_trunc(float x) {
-    return truncf(x);
-}
-
 template <float (*op)(float), typename src0_t, typename dst_t>
 static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
     constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
@@ -146,86 +121,6 @@ static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
     }
 }
 
-template <float (*op)(float, ggml_tensor *)>
-static void unary_op_params(const ggml_compute_params * params, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-
-    /*  */ if (src0->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) { // all f32
-        apply_unary_op<op, float, float>(params, dst);
-    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F16) { // all f16
-        apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst);
-    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
-        apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst);
-    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
-        apply_unary_op<op, ggml_bf16_t, float>(params, dst);
-    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F32) {
-        apply_unary_op<op, ggml_fp16_t, float>(params, dst);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type));
-        GGML_ABORT("fatal error");
-    }
-}
-
-// Extend vec_unary_op to support functors
-template <typename Op, typename src0_t, typename dst_t>
-static inline void vec_unary_op_functor(int64_t n, dst_t * y, const src0_t * x, Op op) {
-    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
-    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
-
-    for (int i = 0; i < n; i++) {
-        y[i] = f32_to_dst(op(src0_to_f32(x[i])));
-    }
-}
-
-// Extend apply_unary_op to support functors
-template <typename Op, typename src0_t, typename dst_t>
-static void apply_unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) {
-    const ggml_tensor * src0 = dst->src[0];
-
-    GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    GGML_ASSERT( nb0 == sizeof(dst_t));
-    GGML_ASSERT(nb00 == sizeof(src0_t));
-
-    const auto [ir0, ir1] = get_thread_range(params, src0);
-
-    for (int64_t ir = ir0; ir < ir1; ++ir) {
-        const int64_t i03 = ir/(ne02*ne01);
-        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
-        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
-
-        dst_t        * dst_ptr  = (dst_t  *)       ((char *)       dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
-        const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
-
-        vec_unary_op_functor(ne0, dst_ptr, src0_ptr, op);
-    }
-}
-
-// Generic dispatcher for functors
-template <typename Op>
-static void unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) {
-    const ggml_tensor * src0 = dst->src[0];
-
-    /*  */ if (src0->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) { // all f32
-        apply_unary_op_functor<Op, float, float>(params, dst, op);
-    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F16) { // all f16
-        apply_unary_op_functor<Op, ggml_fp16_t, ggml_fp16_t>(params, dst, op);
-    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
-        apply_unary_op_functor<Op, ggml_bf16_t, ggml_bf16_t>(params, dst, op);
-    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
-        apply_unary_op_functor<Op, ggml_bf16_t, float>(params, dst, op);
-    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F32) {
-        apply_unary_op_functor<Op, ggml_fp16_t, float>(params, dst, op);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type));
-        GGML_ABORT("fatal error");
-    }
-}
-
 void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) {
     unary_op<op_abs>(params, dst);
 }
@@ -289,33 +184,3 @@ void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor *
 void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) {
     unary_op<op_log>(params, dst);
 }
-
-void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
-    unary_op<op_floor>(params, dst);
-}
-
-void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
-    unary_op<op_ceil>(params, dst);
-}
-
-void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
-    unary_op<op_round>(params, dst);
-}
-
-void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
-    unary_op<op_trunc>(params, dst);
-}
-
-void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
-    const float alpha_n = ggml_get_op_params_f32(dst, 1);
-    const float alpha_p = ggml_get_op_params_f32(dst, 2);
-    const float beta = ggml_get_op_params_f32(dst, 3);
-    const float eps = ggml_get_op_params_f32(dst, 4);
-
-    const auto xielu_op_params = [alpha_n, alpha_p, beta, eps](float f) {
-        return op_xielu(f, alpha_n, alpha_p, beta, eps);
-    };
-
-    unary_op_functor(params, dst, xielu_op_params);
-}
-

ggml/src/ggml-cpu/unary-ops.h

@@ -22,11 +22,6 @@ void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-cpu/vec.cpp

@@ -18,73 +18,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
 #if defined(GGML_SIMD)
     float sumf = 0.0f;
 
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
-        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
-        const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
-
-        const int np = (n & ~(ggml_f32_step - 1));
-        svfloat32_t sum1 = svdup_n_f32(0.0f);
-        svfloat32_t sum2 = svdup_n_f32(0.0f);
-        svfloat32_t sum3 = svdup_n_f32(0.0f);
-        svfloat32_t sum4 = svdup_n_f32(0.0f);
-        svfloat32_t sum5 = svdup_n_f32(0.0f);
-        svfloat32_t sum6 = svdup_n_f32(0.0f);
-        svfloat32_t sum7 = svdup_n_f32(0.0f);
-        svfloat32_t sum8 = svdup_n_f32(0.0f);
-        svfloat32_t ax1,ax2,ax3,ax4,ax5,ax6,ax7,ax8;
-        svfloat32_t ay1,ay2,ay3,ay4,ay5,ay6,ay7,ay8;
-        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(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(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(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(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(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(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(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(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
-        const int np2 = (n & ~(ggml_f32_epr - 1));
-        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(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) {
-            svbool_t pg = svwhilelt_b32(np2, n);
-            ax1 = svld1_f32(pg, x + np2);
-            ay1 = svld1_f32(pg, y + np2);
-            sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
-        }
-        // reduce sum1,sum2 to sum1
-        GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
-    #elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_v_intrinsic)
         int vl = __riscv_vsetvlmax_e32m8();
         vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
         vfloat32m8_t vsum;
@@ -215,69 +149,7 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
 
 #if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8; //get vector length
-        const int ggml_f16_epr = sve_register_length / 16; // running when 16
-        const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
-
-        const int np= (n & ~(ggml_f16_step - 1));
-        svfloat16_t sum1 = svdup_n_f16(0.0f);
-        svfloat16_t sum2 = svdup_n_f16(0.0f);
-        svfloat16_t sum3 = svdup_n_f16(0.0f);
-        svfloat16_t sum4 = svdup_n_f16(0.0f);
-
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
-            sum1 = GGML_F16x_VEC_FMA(sum1, ax1, ay1);
-
-            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
-            sum2 = GGML_F16x_VEC_FMA(sum2, ax2, ay2);
-
-            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-            sum3 = GGML_F16x_VEC_FMA(sum3, ax3, ay3);
-
-            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-            sum4 = GGML_F16x_VEC_FMA(sum4, ax4, ay4);
-
-            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-            sum1 = GGML_F16x_VEC_FMA(sum1, ax5, ay5);
-
-            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-            sum2 = GGML_F16x_VEC_FMA(sum2, ax6, ay6);
-
-            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-            sum3 = GGML_F16x_VEC_FMA(sum3, ax7, ay7);
-
-            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-            sum4 = GGML_F16x_VEC_FMA(sum4, ax8, ay8);
-        }
-
-        const int np2 = (n & ~(ggml_f16_epr - 1)); // round down to multiple of 8
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-            sum1 = GGML_F16x_VEC_FMA(sum1, rx, ry);
-        }
-
-        if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
-
-            sum1 = svmad_f16_x(pg, hx, hy, sum1);
-        }
-        GGML_F16x_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4);
-    #elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_v_intrinsic)
         #if defined(__riscv_zvfh)
             int vl = __riscv_vsetvlmax_e32m2();
             vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
@@ -404,72 +276,6 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
     }
 }
 
-ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean) {
-    int i = 0;
-    ggml_float sum = 0;
-// TODO: optimize to process the remaining elements in groups using the smaller vector sizes from AVX2 and SSE
-// ref: https://github.com/ggml-org/llama.cpp/pull/15953#pullrequestreview-3310928344
-#if defined(__AVX512F__) && defined(__AVX512DQ__)
-    for (; i + 15 < n; i += 16) {
-        __m512 val = _mm512_sub_ps(_mm512_loadu_ps(x + i),
-                                   _mm512_set1_ps(mean));
-        _mm512_storeu_ps(y + i, val);
-        sum += (ggml_float)_mm512_reduce_add_ps(_mm512_mul_ps(val, val));
-    }
-#elif defined(__AVX2__) && defined(__FMA__)
-    for (; i + 7 < n; i += 8) {
-        __m256 val = _mm256_sub_ps(_mm256_loadu_ps(x + i),
-                                   _mm256_set1_ps(mean));
-        _mm256_storeu_ps(y + i, val);
-        val = _mm256_mul_ps(val,val);
-        __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
-                                 _mm256_castps256_ps128(val));
-        val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
-        val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
-        sum += (ggml_float)_mm_cvtss_f32(val2);
-    }
-#elif defined(__SSE2__)
-    for (; i + 3 < n; i += 4) {
-        __m128 val = _mm_sub_ps(_mm_loadu_ps(x + i),
-                                _mm_set1_ps(mean));
-        _mm_storeu_ps(y + i, val);
-        val = _mm_mul_ps(val, val);
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-        val = _mm_add_ps(val, _mm_movehl_ps(val, val));
-        val = _mm_add_ss(val, _mm_movehdup_ps(val));
-#else
-        __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
-        val = _mm_add_ps(val, tmp);
-        tmp = _mm_movehl_ps(tmp, val);
-        val = _mm_add_ss(val, tmp);
-#endif  // __AVX__ || __AVX2__ || __AVX512F__
-        sum += (ggml_float)_mm_cvtss_f32(val);
-    }
-#elif defined(__ARM_NEON) && defined(__aarch64__)
-    for (; i + 3 < n; i += 4) {
-        float32x4_t val = vsubq_f32(vld1q_f32(x + i),
-                                    vdupq_n_f32(mean));
-        vst1q_f32(y + i, val);
-        val = vmulq_f32(val, val);
-        sum += (ggml_float)vaddvq_f32(val);
-    }
-#elif defined(__VXE__) || defined(__VXE2__)
-    for (; i + 3 < n; i += 4) {
-        float32x4_t val = vec_sub(vec_xl(0, x + i), vec_splats(mean));
-        vec_xst(val, 0, y + i);
-        val = vec_mul(val, val);
-        sum += (ggml_float)vec_hsum_f32x4(val);
-    }
-#endif
-    for (; i < n; ++i) {
-        float val = x[i] - mean;
-        y[i] = val;
-        val *= val;
-        sum += (ggml_float)val;
-    }
-    return sum/n;
-}
-
 ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
     int i = 0;
     ggml_float sum = 0;

ggml/src/ggml-cpu/vec.h

@@ -44,7 +44,6 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
 void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
 
 void ggml_vec_silu_f32(const int n, float * y, const float * x);
-ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean); //it will also center y ( y = y - mean )
 ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max);
 ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max);
 
@@ -119,150 +118,37 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
     }
 
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
+#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
+    const int np = (n & ~(GGML_F16_STEP - 1));
 
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16; // running when 16
-        const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
+    GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
 
-        const int np = (n & ~(ggml_f16_step - 1));
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
 
-        svfloat16_t sum_00 = svdup_n_f16(0.0f);
-        svfloat16_t sum_01 = svdup_n_f16(0.0f);
-        svfloat16_t sum_02 = svdup_n_f16(0.0f);
-        svfloat16_t sum_03 = svdup_n_f16(0.0f);
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
 
-        svfloat16_t sum_10 = svdup_n_f16(0.0f);
-        svfloat16_t sum_11 = svdup_n_f16(0.0f);
-        svfloat16_t sum_12 = svdup_n_f16(0.0f);
-        svfloat16_t sum_13 = svdup_n_f16(0.0f);
+            for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+                ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
 
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0); // 8 elements
-
-            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elements
-            sum_00 = GGML_F16x_VEC_FMA(sum_00, ax1, ay1);     // sum_00 = sum_00+ax1*ay1
-            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 0*ggml_f16_epr, 0); // 8 elements
-            sum_10 = GGML_F16x_VEC_FMA(sum_10, ax1, ay1);
-
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1); // next 8 elements
-
-            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 elements
-            sum_01 = GGML_F16x_VEC_FMA(sum_01, ax2, ay2);
-            ax2 = GGML_F16x_VEC_LOAD(x[1] + i + 1*ggml_f16_epr, 1);
-            sum_11 = GGML_F16x_VEC_FMA(sum_11, ax2, ay2);
-
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-
-            ax3 = GGML_F16x_VEC_LOAD(x[0] + i + 2*ggml_f16_epr, 2);
-            sum_02 = GGML_F16x_VEC_FMA(sum_02, ax3, ay3);
-            ax3 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
-            sum_12 = GGML_F16x_VEC_FMA(sum_12, ax3, ay3);
-
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-
-            ax4 = GGML_F16x_VEC_LOAD(x[0] + i + 3*ggml_f16_epr, 3);
-            sum_03 = GGML_F16x_VEC_FMA(sum_03, ax4, ay4);
-            ax4 = GGML_F16x_VEC_LOAD(x[1] + i + 3*ggml_f16_epr, 3);
-            sum_13 = GGML_F16x_VEC_FMA(sum_13, ax4, ay4);
-
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-
-            ax5 = GGML_F16x_VEC_LOAD(x[0] + i + 4*ggml_f16_epr, 4);
-
-            sum_00 = GGML_F16x_VEC_FMA(sum_00, ax5, ay5);
-            ax5 = GGML_F16x_VEC_LOAD(x[1] + i + 4*ggml_f16_epr, 4);
-            sum_10 = GGML_F16x_VEC_FMA(sum_10, ax5, ay5);
-
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-
-            ax6 = GGML_F16x_VEC_LOAD(x[0] + i + 5*ggml_f16_epr, 5);
-
-            sum_01 = GGML_F16x_VEC_FMA(sum_01, ax6, ay6);
-            ax6 = GGML_F16x_VEC_LOAD(x[1] + i + 5*ggml_f16_epr, 5);
-            sum_11 = GGML_F16x_VEC_FMA(sum_11, ax6, ay6);
-
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-
-            ax7 = GGML_F16x_VEC_LOAD(x[0] + i + 6*ggml_f16_epr, 6);
-
-            sum_02 = GGML_F16x_VEC_FMA(sum_02, ax7, ay7);
-            ax7 = GGML_F16x_VEC_LOAD(x[1] + i + 6*ggml_f16_epr, 6);
-            sum_12 = GGML_F16x_VEC_FMA(sum_12, ax7, ay7);
-
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-
-            ax8 = GGML_F16x_VEC_LOAD(x[0] + i + 7*ggml_f16_epr, 7);
-
-            sum_03 = GGML_F16x_VEC_FMA(sum_03, ax8, ay8);
-            ax8 = GGML_F16x_VEC_LOAD(x[1] + i + 7*ggml_f16_epr, 7);
-            sum_13 = GGML_F16x_VEC_FMA(sum_13, ax8, ay8);
-        }
-
-        const int np2 = (n & ~(ggml_f16_epr - 1));
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x[0] + k, 0);
-            sum_00 = GGML_F16x_VEC_FMA(sum_00, rx, ry);
-            rx = GGML_F16x_VEC_LOAD(x[1] + k, 0);
-            sum_10 = GGML_F16x_VEC_FMA(sum_10, rx, ry);
-        }
-
-        if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx_0 = svld1_f16(pg, (const __fp16 *)(x[0] + np2));
-            svfloat16_t hx_1 = svld1_f16(pg, (const __fp16 *)(x[1] + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
-
-            sum_00 = svmad_f16_x(pg, hx_0, hy, sum_00);
-            sum_10 = svmad_f16_x(pg, hx_1, hy, sum_10);
-        }
-        GGML_F16x_VEC_REDUCE(sumf[0], sum_00, sum_01, sum_02, sum_03);
-        GGML_F16x_VEC_REDUCE(sumf[1], sum_10, sum_11, sum_12, sum_13);
-    #elif defined(__riscv_v_intrinsic)
-      // todo: RVV impl
-      for (int i = 0; i < n; ++i) {
-          for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-              sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
-          }
-      }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
-
-        GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
-
-        GGML_F16_VEC ax[GGML_F16_ARR];
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-
-                for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
-                    ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
-
-                    sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
-                }
+                sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
             }
         }
+    }
 
