CUDA: add fp kernel for larger batch size MoE (#16512)

* CUDA: kernel for larger batch sizes for MoE

* WIP

* WIP

* WIP

* WIP

* WIP

* WIP

* fixup

* tests

* Move mmq_ids_helper to mmid

* cleanup

* Remove redundant checks

.devops/intel.Dockerfile

@@ -1,8 +1,8 @@
-ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.2.2-0-devel-ubuntu24.04
 
 ## Build Image
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+FROM intel/deep-learning-essentials:$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/oneapi-basekit:$ONEAPI_VERSION AS base
+FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
 
 RUN apt-get update \
     && apt-get install -y libgomp1 curl\

.devops/nix/package.nix

@@ -128,10 +128,6 @@ 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=6.4
-ARG AMDGPU_VERSION=6.4
+ARG ROCM_VERSION=7.0
+ARG AMDGPU_VERSION=7.0
 
 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
@@ -13,9 +13,8 @@ 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, 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
+# 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
 
 ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
 #ARG ROCM_DOCKER_ARCH='gfx1151'
@@ -36,13 +35,10 @@ 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

@@ -0,0 +1,36 @@
+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

@@ -0,0 +1,20 @@
+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

@@ -0,0 +1,20 @@
+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

@@ -0,0 +1,27 @@
+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

@@ -0,0 +1,15 @@
+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-cache.yml

@@ -0,0 +1,89 @@
+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

@@ -258,31 +258,29 @@ jobs:
     runs-on: ubuntu-24.04
 
     env:
+      # Make sure this is in sync with build-cache.yml
       SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
-      SPACEMIT_IME_TOOLCHAIN_PATH: "spacemit-toolchain-linux-glibc-x86_64"
 
     steps:
       - uses: actions/checkout@v4
 
-      - name: Cache Toolchain
+      - name: Use SpacemiT Toolchain Cache
         uses: actions/cache@v4
-        id: cache-spacemit-ime-cross-toolchain
+        id: cache-toolchain
         with:
-          path: ./${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
-          key: ${{ runner.os }}-spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
+          path: ./spacemit_toolchain
+          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
 
-      - name: Setup Toolchain
-        if: steps.cache-spacemit-ime-cross-toolchain.outputs.cache-hit != 'true'
-        run: |
-          wget --quiet --no-check-certificate https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}.tar.xz -O ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}.tar.xz
-          rm -rf ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
-          mkdir -p ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
-          tar xf ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}.tar.xz -C ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }} --strip-components=1
-          rm -rf ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}.tar.xz
+      - 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=${PWD}/${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
+          export RISCV_ROOT_PATH=${PWD}/spacemit_toolchain
           cmake -B build -DLLAMA_CURL=OFF \
                          -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_OPENMP=OFF \

.github/workflows/build.yml

@@ -97,7 +97,7 @@ jobs:
           ctest -L 'main|curl' --verbose --timeout 900
 
   macOS-latest-cmake-x64:
-    runs-on: macos-13
+    runs-on: macos-15-intel
 
     steps:
       - name: Clone
@@ -207,7 +207,7 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-cpu-cmake
+          key: ubuntu-cpu-cmake-${{ matrix.build }}
           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-22-cmake-vulkan:
-    runs-on: ubuntu-22.04
+  ubuntu-24-cmake-vulkan-deb:
+    runs-on: ubuntu-24.04
 
     steps:
       - name: Clone
@@ -398,20 +398,72 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-22-cmake-vulkan
+          key: ubuntu-24-cmake-vulkan-deb
           evict-old-files: 1d
 
       - name: Dependencies
         id: depends
         run: |
-          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 vulkan-sdk libcurl4-openssl-dev
+          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 $(nproc)
+
+  ubuntu-24-cmake-vulkan:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-24-cmake-vulkan
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo add-apt-repository -y ppa:kisak/kisak-mesa
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
+
+      - name: Get latest Vulkan SDK version
+        id: vulkan_sdk_version
+        run: |
+          echo "VULKAN_SDK_VERSION=$(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 }}
+
+      - 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)
@@ -421,11 +473,12 @@ 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-22-cmake-webgpu:
-    runs-on: ubuntu-22.04
+  ubuntu-24-cmake-webgpu:
+    runs-on: ubuntu-24.04
 
     steps:
       - name: Clone
@@ -435,16 +488,34 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-22-cmake-webgpu
+          key: ubuntu-24-cmake-webgpu
           evict-old-files: 1d
 