-        // reduce sum0..sum3 to sum0
-        for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
-            GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
-        }
+    // reduce sum0..sum3 to sum0
+    for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+        GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
+    }
 
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-                sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
-            }
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
-    #endif
+    }
 #else
     for (int i = 0; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
@@ -278,86 +164,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 
 inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-
-        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
-        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
-        const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
-        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
-
-        const int np = (n & ~(ggml_f32_step - 1));
-        svfloat32_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat32_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-        for (int i = 0; i < np; i += ggml_f32_step) {
-
-            ax1 = GGML_F32_VEC_LOAD(x + i);
-            ay1 = GGML_F32_VEC_LOAD(y + i);
-            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(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(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(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(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(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(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(ay8, ax8, vx);
-
-            GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
-        }
-        // leftovers
-        // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
-        const int np2 = (n & ~(ggml_f32_epr - 1));
-        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(ay1, ax1, vx);
-
-            GGML_F32_VEC_STORE(y + i, ay1);
-        }
-        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
-        if (np2 < n) {
-            svbool_t pg =svwhilelt_b32(np2, n);
-            ax1 = svld1_f32(pg, x + np2);
-            ay1 = svld1_f32(pg, y + np2);
-            ay1 = svmad_f32_m(pg, ax1, vx, ay1);
-
-            svst1_f32(pg, y + np2, ay1);
-        }
-    #elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_v_intrinsic)
         for (int i = 0, avl; i < n; i += avl) {
             avl = __riscv_vsetvl_e32m8(n - i);
             vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
@@ -397,113 +204,28 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 }
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16;
-        const int ggml_f16_step = 8 * ggml_f16_epr;
+#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
+    const int np = (n & ~(GGML_F16_STEP - 1));
 
-        GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
 
-        const int np= (n & ~(ggml_f16_step - 1));
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
 
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
 
-            GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
-
-            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
-
-            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
-
-            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
-
-            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
-
-            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
-
-            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
-
-            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
-
-            GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        const int np2 = (n & ~(ggml_f16_epr - 1));
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-            ry = GGML_F16x_VEC_FMA(ry, rx, vx);
+    }
 
-            GGML_F16x_VEC_STORE(y + k, ry, 0);
-        }
-
-        if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
-            hy = svmad_f16_x(pg, hx, vx, hy);
-            svst1_f16(pg, (__fp16 *)(y + np2), hy);
-        }
-
-    #elif defined(__riscv_v_intrinsic)
-        // todo: RVV impl
-        // scalar
-        for (int i = 0; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
-
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ax[GGML_F16_ARR];
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
-
-                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
-            }
-        }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #endif
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        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) {
@@ -524,14 +246,7 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
     }
 
 #if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        // scalar Route to scalar implementation       //TODO: Write SVE code
-        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-            for (int i = 0; i < n; ++i) {
-                y[i] += x[k][i]*v[k][0];
-            }
-        }
-    #elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_v_intrinsic)
         for (int i = 0, avl; i < n; i += avl) {
             avl = __riscv_vsetvl_e32m8(n - i);
             vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
@@ -587,12 +302,7 @@ inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, co
 #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;
-        }
-    #elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_v_intrinsic)
         for (int i = 0, avl; i < n; i += avl) {
             avl = __riscv_vsetvl_e32m8(n - i);
             vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
@@ -611,7 +321,7 @@ inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, co
         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(vb, ay[j], vs);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], vs, vb);
 
                 GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
             }
@@ -635,33 +345,7 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)
     vDSP_vsmul(y, 1, &v, y, 1, n);
 #elif defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
-        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
-        const int ggml_f32_step = 2 * ggml_f32_epr;
-
-        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
-        const int np = (n & ~(ggml_f32_step - 1));
-        svfloat32_t ay1;
-        svfloat32_t ay2;
-        for (int i = 0; i < np; i += ggml_f32_step) {
-            ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_MUL(ay1, vx);
-            GGML_F32_VEC_STORE(y + i, ay1);
-
-            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            ay2 = GGML_F32_VEC_MUL(ay2, vx);
-            GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
-        }
-        // leftovers
-        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
-        for (int i = np; i < n; i += ggml_f32_epr) {
-            svbool_t pg = svwhilelt_b32(i, n);
-            ay1 = svld1_f32(pg, y + i);
-            ay1 = svmul_f32_m(pg, ay1, vx);
-            svst1_f32(pg, y + i, ay1);
-        }
-    #elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_v_intrinsic)
         for (int i = 0, avl; i < n; i += avl) {
             avl = __riscv_vsetvl_e32m8(n - i);
             vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
@@ -698,60 +382,26 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 }
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16;
-        const int ggml_f16_step = 2 * ggml_f16_epr;
+#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
+    const int np = (n & ~(GGML_F16_STEP - 1));
 
-        GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
-        const int np = (n & ~(ggml_f16_step - 1));
-        svfloat16_t ay1, ay2;
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
 
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_MUL(ay1, vx);
-            GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
+    GGML_F16_VEC ay[GGML_F16_ARR];
 
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_MUL(ay2, vx);
-            GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        // leftovers
-        // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
-        if (np < n) {
-            svbool_t pg = svwhilelt_b16(np, n);
-            svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
-            svfloat16_t out = svmul_f16_m(pg, hy, vx);
-            svst1_f16(pg, (__fp16 *)(y + np), out);
-        }
-    #elif defined(__riscv_v_intrinsic)
-        // todo: RVV impl
-        // scalar
-        for (int i = 0; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
+    }
 
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
-
-                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
-            }
-        }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-        }
-    #endif
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+    }
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -820,8 +470,7 @@ inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_f
 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) {
-        const float v = GGML_CPU_FP16_TO_FP32(x[i]);
-        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : expm1f(v));
+        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; }

ggml/src/ggml-cuda/CMakeLists.txt

@@ -44,8 +44,6 @@ if (CUDAToolkit_FOUND)
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
 
     file(GLOB   GGML_SOURCES_CUDA "*.cu")
-    file(GLOB   SRCS "template-instances/fattn-tile*.cu")
-    list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/fattn-mma*.cu")
     list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/mmq*.cu")

ggml/src/ggml-cuda/argsort.cu

@@ -1,81 +1,5 @@
 #include "argsort.cuh"
 
-#ifdef GGML_CUDA_USE_CUB
-#    include <cub/cub.cuh>
-using namespace cub;
-#endif  // GGML_CUDA_USE_CUB
-
-static __global__ void init_indices(int * indices, const int ncols, const int nrows) {
-    const int col = blockIdx.x * blockDim.x + threadIdx.x;
-    const int row = blockIdx.y;
-
-    if (col < ncols && row < nrows) {
-        indices[row * ncols + col] = col;
-    }
-}
-
-static __global__ void init_offsets(int * offsets, const int ncols, const int nrows) {
-    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
-    if (idx <= nrows) {
-        offsets[idx] = idx * ncols;
-    }
-}
-
-#ifdef GGML_CUDA_USE_CUB
-static void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
-                                     const float *    x,
-                                     int *            dst,
-                                     const int        ncols,
-                                     const int        nrows,
-                                     ggml_sort_order  order,
-                                     cudaStream_t     stream) {
-    ggml_cuda_pool_alloc<int>   temp_indices_alloc(pool, ncols * nrows);
-    ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
-    ggml_cuda_pool_alloc<int>   offsets_alloc(pool, nrows + 1);
-
-    int *   temp_indices = temp_indices_alloc.get();
-    float * temp_keys    = temp_keys_alloc.get();
-    int *   d_offsets    = offsets_alloc.get();
-
-    static const int block_size = 256;
-    const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
-    init_indices<<<grid_size, block_size, 0, stream>>>(temp_indices, ncols, nrows);
-
-    const dim3 offset_grid((nrows + block_size - 1) / block_size);
-    init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
-
-    cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream);
-
-    size_t temp_storage_bytes = 0;
-
-    if (order == GGML_SORT_ORDER_ASC) {
-        DeviceSegmentedRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
-                                            temp_indices, dst,                                  // values (indices)
-                                            ncols * nrows, nrows,                            // num items, num segments
-                                            d_offsets, d_offsets + 1, 0, sizeof(float) * 8,  // all bits
-                                            stream);
-    } else {
-        DeviceSegmentedRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
-                                                      dst, ncols * nrows, nrows, d_offsets, d_offsets + 1, 0,
-                                                      sizeof(float) * 8, stream);
-    }
-
-    ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
-    void *                        d_temp_storage = temp_storage_alloc.get();
-
-    if (order == GGML_SORT_ORDER_ASC) {
-        DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
-                                            ncols * nrows, nrows, d_offsets, d_offsets + 1, 0, sizeof(float) * 8,
-                                            stream);
-    } else {
-        DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
-                                                      temp_indices, dst, ncols * nrows, nrows, d_offsets, d_offsets + 1,
-                                                      0, sizeof(float) * 8, stream);
-    }
-}
-#endif  // GGML_CUDA_USE_CUB
-
-// Bitonic sort implementation
 template<typename T>
 static inline __device__ void ggml_cuda_swap(T & a, T & b) {
     T tmp = a;
@@ -87,7 +11,7 @@ template<ggml_sort_order order>
 static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
     // bitonic sort
     int col = threadIdx.x;
-    int row = blockIdx.x;
+    int row = blockIdx.y;
 
     if (col >= ncols_pad) {
         return;
@@ -141,28 +65,21 @@ static int next_power_of_2(int x) {
     return n;
 }
 
-static void argsort_f32_i32_cuda_bitonic(const float *   x,
-                                         int *           dst,
-                                         const int       ncols,
-                                         const int       nrows,
-                                         ggml_sort_order order,
-                                         cudaStream_t    stream) {
+static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
     // bitonic sort requires ncols to be power of 2
     const int ncols_pad = next_power_of_2(ncols);
 
     const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(nrows, 1, 1);
+    const dim3 block_nums(1, nrows, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     // FIXME: this limit could be raised by ~2-4x on Ampere or newer
     GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
 
     if (order == GGML_SORT_ORDER_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_ASC>
-            <<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else if (order == GGML_SORT_ORDER_DESC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_DESC>
-            <<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else {
         GGML_ABORT("fatal error");
     }
@@ -183,18 +100,5 @@ void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
-#ifdef GGML_CUDA_USE_CUB
-    const int    ncols_pad      = next_power_of_2(ncols);
-    const size_t shared_mem     = ncols_pad * sizeof(int);
-    const size_t max_shared_mem = ggml_cuda_info().devices[ggml_cuda_get_device()].smpb;
-
-    if (shared_mem > max_shared_mem || ncols > 1024) {
-        ggml_cuda_pool & pool = ctx.pool();
-        argsort_f32_i32_cuda_cub(pool, src0_d, (int *) dst_d, ncols, nrows, order, stream);
-    } else {
-        argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
-    }
-#else
-    argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
-#endif
+    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
 }

ggml/src/ggml-cuda/binbcast.cu

@@ -272,7 +272,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         const uint3 ne12 = init_fastdiv_values((uint32_t) cne1[2]);
         const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);
 
-        if (block_nums.z > 65535 || block_nums.y > 65535) {
+        if (block_nums.z > 65535) {
             int         block_num  = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
             const uint3 prod_012    = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
             const uint3 prod_01     = init_fastdiv_values((uint32_t) (ne0 * ne1));

ggml/src/ggml-cuda/common.cuh

@@ -220,6 +220,14 @@ static const char * cu_get_error_str(CUresult err) {
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
+#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+#define FP16_MMA_AVAILABLE
+#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(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
+#define FP16_MMA_AVAILABLE
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
+
 #if defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
@@ -245,8 +253,7 @@ static bool fp16_available(const int cc) {
 }
 
 static bool fast_fp16_available(const int cc) {
-    return GGML_CUDA_CC_IS_AMD(cc) ||
-        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -255,6 +262,27 @@ static bool fast_fp16_hardware_available(const int cc) {
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }
 
+// Any FP16 tensor core instructions are available for ggml code.
+static bool fp16_mma_available(const int cc) {
+#if defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
+    return false;
+#else
+    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(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) ||
@@ -572,10 +600,6 @@ static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v,
 }
 
 // Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
-// Important: do not use this function if dst and src both point at registers.
-//     Due to the strict aliasing rule the compiler can do incorrect optimizations if src and dst have different types.
-//     The function is intended for copies between registers and SRAM/VRAM to make the compiler emit the right instructions.
-//     If dst and src point at different address spaces then they are guaranteed to not be aliased.
 template <int nbytes, int alignment = 0>
 static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
     if constexpr (alignment != 0) {
@@ -625,11 +649,8 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 // and a shift:
 //
 // n/d = (mulhi(n, mp) + n) >> L;
-static const uint3 init_fastdiv_values(uint64_t d_64) {
-    GGML_ASSERT(d_64 != 0);
-    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());
-
-    uint32_t d = (uint32_t)d_64;
+static const uint3 init_fastdiv_values(uint32_t d) {
+    GGML_ASSERT(d != 0);
 
     // compute L = ceil(log2(d));
     uint32_t L = 0;
@@ -947,6 +968,13 @@ struct ggml_cuda_graph {
     bool disable_due_to_failed_graph_capture = false;
     int number_consecutive_updates = 0;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
+    bool use_cpy_indirection = false;
+    std::vector<char *> cpy_dest_ptrs;
+    char ** dest_ptrs_d;
+    int dest_ptrs_size = 0;
+    // Index to allow each cpy kernel to be aware of it's position within the graph
+    // relative to other cpy nodes.
+    int graph_cpynode_index = -1;
 #endif
 };
 
@@ -1008,16 +1036,3 @@ struct ggml_backend_cuda_context {
         return pool(device);
     }
 };
-
-struct ggml_cuda_mm_fusion_args_host {
-    const ggml_tensor * x_bias = nullptr;
-    const ggml_tensor * gate = nullptr;
-    const ggml_tensor * gate_bias = nullptr;
-    ggml_glu_op glu_op;
-};
-struct ggml_cuda_mm_fusion_args_device {
-    const void * x_bias = nullptr;
-    const void * gate = nullptr;
-    const void * gate_bias = nullptr;
-    ggml_glu_op glu_op;
-};

ggml/src/ggml-cuda/convert.cuh

@@ -1,4 +1,3 @@
-#pragma once
 #include "common.cuh"
 