-      - name: Vulkan SDK Dependencies
-        id: vulkan-depends
+      - name: Dependencies
+        id: depends
         run: |
-          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 add-apt-repository -y ppa:kisak/kisak-mesa
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+          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 }}
 
       - name: Dawn Dependency
         id: dawn-depends
@@ -487,7 +558,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev rocwmma-dev
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1059,7 +1130,7 @@ jobs:
         shell: bash
 
     env:
-      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_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_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:
@@ -1090,6 +1161,7 @@ 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:
@@ -1097,38 +1169,25 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      - name: Clone rocWMMA repository
-        id: clone_rocwmma
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
         run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-${{ env.ROCM_VERSION }} --depth 1
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/${{ 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
 
-      - name: Cache ROCm Installation
-        id: cache-rocm
+      - name: Use ROCm Installation Cache
         uses: actions/cache@v4
+        id: cache-rocm
         with:
           path: C:\Program Files\AMD\ROCm
           key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
-      - name: Install ROCm
+      - name: Setup ROCm
         if: steps.cache-rocm.outputs.cache-hit != 'true'
-        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"
+        uses: ./.github/actions/windows-setup-rocm
+        with:
+          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
 
       - name: Verify ROCm
         id: verify
@@ -1161,8 +1220,9 @@ 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('\', '/'))/rocwmma/library/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
             -DCMAKE_BUILD_TYPE=Release `
+            -DROCM_DIR="${env:HIP_PATH}" `
             -DGGML_HIP=ON `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_RPC=ON `
@@ -1488,3 +1548,29 @@ 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

@@ -89,12 +89,15 @@ 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=$SERVERTAGS"  # print out for debugging
@@ -131,11 +134,14 @@ jobs:
           target: full
           provenance: false
           # using github experimental cache
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ 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: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
@@ -150,11 +156,14 @@ jobs:
           target: light
           provenance: false
           # using github experimental cache
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ 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: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
@@ -169,11 +178,14 @@ jobs:
           target: server
           provenance: false
           # using github experimental cache
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ 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-13
+    runs-on: macos-15-intel
 
     steps:
       - name: Clone
@@ -150,7 +150,7 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-cpu-cmake
+          key: ubuntu-cpu-cmake-${{ matrix.build }}
           evict-old-files: 1d
 
       - name: Dependencies
@@ -462,7 +462,7 @@ jobs:
         shell: bash
 
     env:
-      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_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_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"
 
@@ -505,6 +505,7 @@ 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 "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
@@ -513,10 +514,15 @@ jobs:
           cp "${{ 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/*
 
@@ -543,10 +549,12 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      - name: Clone rocWMMA repository
-        id: clone_rocwmma
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
         run: |
-          git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+          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
 
       - name: Cache ROCm Installation
         id: cache-rocm
@@ -601,7 +609,7 @@ 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('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.0.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `

CODEOWNERS

@@ -2,7 +2,7 @@
 # multiplie collaborators per item can be specified
 
 /.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @slaren
+/.github/actions/                       @slaren @CISC
 /.github/workflows/                     @CISC
 /.github/workflows/release.yml          @slaren
 /.github/workflows/winget.yml           @slaren
@@ -14,6 +14,7 @@
 /common/build-info.*                    @ggerganov
 /common/common.*                        @ggerganov
 /common/console.*                       @ggerganov
+/common/http.*                          @angt
 /common/llguidance.*                    @ggerganov
 /common/log.*                           @ggerganov
 /common/sampling.*                      @ggerganov
@@ -58,6 +59,9 @@
 /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
@@ -66,6 +70,7 @@
 /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

ci/run.sh

@@ -22,6 +22,9 @@
 # # 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>"
@@ -34,9 +37,9 @@ 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/../
@@ -115,6 +118,34 @@ if [ ! -z ${GG_BUILD_NO_SVE} ]; then
     CMAKE_EXTRA="${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
@@ -512,12 +543,7 @@ 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/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
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.json
 
     path_models="../models-mnt/rerank-tiny"
 
@@ -607,6 +633,7 @@ 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
 
@@ -624,4 +651,6 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run ctest_with_model_release
 fi
 