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256

ggml/src/ggml-cuda/cpy.cu

@@ -8,16 +8,18 @@
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne,
                                const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                               const int nb12, const int nb13) {
+                               const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
     // then combine those indices with the corresponding byte offsets to get the total offsets
     const int64_t i03 = i/(ne00 * ne01 * ne02);
@@ -61,16 +63,18 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13) {
+                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -87,16 +91,18 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13) {
+                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -112,71 +118,67 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
-template<typename src_t, typename dst_t>
-static __global__ void cpy_flt_contiguous(const char * cx, char * cdst, const int64_t ne) {
-    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
+// Copy destination pointers to GPU to be available when pointer indirection is in use
 
-    if (i >= ne) {
-        return;
+void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) {
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
+    if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers
+        CUDA_CHECK(cudaStreamSynchronize(stream));
+        if (cuda_graph->dest_ptrs_d != nullptr) {
+            CUDA_CHECK(cudaFree(cuda_graph->dest_ptrs_d));
+        }
+        CUDA_CHECK(cudaMalloc(&cuda_graph->dest_ptrs_d, host_dest_ptrs_size*sizeof(char *)));
+        cuda_graph->dest_ptrs_size = host_dest_ptrs_size;
     }
-
-    const src_t * x = (const src_t *) cx;
-    dst_t *     dst = (dst_t *) cdst;
-
-    dst[i] = ggml_cuda_cast<dst_t>(x[i]);
-}
-
-template<typename src_t, typename dst_t>
-static void ggml_cpy_flt_contiguous_cuda(
-    const char * cx, char * cdst, const int64_t ne,
-cudaStream_t stream) {
-
-    const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_flt_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne);
+    // copy destination pointers to GPU
+    CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
+    cuda_graph->graph_cpynode_index = 0; // reset index
+#else
+    GGML_UNUSED_VARS(cuda_graph, host_dest_ptrs, host_dest_ptrs_size, stream);
+#endif
 }
 
 template<typename src_t, typename dst_t>
 static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
     cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q8_0_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q4_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q4_0_f32_cuda(
@@ -185,22 +187,22 @@ static void ggml_cpy_q4_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q4_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q4_1_f32_cuda(
@@ -209,22 +211,22 @@ static void ggml_cpy_q4_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q5_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q5_0_f32_cuda(
@@ -233,22 +235,22 @@ static void ggml_cpy_q5_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q5_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
     cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q5_1_f32_cuda(
@@ -257,25 +259,25 @@ static void ggml_cpy_q5_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_iq4_nl_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
     cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -309,9 +311,17 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
 
-    const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
-
-    if (src0->type == src1->type && contiguous_srcs) {
+    char ** dest_ptrs_d = nullptr;
+    int graph_cpynode_index = -1;
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
+        dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
+        graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
+    }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
+#endif
+    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
 #if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
         if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
@@ -322,91 +332,121 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<float, float>           (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<float, half>        (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<float, half>        (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<half, half>               (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<half, nv_bfloat16>  (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<half, nv_bfloat16>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<half, float>        (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<half, float>          (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, half>  (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<nv_bfloat16, half>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<nv_bfloat16, float>   (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<float, int32_t>     (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<float, int32_t>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<float, int32_t> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<int32_t, float>     (src0_ddc, src1_ddc, ne, main_stream);
-        } else {
-            ggml_cpy_flt_cuda<int32_t, float>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-        }
+        ggml_cpy_flt_cuda<int32_t, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else {
+        GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+    }
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
+        ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
+    }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
+#endif
+
+}
+
+void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    bool disable_indirection = true;
+    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
+}
+
+void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
+    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        return nullptr;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<float, float>>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<float, nv_bfloat16>>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_flt<cpy_1_flt<float, half>>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
+    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>;
+    } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>;
+    } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>;
+    } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
+        return (void*) cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>;
+    } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_flt<cpy_1_flt<half, half>>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<half, nv_bfloat16>>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<half, float>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, half>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
+        return (void*) cpy_flt<cpy_1_flt<float, int32_t>>;
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<int32_t, float>>;
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
 }
-
-void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    ggml_cuda_cpy(ctx, src0, dst);
-}

ggml/src/ggml-cuda/cpy.cuh

@@ -2,6 +2,10 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);
+
+void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream);

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

@@ -793,6 +793,8 @@ void launch_fattn(
     GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
         "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
 
+    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
     const int id  = ggml_cuda_get_device();
@@ -876,7 +878,7 @@ void launch_fattn(
     // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
     // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
     //     multiple sequences of possibly different lengths.
-    if (mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+    if (mask && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
         const int s31 = mask->nb[1] / sizeof(half2);
         const int s33 = mask->nb[3] / sizeof(half2);
 
@@ -895,7 +897,6 @@ void launch_fattn(
     const dim3 block_dim(warp_size, nwarps, 1);
     int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
     CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
-    GGML_ASSERT(max_blocks_per_sm > 0);
     int parallel_blocks = max_blocks_per_sm;
 
     dim3 blocks_num;
@@ -915,7 +916,8 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
-        const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 1) / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+        GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
+        const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
 
         // parallel_blocks must not be larger than what the tensor size allows:
         parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
@@ -944,7 +946,7 @@ void launch_fattn(
 
         blocks_num.x = ntiles_x;
         blocks_num.y = parallel_blocks;
-        blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
+        blocks_num.z = Q->ne[2]*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));

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

@@ -1,45 +1,755 @@
 #include "common.cuh"
+#include "fattn-common.cuh"
 #include "fattn-tile.cuh"
-#include "fattn-wmma-f16.cuh"
 
-void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-    switch (K->ne[0]) {
-        case  40: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case< 40,  40>(ctx, dst);
-        } break;
+// kq_stride == number of KQ rows to process per iteration
+// kq_nbatch == number of K columns to load in parallel for KQ calculation
+
+static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int cc, const int warp_size) {
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA(cc)) {
+            switch (D) {
+                case 64:
+                    return 128;
+                case 128:
+                case 256:
+                    return ncols <= 16 ? 128 : 64;
+                default:
+                    GGML_ABORT("fatal error");
+                    return -1;
+            }
+        }
+        switch (D) {
+            case 64:
+                return ncols == 32 ? 128 : 64;
+            case 128:
+                return ncols == 32 ? 64 : 32;
+            case 256:
+                return 32;
+            default:
+                GGML_ABORT("fatal error");
+                return -1;
+        }
+    }
+    if (fast_fp16_available(cc)) {
+        switch (D) {
+            case 64:
+            case 128:
+            case 256:
+                return ncols <= 16 ? 128 : 64;
+            default:
+                GGML_ABORT("fatal error");
+                return -1;
+        }
+    }
+    switch (D) {
+        case 64:
+            return ncols <= 16 ? 128 : 64;
+        case 128:
+            return ncols <= 16 ? 64 : 32;
+        case 256:
+            return 32;
+        default:
+            GGML_ABORT("fatal error");
+            return -1;
+    }
+    GGML_UNUSED(warp_size);
+}
+
+static constexpr __device__ int fattn_tile_get_kq_stride_device(int D, int ncols, int warp_size) {
+#ifdef GGML_USE_HIP
+#ifdef RDNA
+    switch (D) {
+        case 64:
+            return 128;
+        case 128:
+        case 256:
+            return ncols <= 16 ? 128 : 64;
+        default:
+            return -1;
+    }
+#else
+    switch (D) {
+        case 64:
+            return ncols == 32 ? 128 : 64;
+        case 128:
+            return ncols == 32 ? 64 : 32;
+        case 256:
+            return 32;
+        default:
+            return -1;
+    }
+#endif // RDNA
+#else
+#ifdef FAST_FP16_AVAILABLE
+    switch (D) {
+        case 64:
+        case 128:
+        case 256:
+            return ncols <= 16 ? 128 : 64;
+        default:
+            return -1;
+    }
+#else
+    switch (D) {
+        case 64:
+            return ncols <= 16 ? 128 : 64;
+        case 128:
+            return ncols <= 16 ? 64 : 32;
+        case 256:
+            return 32;
+        default:
+            return -1;
+    }
+#endif // FAST_FP16_AVAILABLE
+#endif // GGML_USE_HIP
+    GGML_UNUSED_VARS(ncols, warp_size);
+}
+
+static constexpr __device__ int fattn_tile_get_kq_nbatch_device(int D, int ncols, int warp_size) {
+#ifdef GGML_USE_HIP
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+        case 256:
+            return 128;
+        default:
+            return -1;
+    }
+#else
+#ifdef FAST_FP16_AVAILABLE
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+        case 256:
+            return 128;
+        default:
+            return -1;
+    }
+#else
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+            return 128;
+        case 256:
+            return ncols <= 16 ? 128 : 64;
+        default:
+            return -1;
+    }
+#endif // FAST_FP16_AVAILABLE
+#endif // GGML_USE_HIP
+    GGML_UNUSED_VARS(ncols, warp_size);
+}
+
+static int fattn_tile_get_nthreads_host(const int cc, const int ncols) {
+    return 256;
+    GGML_UNUSED_VARS(cc, ncols);
+}
+
+static constexpr __device__ int fattn_tile_get_nthreads_device(int ncols) {
+    return 256;
+    GGML_UNUSED(ncols);
+}
+
+static constexpr __device__ int fattn_tile_get_occupancy_device(int ncols) {
+#ifdef RDNA
+    return 3;
+#else
+    return ncols <= 16 ? 3 : 2;
+#endif // RDNA
+    GGML_UNUSED(ncols);
+}
+
+template<int D, int ncols, bool use_logit_softcap> // D == head size
+__launch_bounds__(fattn_tile_get_nthreads_device(ncols), fattn_tile_get_occupancy_device(ncols))
+static __global__ void flash_attn_tile(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        const char * __restrict__ sinks,
+        const int  * __restrict__ KV_max,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
+#ifdef FLASH_ATTN_AVAILABLE
+
+    // Skip unused kernel variants for faster compilation:
+#ifdef FP16_MMA_AVAILABLE
+    NO_DEVICE_CODE;
+    return;
+#endif // FP16_MMA_AVAILABLE
+
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int warp_size = 32;
+    constexpr int nwarps    = fattn_tile_get_nthreads_device(ncols) / warp_size;
+    constexpr int kq_stride = fattn_tile_get_kq_stride_device(D, ncols, warp_size);
+    static_assert(kq_stride % warp_size == 0, "kq_stride not divisable by warp_size.");
+    constexpr int kq_nbatch = fattn_tile_get_kq_nbatch_device(D, ncols, warp_size);
+    static_assert(kq_nbatch % (2*warp_size) == 0, "bad kq_nbatch");
+
+    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
+
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
+    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    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2 * K_h2   = (const half2 *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2 * V_h2   = (const half2 *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
+    const float * sinksf = (const float *) (sinks);
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
+
+    constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
+    constexpr int cpy_ne = cpy_nb / 4;
+
+    constexpr int cpw = ncols/nwarps; // cols per warp
+
+    // softmax_iter_j == number of KQ columns for which to calculate softmax in parallel.
+    // KQ is originall 2D but uses a Z-shaped memory pattern for larger reads/writes.
+#ifdef FAST_FP16_AVAILABLE
+    constexpr int softmax_iter_j = cpw < 2*cpy_ne ? cpw : 2*cpy_ne;
+
+    __shared__ half  KQ[ncols/softmax_iter_j][kq_stride][softmax_iter_j];
+    __shared__ half2 Q_tmp[ncols][D/2];
+    __shared__ half2 KV_tmp[kq_stride * (kq_nbatch/2 + cpy_ne)]; // Padded to avoid memory bank conflicts.
+    half2 VKQ[cpw][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
+#else
+    constexpr int softmax_iter_j = cpw < 1*cpy_ne ? cpw : 1*cpy_ne;
+
+    __shared__ float KQ[ncols/softmax_iter_j][kq_stride][softmax_iter_j];
+    __shared__ float Q_tmp[ncols][D];
+    __shared__ float KV_tmp[kq_stride * (kq_nbatch + cpy_ne)]; // Padded to avoid memory bank conflicts.
+    float2 VKQ[cpw][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
+#endif // FAST_FP16_AVAILABLE
+    static_assert(cpw % softmax_iter_j == 0, "bad softmax_iter_j");
+
+    float KQ_max[cpw];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        KQ_max[j0/nwarps] = -FLT_MAX/2.0f;
+    }
+    float KQ_sum[cpw] = {0.0f};
+
+    // Load Q data, convert to FP16 if fast.
+#pragma unroll
+    for (int j0 = 0; j0 < cpw; ++j0) {
+        const int j = j0 + threadIdx.y*cpw;
+
+        constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
+            float tmp_f[cpy_ne_D] = {0.0f};
+            if (ic0 + j < ne01) {
+                ggml_cuda_memcpy_1<sizeof(tmp_f)>(tmp_f, &Q_f[j*(nb01/sizeof(float)) + i0 + threadIdx.x*cpy_ne_D]);
+            }
+
+#pragma unroll
+            for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
+                tmp_f[i1] *= scale;
+            }
+
+#ifdef FAST_FP16_AVAILABLE
+            half2 tmp_h2[cpy_ne_D/2];
+#pragma unroll
+            for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
+                tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
+            }
+            ggml_cuda_memcpy_1<sizeof(tmp_h2)>(&Q_tmp[j][i0/2 + threadIdx.x*(cpy_ne_D/2)], tmp_h2);
+#else
+            ggml_cuda_memcpy_1<sizeof(tmp_f)> (&Q_tmp[j][i0   + threadIdx.x* cpy_ne_D],    tmp_f);
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+    __syncthreads();
+
+    // Main loop over KV cache:
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*kq_stride; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*kq_stride) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float KQ_max_new[cpw];