+cat $OUT/README.md
+
 exit $ret

common/CMakeLists.txt

@@ -56,6 +56,7 @@ 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

@@ -32,13 +32,11 @@
 #include <thread>
 #include <vector>
 
-//#define LLAMA_USE_CURL
-
 #if defined(LLAMA_USE_CURL)
 #include <curl/curl.h>
 #include <curl/easy.h>
 #else
-#include <cpp-httplib/httplib.h>
+#include "http.h"
 #endif
 
 #ifdef __linux__
@@ -54,6 +52,13 @@
 #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 = {
@@ -100,6 +105,14 @@ 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;
@@ -581,78 +594,11 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 
 #else
 
-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);
+static void print_progress(size_t current, size_t total) {
+    if (!is_output_a_tty()) {
+        return;
     }
 
-    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;
     }
@@ -740,7 +686,7 @@ static bool common_download_file_single_online(const std::string & url,
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
 
-    auto [cli, parts] = http_client(url);
+    auto [cli, parts] = common_http_client(url);
 
     httplib::Headers default_headers = {{"User-Agent", "llama-cpp"}};
     if (!bearer_token.empty()) {
@@ -820,7 +766,7 @@ 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__, show_masked_url(parts).c_str(), path_temporary.c_str(), etag.c_str());
+                __func__, common_http_show_masked_url(parts).c_str(), path_temporary.c_str(), etag.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) {
@@ -848,7 +794,7 @@ 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] = http_client(url);
+    auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {{"User-Agent", "llama-cpp"}};
     for (const auto & header : params.headers) {
@@ -1669,18 +1615,14 @@ static void add_rpc_devices(const std::string & servers) {
     if (!rpc_reg) {
         throw std::invalid_argument("failed to find RPC backend");
     }
-    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");
+    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");
     }
     for (const auto & server : rpc_servers) {
-        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");
-        }
+        auto reg = ggml_backend_rpc_add_server_fn(server.c_str());
+        ggml_backend_register(reg);
     }
 }
 
@@ -1986,13 +1928,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
-        {"--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),
+        {"--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),
         [](common_params & params, int value) {
-            params.n_swa_checkpoints = value;
+            params.n_ctx_checkpoints = value;
         }
-    ).set_env("LLAMA_ARG_SWA_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).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}));
     add_opt(common_arg(
         {"--kv-unified", "-kvu"},
         string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
@@ -2642,6 +2592,13 @@ 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(
@@ -3401,7 +3358,7 @@ 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();
@@ -3483,7 +3440,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--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` (except in streaming mode, which behaves as `none`)\n"
+        "- deepseek: puts thoughts in `message.reasoning_content`\n"
+        "- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`\n"
         "(default: auto)",
         [](common_params & params, const std::string & value) {
             params.reasoning_format = common_reasoning_format_from_name(value);
@@ -3612,21 +3570,23 @@ 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)",
@@ -3892,7 +3852,87 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
-    // model-specific
+    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
     add_opt(common_arg(
         {"--tts-oute-default"},
         string_format("use default OuteTTS models (note: can download weights from the internet)"),
@@ -3905,42 +3945,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
     add_opt(common_arg(
-        {"--embd-bge-small-en-default"},
-        string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
+        {"--embd-gemma-default"},
+        string_format("use default EmbeddingGemma model (note: can download weights from the internet)"),
         [](common_params & params) {
-            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.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.verbose_prompt = true;
             params.embedding = true;
         }
@@ -4035,96 +4049,65 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
     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 }));
+        {"--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}));
 
     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 }));
+        {"--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}));
 
     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 }));
+        {"--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}));
 
-
-    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 }));
+        {"--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}));
 
     return ctx_arg;
 }

common/chat-parser.cpp

@@ -3,9 +3,12 @@
 #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;
@@ -75,6 +78,35 @@ 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_));
@@ -137,6 +169,27 @@ 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()) {
@@ -149,28 +202,116 @@ 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);
         }
     };
-    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();
+
+    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 (!rest.empty()) {
                 handle_reasoning(rest, /* closed */ !is_partial());
             }
-            // 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);
-            // }
+            move_to(input_.size());
             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() {
@@ -291,7 +432,7 @@ 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(),
+                partial->json.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true),
                 /* .is_partial = */ false,
             };
         }
@@ -303,7 +444,7 @@ 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();
+            auto arguments = j.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true);
             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