+#pragma unroll
+        for (int j = 0; j < cpw; ++j) {
+            KQ_max_new[j] = KQ_max[j];
+        }
+
+        float KQ_acc[kq_stride/warp_size][cpw] = {{0.0f}}; // Accumulators for KQ matrix multiplication.
+
+        // KQ = K @ Q matrix multiplication:
+#pragma unroll
+        for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += kq_nbatch) {
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += nwarps) {
+                const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#ifdef FAST_FP16_AVAILABLE
+                constexpr int cpy_ne_kqnb = cpy_ne < kq_nbatch/(2*warp_size) ? cpy_ne : kq_nbatch/(2*warp_size);
+#pragma unroll
+                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += warp_size*cpy_ne_kqnb) {
+                    ggml_cuda_memcpy_1<cpy_ne_kqnb*4>(
+                        &KV_tmp[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1 + threadIdx.x*cpy_ne_kqnb],
+                        &K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1 + threadIdx.x*cpy_ne_kqnb]);
+                }
+#else
+                constexpr int cpy_ne_kqnb = cpy_ne < kq_nbatch/warp_size ? cpy_ne : kq_nbatch/warp_size;
+#pragma unroll
+                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += warp_size*cpy_ne_kqnb) {
+                    half2 tmp_h2[cpy_ne_kqnb/2];
+                    ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
+                        tmp_h2, &K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1/2 + threadIdx.x*(cpy_ne_kqnb/2)]);
+
+                    float2 tmp_f2[cpy_ne_kqnb/2];
+#pragma unroll
+                    for (int k_KQ_2 = 0; k_KQ_2 < cpy_ne_kqnb/2; ++k_KQ_2) {
+                        tmp_f2[k_KQ_2] = __half22float2(tmp_h2[k_KQ_2]);
+                    }
+                    ggml_cuda_memcpy_1<sizeof(tmp_f2)>(
+                        &KV_tmp[i_KQ*(kq_nbatch + cpy_ne) + k_KQ_1 + threadIdx.x*cpy_ne_kqnb], tmp_f2);
+                }
+#endif // FAST_FP16_AVAILABLE
+            }
+
+            __syncthreads();
+
+#ifdef FAST_FP16_AVAILABLE
+#pragma unroll
+            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += cpy_ne) {
+                half2 K_k[kq_stride/warp_size][cpy_ne];
+                half2 Q_k[cpw][cpy_ne];
+#else
+#pragma unroll
+            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += cpy_ne) {
+                float K_k[kq_stride/warp_size][cpy_ne];
+                float Q_k[cpw][cpy_ne];
+#endif // FAST_FP16_AVAILABLE
+
+#pragma unroll
+                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
+                    const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#ifdef FAST_FP16_AVAILABLE
+                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1]);
+#else
+                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp[i_KQ*(kq_nbatch   + cpy_ne) + k_KQ_1]);
+#endif // FAST_FP16_AVAILABLE
+                }
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
+                    const int j_KQ = j_KQ_0 + threadIdx.y*cpw;
+
+#ifdef FAST_FP16_AVAILABLE
+                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0], &Q_tmp[j_KQ][k_KQ_0/2 + k_KQ_1]);
+#else
+                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0], &Q_tmp[j_KQ][k_KQ_0   + k_KQ_1]);
+#endif // FAST_FP16_AVAILABLE
+                }
+
+#pragma unroll
+                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
+#pragma unroll
+                    for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
+#pragma unroll
+                        for (int k = 0; k < cpy_ne; ++k) {
+                            ggml_cuda_mad(KQ_acc[i_KQ_0/warp_size][j_KQ_0], K_k[i_KQ_0/warp_size][k], Q_k[j_KQ_0][k]);
+                        }
+                    }
+                }
+            }
+
+            if (k_KQ_0 + kq_nbatch < D) {
+                __syncthreads(); // Sync not needed on last iteration.
+            }
+        }
+
+        // Apply logit softcap, mask, update KQ_max:
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
+                const int j_KQ = j_KQ_0 + threadIdx.y*cpw;
+
+                if (use_logit_softcap) {
+                    KQ_acc[i_KQ_0/warp_size][j_KQ_0] = logit_softcap * tanhf(KQ_acc[i_KQ_0/warp_size][j_KQ_0]);
+                }
+
+                KQ_acc[i_KQ_0/warp_size][j_KQ_0] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                KQ_max_new[j_KQ_0] = fmaxf(KQ_max_new[j_KQ_0], KQ_acc[i_KQ_0/warp_size][j_KQ_0]);
+            }
+        }
+
+        __syncthreads();
+
+        // Calculate KQ softmax, write to shared KQ buffer, re-scale VKQ accumulators:
+#pragma unroll
+        for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
+#ifdef FAST_FP16_AVAILABLE
+            half  tmp[kq_stride/warp_size][softmax_iter_j];
+#else
+            float tmp[kq_stride/warp_size][softmax_iter_j];
+#endif // FAST_FP16_AVAILABLE
+
+#pragma unroll
+            for (int j1 = 0; j1 < softmax_iter_j; ++j1) {
+                KQ_max_new[j0+j1] = warp_reduce_max<warp_size>(KQ_max_new[j0+j1]);
+                const float KQ_max_scale = expf(KQ_max[j0+j1] - KQ_max_new[j0+j1]);
+                KQ_max[j0+j1] = KQ_max_new[j0+j1];
+
+                float KQ_sum_add = 0.0f;
+#pragma unroll
+                for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
+                    const float val = expf(KQ_acc[i0/warp_size][j0+j1] - KQ_max[j0+j1]);
+                    KQ_sum_add += val;
+                    tmp[i0/warp_size][j1] = val;
+                }
+                KQ_sum[j0+j1] = KQ_sum[j0+j1]*KQ_max_scale + KQ_sum_add;
+
+#ifdef FAST_FP16_AVAILABLE
+                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                    VKQ[j0+j1][i0/warp_size] *= KQ_max_scale_h2;
+                }
+#else
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                    VKQ[j0+j1][i0/warp_size].x *= KQ_max_scale;
+                    VKQ[j0+j1][i0/warp_size].y *= KQ_max_scale;
+                }
+#endif // FAST_FP16_AVAILABLE
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
+                const int i = i0 + threadIdx.x;
+
+                ggml_cuda_memcpy_1<sizeof(tmp[0])>(
+                    KQ[j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j)][i], tmp[i0/warp_size]);
+            }
+        }
+
+        // VKQ = V @ KQ matrix multiplication:
+        constexpr int V_cols_per_iter = kq_stride*kq_nbatch / D; // Number of V columns that fit in SRAM for K.
+        static_assert(kq_stride % V_cols_per_iter == 0, "bad V_cols_per_iter");
+#pragma unroll
+        for (int k0 = 0; k0 < kq_stride; k0 += V_cols_per_iter) {
+#pragma unroll
+            for (int k1 = 0; k1 < V_cols_per_iter; k1 += nwarps) {
+                const int k_tile = k1 + threadIdx.y;
+
+#ifdef FAST_FP16_AVAILABLE
+                constexpr int cpy_ne_D = cpy_ne < D/(2*warp_size) ? cpy_ne : D/(2*warp_size);
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
+                    ggml_cuda_memcpy_1<cpy_ne_D*4>(
+                        &KV_tmp[k_tile*(D/2) + i0 + threadIdx.x*cpy_ne_D],
+                        &V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i0 + threadIdx.x*cpy_ne_D]);
+                }
+#else
+                constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
+#pragma unroll
+                for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
+                    half2 tmp_h2[cpy_ne_D/2];
+                    ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
+                        tmp_h2, &V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i0/2 + threadIdx.x*(cpy_ne_D/2)]);
+
+                    float2 tmp_f2[cpy_ne_D/2];
+#pragma unroll
+                    for (int i1 = 0; i1 < cpy_ne_D/2; ++i1) {
+                        tmp_f2[i1] = __half22float2(tmp_h2[i1]);
+                    }
+                    ggml_cuda_memcpy_1<sizeof(tmp_f2)>(
+                        &KV_tmp[k_tile*D + i0 + threadIdx.x*cpy_ne_D], tmp_f2);
+                }
+#endif // FAST_FP16_AVAILABLE
+            }
+
+            __syncthreads();
+
+#ifdef FAST_FP16_AVAILABLE
+#pragma unroll
+            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
+                half2 V_k[(D/2)/warp_size];