@@ -64,6 +64,9 @@ 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

@@ -625,6 +625,7 @@ 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";
@@ -638,6 +639,7 @@ 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";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -801,6 +803,7 @@ 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);
     }
@@ -982,6 +985,65 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     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());
@@ -992,6 +1054,18 @@ 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;
 
@@ -1264,7 +1338,78 @@ 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;
@@ -2323,6 +2468,37 @@ 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_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>");
@@ -2567,6 +2743,11 @@ 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);
+    }
+
     // 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())) {
@@ -2595,6 +2776,10 @@ 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);
@@ -2679,6 +2864,7 @@ 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());
 }
 
@@ -2695,6 +2881,9 @@ 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;
@@ -2734,6 +2923,9 @@ 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;
         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 ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+    void set_tool_call_ids(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;
@@ -101,6 +101,7 @@ 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,
@@ -114,6 +115,7 @@ 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_COUNT, // Not a format, just the # formats
 };

common/common.cpp

@@ -1133,6 +1133,7 @@ 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
 
@@ -392,6 +392,7 @@ 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
 
@@ -424,7 +425,8 @@ 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_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
+    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.
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -432,7 +434,7 @@ 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_AUTO;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 

common/http.h

@@ -0,0 +1,73 @@
+#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

@@ -5,6 +5,7 @@
 #include <nlohmann/json.hpp>
 
 #include <string>
+#include <regex>
 
 using json = nlohmann::ordered_json;
 
@@ -168,6 +169,47 @@ 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) {
@@ -186,6 +228,9 @@ 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(':');
@@ -205,6 +250,9 @@ 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;
@@ -230,6 +278,9 @@ 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) {

convert_hf_to_gguf.py

@@ -93,13 +93,15 @@ class ModelBase:
     # Mistral format specifics
     is_mistral_format: bool = False
     disable_mistral_community_chat_template: bool = False
+    sentence_transformers_dense_modules: bool = False
 
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
                  small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
-                 disable_mistral_community_chat_template: bool = False):
+                 disable_mistral_community_chat_template: bool = False,
+                 sentence_transformers_dense_modules: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -114,6 +116,7 @@ class ModelBase:
         self.lazy = not eager or (remote_hf_model_id is not None)
         self.dry_run = dry_run
         self.remote_hf_model_id = remote_hf_model_id
+        self.sentence_transformers_dense_modules = sentence_transformers_dense_modules
         if remote_hf_model_id is not None:
             self.is_safetensors = True
 
@@ -891,6 +894,9 @@ class TextModel(ModelBase):
         if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
             # ref: https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base
             res = "llada-moe"
+        if chkhsh == "53e325976a6e142379c19b09afcae354f2f496f147afa8f9e189a33fe4e3024e":
+            # ref: https://huggingface.co/ibm-granite/granite-docling-258M
+            res = "granite-docling"
 
         if res is None:
             logger.warning("\n")
@@ -1325,6 +1331,7 @@ class MmprojModel(ModelBase):
         self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
 
         # load preprocessor config
+        self.preprocessor_config = {}
         if not self.is_mistral_format:
             with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
                 self.preprocessor_config = json.load(f)
@@ -1347,7 +1354,8 @@ class MmprojModel(ModelBase):
             self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
 
             # vision config
-            self.gguf_writer.add_vision_image_size(self.find_vparam(["image_size"]))
+            self.image_size = self.find_vparam(["image_size"])
+            self.gguf_writer.add_vision_image_size(self.image_size)
             self.gguf_writer.add_vision_patch_size(self.find_vparam(["patch_size"]))
             self.gguf_writer.add_vision_embedding_length(self.find_vparam(["hidden_size"]))
             self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size"]))
@@ -2378,6 +2386,10 @@ class SmolVLMModel(MmprojModel):
         self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("scale_factor", 2))
         self.gguf_writer.add_vision_use_gelu(True)
 
+        # Add the preprocessor longest edge size
+        preproc_image_size = self.preprocessor_config.get("size", {}).get("longest_edge", self.image_size)
+        self.gguf_writer.add_vision_preproc_image_size(preproc_image_size)
+
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         if ".embeddings." in name:
             return gguf.GGMLQuantizationType.F32
@@ -4250,7 +4262,8 @@ class Plamo2Model(TextModel):
         # This logic matches modeling_plamo.py's is_mamba function
         mamba_step = hparams.get("mamba_step", 2)
         mamba_enabled = hparams.get("mamba_enabled", True)
-        mamba_layers = []
+        num_key_value_heads = []
+        num_attention_heads = []
 
         if mamba_enabled:
             for i in range(block_count):
@@ -4260,17 +4273,21 @@ class Plamo2Model(TextModel):
                 else:
                     is_mamba = (i % mamba_step) != (mamba_step // 2)
                 if is_mamba:
-                    mamba_layers.append(0)
+                    num_key_value_heads.append(0)
+                    num_attention_heads.append(0)
                 else:
-                    mamba_layers.append(hparams.get("num_key_value_heads", 4))
+                    num_key_value_heads.append(hparams.get("num_key_value_heads", 4))
+                    num_attention_heads.append(hparams.get("num_attention_heads", 32))
 
-        if mamba_layers:
-            self.gguf_writer.add_head_count_kv(mamba_layers)
+        if num_key_value_heads and num_attention_heads:
+            self.gguf_writer.add_head_count_kv(num_key_value_heads)
+            self.gguf_writer.add_head_count(num_attention_heads)
 
         self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 2048))
         self.gguf_writer.add_embedding_length(hparams.get("hidden_size", 4096))
+        self.gguf_writer.add_key_length(hparams.get("hidden_size_per_head", 128))
+        self.gguf_writer.add_value_length(hparams.get("hidden_size_per_head", 128))
         self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 32))
         self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 10000))
 
@@ -5255,6 +5272,53 @@ class Gemma3Model(TextModel):
 @ModelBase.register("Gemma3TextModel")
 class EmbeddingGemma(Gemma3Model):
     model_arch = gguf.MODEL_ARCH.GEMMA_EMBEDDING
+    module_paths = []
+    dense_features_dims = {}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.sentence_transformers_dense_modules:
+            # read modules.json to determine if model has Dense layers
+            modules_file = self.dir_model / "modules.json"
+            if modules_file.is_file():
+                with open(modules_file, encoding="utf-8") as modules_json_file:
+                    mods = json.load(modules_json_file)
+                for mod in mods:
+                    if mod["type"] == "sentence_transformers.models.Dense":
+                        mod_path = mod["path"]
+                        # check if model.safetensors file for Dense layer exists
+                        model_tensors_file = self.dir_model / mod_path / "model.safetensors"
+                        if model_tensors_file.is_file():
+                            self.module_paths.append(mod_path)
+                            # read config.json of the Dense layer to get in/out features
+                            mod_conf_file = self.dir_model / mod_path / "config.json"
+                            if mod_conf_file.is_file():
+                                with open(mod_conf_file, encoding="utf-8") as mod_conf_json_file:
+                                    mod_conf = json.load(mod_conf_json_file)
+                                    # hparams dense_2_feat_out and dense_3_feat_in are required when loading model's dense weights
+                                    prefix = self._get_dense_prefix(mod_path)
+                                    if mod_conf["in_features"] is not None and mod_conf["out_features"] is not None:
+                                        self.dense_features_dims[prefix] = (mod_conf["in_features"], mod_conf["out_features"])
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        from safetensors.torch import load_file
+        module_paths = list(self.module_paths)
+        for i, module_path in enumerate(module_paths):
+            tensors_file = self.dir_model / module_path / "model.safetensors"
+            local_tensors = load_file(tensors_file)
+            tensor_name = self._get_dense_prefix(module_path)
+            for name, local_tensor in local_tensors.items():
+                if not name.endswith(".weight"):
+                    continue
+                orig_name = name.replace("linear", tensor_name)
+                name = self.map_tensor_name(orig_name)
+                yield name, local_tensor.clone()
+
+    @staticmethod
+    def _get_dense_prefix(module_path) -> str:
+        """Get the tensor name prefix for the Dense layer from module path."""
+        tensor_name = "dense_2" if module_path == "2_Dense" else "dense_3"
+        return tensor_name
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -5271,6 +5335,10 @@ class EmbeddingGemma(Gemma3Model):
             logger.info(f"Using original sliding_window from config: {orig_sliding_window} "
                         f"instead of {self.hparams['sliding_window']}")
             self.gguf_writer.add_sliding_window(orig_sliding_window)
+        if self.sentence_transformers_dense_modules:
+            for dense, dims in self.dense_features_dims.items():
+                logger.info(f"Setting dense layer {dense} in/out features to {dims}")
+                self.gguf_writer.add_dense_features_dims(dense, dims[0], dims[1])
 
         self._try_set_pooling_type()
 