+                half2 KQ_k[cpw];
+
+                constexpr int cpy_ne_D = cpy_ne/2 < (D/2)/warp_size ? cpy_ne/2 : (D/2)/warp_size;
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
+                    ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/warp_size], &KV_tmp[k1*(D/2) + i0 + threadIdx.x*cpy_ne_D]);
+                }
+#pragma unroll
+                for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
+                    const int j = j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j);
+
+                    half tmp[softmax_iter_j];
+                    ggml_cuda_memcpy_1<softmax_iter_j*sizeof(half)>(
+                        &tmp, KQ[j][k0 + k1]);
+#pragma unroll
+                    for (int j1 = 0; j1 < softmax_iter_j; ++j1) {
+                        KQ_k[j0+j1] = __half2half2(tmp[j1]);
+                    }
+                }
+
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+#pragma unroll
+                    for (int j0 = 0; j0 < cpw; ++j0) {
+                        VKQ[j0][i0/warp_size] += V_k[i0/warp_size]*KQ_k[j0];
+                    }
+                }
+            }
+#else
+#pragma unroll
+            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
+                float2 V_k[(D/2)/warp_size];
+                float  KQ_k[cpw];
+
+                constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
+#pragma unroll
+                for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
+                    ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/(2*warp_size)], &KV_tmp[k1*D + i0 + threadIdx.x*cpy_ne_D]);
+                }
+#pragma unroll
+                for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
+                    const int j = j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j);
+
+                    ggml_cuda_memcpy_1<softmax_iter_j*sizeof(float)>(
+                        &KQ_k[j0], KQ[j][k0 + k1]);
+                }
+
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+#pragma unroll
+                    for (int j0 = 0; j0 < cpw; ++j0) {
+                        VKQ[j0][i0/warp_size].x += V_k[i0/warp_size].x*KQ_k[j0];
+                        VKQ[j0][i0/warp_size].y += V_k[i0/warp_size].y*KQ_k[j0];
+                    }
+                }
+            }
+#endif // FAST_FP16_AVAILABLE
+
+            __syncthreads();
+        }
+    }
+
+
+    // Attention sink: adjust running max and sum once per head
+    if (sinksf && blockIdx.y == 0) {
+        const float sink = sinksf[head];
+
+#pragma unroll
+        for (int j0 = 0; j0 < cpw; ++j0) {
+            float KQ_max_new_j = fmaxf(KQ_max[j0], sink);
+            KQ_max_new_j = warp_reduce_max<warp_size>(KQ_max_new_j);
+
+            const float KQ_max_scale = expf(KQ_max[j0] - KQ_max_new_j);
+            KQ_max[j0] = KQ_max_new_j;
+
+            const float val = expf(sink - KQ_max[j0]);
+            KQ_sum[j0] = KQ_sum[j0] * KQ_max_scale;
+            if (threadIdx.x == 0) {
+                KQ_sum[j0] += val;
+            }
+
+#ifdef FAST_FP16_AVAILABLE
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                VKQ[j0][i0/warp_size] *= KQ_max_scale_h2;
+            }
+#else
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                VKQ[j0][i0/warp_size].x *= KQ_max_scale;
+                VKQ[j0][i0/warp_size].y *= KQ_max_scale;
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
+        KQ_sum[j_VKQ_0] = warp_reduce_sum<warp_size>(KQ_sum[j_VKQ_0]);
+    }
+    if (gridDim.y == 1) {
+#pragma unroll
+        for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
+#ifdef FAST_FP16_AVAILABLE
+            const half2 KQ_sum_j_inv = make_half2(1.0f/KQ_sum[j_VKQ_0], 1.0f/KQ_sum[j_VKQ_0]);
+#pragma unroll
+            for (int i = 0; i < (D/2)/warp_size; ++i) {
+                VKQ[j_VKQ_0][i] *= KQ_sum_j_inv;
+            }
+#else
+            const float KQ_sum_j_inv = 1.0f/KQ_sum[j_VKQ_0];
+#pragma unroll
+            for (int i = 0; i < (D/2)/warp_size; ++i) {
+                VKQ[j_VKQ_0][i].x *= KQ_sum_j_inv;
+                VKQ[j_VKQ_0][i].y *= KQ_sum_j_inv;
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+    // Write back results:
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y*cpw;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
+
+#ifdef FAST_FP16_AVAILABLE
+        constexpr int cpy_ne_D = cpy_ne/2 < (D/2)/warp_size ? cpy_ne/2 : (D/2)/warp_size;
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
+            float2 tmp[cpy_ne_D];
+#pragma unroll
+            for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
+                tmp[i1] = __half22float2(VKQ[j_VKQ_0][i0/warp_size + i1]);
+            }
+            ggml_cuda_memcpy_1<sizeof(tmp)>(&dst[j_dst_unrolled*D + 2*i0 + threadIdx.x*(2*cpy_ne_D)], tmp);
+        }
+#else
+        constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
+            ggml_cuda_memcpy_1<cpy_ne_D*4>(
+                &dst[j_dst_unrolled*D + i0 + threadIdx.x*cpy_ne_D], &VKQ[j_VKQ_0][i0/(2*warp_size)]);
+        }
+#endif // FAST_FP16_AVAILABLE
+
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[j_dst_unrolled] = make_float2(KQ_max[j_VKQ_0], KQ_sum[j_VKQ_0]);
+        }
+    }
+#else
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
+    NO_DEVICE_CODE;
+#endif // FLASH_ATTN_AVAILABLE
+}
+
+template <int D, bool use_logit_softcap>
+static void launch_fattn_tile_switch_ncols(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    const int id        = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[id].cc;
+    const int warp_size = 32;
+
+    constexpr size_t nbytes_shared = 0;
+
+#ifdef GGML_USE_HIP
+    if constexpr (D <= 128) {
+        if (Q->ne[1] > 32) {
+            constexpr int cols_per_block = 64;
+            const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
+            fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
+            const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
+            launch_fattn<D, cols_per_block, 1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
+            return;
+        }
+    }
+#endif // GGML_USE_HIP
+
+    if (Q->ne[1] > 16) {
+        constexpr int cols_per_block = 32;
+        const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
+        fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
+        const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
+        launch_fattn<D, cols_per_block, 1>
+            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
+        return;
+    }
+
+    constexpr int cols_per_block = 16;
+    const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
+    fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
+    const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
+    launch_fattn<D, cols_per_block, 1>
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
+}
+
+template <bool use_logit_softcap>
+static void launch_fattn_tile_switch_head_size(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
         case  64: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case< 64,  64>(ctx, dst);
-        } break;
-        case  80: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case< 80,  80>(ctx, dst);
-        } break;
-        case  96: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case< 96,  96>(ctx, dst);
-        } break;
-        case 112: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case<112, 112>(ctx, dst);
+            launch_fattn_tile_switch_ncols< 64, use_logit_softcap>(ctx, dst);
         } break;
         case 128: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case<128, 128>(ctx, dst);
+            launch_fattn_tile_switch_ncols<128, use_logit_softcap>(ctx, dst);
         } break;
         case 256: {
-            GGML_ASSERT(V->ne[0] == K->ne[0]);
-            ggml_cuda_flash_attn_ext_tile_case<256, 256>(ctx, dst);
-        } break;
-        case 576: {
-            GGML_ASSERT(V->ne[0] == 512);
-            ggml_cuda_flash_attn_ext_tile_case<576, 512>(ctx, dst);
+            launch_fattn_tile_switch_ncols<256, use_logit_softcap>(ctx, dst);
         } break;
         default: {
             GGML_ABORT("Unsupported head size");
         } break;
     }
 }
+
+void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
+    }
+}