@@ -5898,20 +5966,12 @@ class Mamba2Model(TextModel):
 class JambaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.JAMBA
 
-    def get_vocab_base_pre(self, tokenizer) -> str:
-        del tokenizer  # unused
-
-        return "gpt-2"
-
     def set_vocab(self):
         if (self.dir_model / "tokenizer.model").is_file():
-            # Using Jamba's tokenizer.json causes errors on model load
-            # (something about "byte not found in vocab"),
-            # but there's a working tokenizer.model
             self._set_vocab_sentencepiece()
         else:
-            # Some Jamba models only have a tokenizer.json, which works.
-            self._set_vocab_gpt2()
+            self._set_vocab_llama_hf()
+            self.gguf_writer.add_add_space_prefix(False)
 
     def set_gguf_parameters(self):
         d_model = self.find_hparam(["hidden_size", "mamba_d_model"])
@@ -8822,6 +8882,75 @@ class LFM2Model(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Lfm2MoeForCausalLM")
+class LFM2MoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LFM2MOE
+
+    def set_gguf_parameters(self):
+        # set num_key_value_heads only for attention layers
+        self.hparams["num_key_value_heads"] = [
+            self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
+            for layer_type in self.hparams["layer_types"]
+        ]
+
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["num_dense_layers"])
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        self.gguf_writer.add_shortconv_l_cache(self.hparams["conv_L_cache"])
+
+    # cache for experts weights for merging
+    _experts_cache: dict[int, dict[str, Tensor]] = {}
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # conv op requires 2d tensor
+        if 'conv.conv' in name:
+            data_torch = data_torch.squeeze(1)
+
+        if name.endswith(".expert_bias"):
+            name = name.replace(".expert_bias", ".expert_bias.bias")
+
+        # merge expert weights
+        if 'experts' in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            expert_cache = self._experts_cache.setdefault(bid, {})
+            expert_cache[name] = data_torch
+            expert_weights = ["w1", "w2", "w3"]
+
+            # not enough expert weights to merge
+            if len(expert_cache) < n_experts * len(expert_weights):
+                return []
+
+            tensors: list[tuple[str, Tensor]] = []
+            for w_name in expert_weights:
+                datas: list[Tensor] = []
+
+                for xid in range(n_experts):
+                    ename = f"model.layers.{bid}.feed_forward.experts.{xid}.{w_name}.weight"
+                    datas.append(expert_cache[ename])
+                    del expert_cache[ename]
+
+                data_torch = torch.stack(datas, dim=0)
+                merged_name = f"layers.{bid}.feed_forward.experts.{w_name}.weight"
+                new_name = self.map_tensor_name(merged_name)
+                tensors.append((new_name, data_torch))
+
+            del self._experts_cache[bid]
+            return tensors
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        assert not self._experts_cache
+
+
 @ModelBase.register("Lfm2VlForConditionalGeneration")
 class LFM2VLModel(MmprojModel):
     def __init__(self, *args, **kwargs):
@@ -8940,6 +9069,43 @@ class SmallThinkerModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("ApertusForCausalLM")
+class ApertusModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.APERTUS
+    undo_permute = False
+
+    _alpha_n = {}
+    _alpha_p = {}
+    _beta = {}
+    _eps = {}
+
+    def modify_tensors(self, data_torch, name, bid):
+        # Handle xIELU activation parameters
+        n_layers = self.hparams["num_hidden_layers"]
+        if name.endswith(".act_fn.alpha_n"):
+            self._alpha_n[bid] = data_torch.to("cpu").float().item()
+            if (len(self._alpha_n) == n_layers):
+                self.gguf_writer.add_xielu_alpha_n([self._alpha_n[k] for k in sorted(self._alpha_n)])
+            return []
+        if name.endswith(".act_fn.alpha_p"):
+            self._alpha_p[bid] = data_torch.to("cpu").float().item()
+            if (len(self._alpha_p) == n_layers):
+                self.gguf_writer.add_xielu_alpha_p([self._alpha_p[k] for k in sorted(self._alpha_p)])
+            return []
+        if name.endswith(".act_fn.beta"):
+            self._beta[bid] = data_torch.to("cpu").float().item()
+            if (len(self._beta) == n_layers):
+                self.gguf_writer.add_xielu_beta([self._beta[k] for k in sorted(self._beta)])
+            return []
+        if name.endswith(".act_fn.eps"):
+            self._eps[bid] = data_torch.to("cpu").float().item()
+            if (len(self._eps) == n_layers):
+                self.gguf_writer.add_xielu_eps([self._eps[k] for k in sorted(self._eps)])
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 class MistralModel(LlamaModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     model_name = "Mistral"
@@ -9107,7 +9273,7 @@ class LazyTorchTensor(gguf.LazyBase):
     def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
         dtype = cls._dtype_str_map[st_slice.get_dtype()]
         shape: tuple[int, ...] = tuple(st_slice.get_shape())
-        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
+        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[...] if len(s.get_shape()) == 0 else s[:])
         return cast(torch.Tensor, lazy)
 
     @classmethod
@@ -9215,6 +9381,13 @@ def parse_args() -> argparse.Namespace:
         )
     )
 
+    parser.add_argument(
+        "--sentence-transformers-dense-modules", action="store_true",
+        help=("Whether to include sentence-transformers dense modules."
+              "It can be used for sentence-transformers models, like google/embeddinggemma-300m"
+              "Default these modules are not included.")
+    )
+
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
         parser.error("the following arguments are required: model")
@@ -9277,9 +9450,13 @@ def main() -> None:
     if args.remote:
         hf_repo_id = args.model
         from huggingface_hub import snapshot_download
+        allowed_patterns = ["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"]
+        if args.sentence_transformers_dense_modules:
+            # include sentence-transformers dense modules safetensors files
+            allowed_patterns.append("*.safetensors")
         local_dir = snapshot_download(
             repo_id=hf_repo_id,
-            allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
+            allow_patterns=allowed_patterns)
         dir_model = Path(local_dir)
         logger.info(f"Downloaded config and tokenizer to {local_dir}")
     else:
@@ -9347,7 +9524,8 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template
+                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template,
+                                     sentence_transformers_dense_modules=args.sentence_transformers_dense_modules
                                      )
 
         if args.vocab_only:

convert_hf_to_gguf_update.py

@@ -140,6 +140,7 @@ models = [
     {"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": "llada-moe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
+    {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/backend/SYCL.md

@@ -145,12 +145,13 @@ 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.
 
@@ -160,7 +161,7 @@ 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 "$(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
+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
 ```
 
 *Notes:*
@@ -215,9 +216,19 @@ 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**
 
-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.
+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.
 
 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)*.
 
@@ -225,6 +236,12 @@ 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.
@@ -255,10 +272,11 @@ 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:
 
 ```
-[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]
+[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]
 ```
 
 - **Nvidia GPU**
@@ -353,7 +371,7 @@ 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/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).
+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).
 
 ##### Check device
 
@@ -466,7 +484,17 @@ If you already have a recent version of Microsoft Visual Studio, you can skip th
 
 3. Install Intel® oneAPI Base toolkit
 
-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.
+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.
 
 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/ops.md

@@ -31,7 +31,7 @@ Legend:
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -51,7 +51,7 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
@@ -65,11 +65,11 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -92,9 +92,9 @@ Legend:
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -102,9 +102,11 @@ Legend:
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
+|                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

examples/model-conversion/Makefile

@@ -116,20 +116,39 @@ 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 $(PROMPTS_FILE),--prompts-file "$(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
 
 embedding-run-converted-model:
 	@./scripts/embedding/run-converted-model.sh $(CONVERTED_EMBEDDING_MODEL) \
-	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
+	$(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
+	@./scripts/embedding/compare-embeddings-logits.sh \
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
+
 embedding-inspect-original-model:
 	$(call validate_embedding_model_path,embedding-inspect-original-model)
 	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}

examples/model-conversion/README.md

@@ -189,6 +189,23 @@ 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:
@@ -208,6 +225,13 @@ 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

@@ -1,4 +1,7 @@
 #include "llama.h"
+#include "common.h"
+
+
 #include <cstdio>
 #include <cstring>
 #include <string>
@@ -8,7 +11,10 @@
 
 static void print_usage(int, char ** argv) {
     printf("\nexample usage:\n");
-    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
+    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");
 }
 
@@ -17,6 +23,8 @@ 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;
@@ -41,9 +49,13 @@ 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 {
-                        embedding_mode = true;
+                        embd_norm = std::stoi(argv[++i]);
                     } catch (...) {
                         print_usage(argc, argv);
                         return 1;
@@ -112,7 +124,7 @@ int main(int argc, char ** argv) {
     ctx_params.no_perf = false;
     if (embedding_mode) {
         ctx_params.embeddings = true;
-        ctx_params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+        ctx_params.pooling_type = pooling_enabled ? LLAMA_POOLING_TYPE_MEAN : LLAMA_POOLING_TYPE_NONE;
         ctx_params.n_ubatch = ctx_params.n_batch;
     }
 
@@ -143,17 +155,27 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    float * logits;
-    int n_logits;
+    float * data_ptr;
+    int data_size;
     const char * type;
+    std::vector<float> embd_out;
 
     if (embedding_mode) {
-        logits = llama_get_embeddings(ctx);
-        n_logits = llama_model_n_embd(model) * batch.n_tokens;
+        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;
         type = "-embeddings";
 
-        const int n_embd = llama_model_n_embd(model);
-        const int n_embd_count = batch.n_tokens;
+        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);
+        }
 
         printf("Embedding dimension: %d\n", n_embd);
         printf("\n");
@@ -164,7 +186,7 @@ int main(int argc, char ** argv) {
 
             // Print first 3 values
             for (int i = 0; i < 3 && i < n_embd; i++) {
-                printf("%9.6f ", logits[j * n_embd + i]);
+                printf("%9.6f ", embeddings[j * n_embd + i]);
             }
 
             printf(" ... ");
@@ -172,7 +194,7 @@ 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 ", logits[j * n_embd + i]);
+                    printf("%9.6f ", embeddings[j * n_embd + i]);
                 }
             }
 
@@ -180,27 +202,33 @@ int main(int argc, char ** argv) {
         }
         printf("\n");
 
-        printf("Embeddings size: %d\n", n_logits);
+        printf("Embeddings size: %d\n", n_embeddings);
+
+        data_ptr = embeddings;
+        data_size = n_embeddings;
     } else {
-        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
-        n_logits = llama_vocab_n_tokens(vocab);
+        float * logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        const int 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 logits to binary file
+    // Save data to binary file
     char bin_filename[512];
     snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
-    printf("Saving logits to %s\n", bin_filename);
+    printf("Saving data 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(logits, sizeof(float), n_logits, f);
+    fwrite(data_ptr, sizeof(float), data_size, f);
     fclose(f);
 
     // Also save as text for debugging
@@ -211,27 +239,27 @@ 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 < n_logits; i++) {
-        fprintf(f, "%d: %.6f\n", i, logits[i]);
+    for (int i = 0; i < data_size; i++) {
+        fprintf(f, "%d: %.6f\n", i, data_ptr[i]);
     }
     fclose(f);
 
     if (!embedding_mode) {
         printf("First 10 logits: ");
-        for (int i = 0; i < 10 && i < n_logits; i++) {
-            printf("%.6f ", logits[i]);
+        for (int i = 0; i < 10 && i < data_size; i++) {
+            printf("%.6f ", data_ptr[i]);
         }
         printf("\n");
 
         printf("Last 10 logits: ");
-        for (int i = n_logits - 10; i < n_logits; i++) {
-            if (i >= 0) printf("%.6f ", logits[i]);
+        for (int i = data_size - 10; i < data_size; i++) {
+            if (i >= 0) printf("%.6f ", data_ptr[i]);
         }
         printf("\n\n");
     }
 
-    printf("Logits saved to %s\n", bin_filename);
-    printf("Logits saved to %s\n", txt_filename);
+    printf("Data saved to %s\n", bin_filename);
+    printf("Data saved to %s\n", txt_filename);
 
     llama_free(ctx);
     llama_model_free(model);

examples/model-conversion/requirements.txt

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

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

@@ -2,6 +2,21 @@
 
 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}"
@@ -15,7 +30,8 @@ echo "Converted model path:: ${CONVERTED_MODEL}"
 python ../../convert_hf_to_gguf.py --verbose \
     ${EMBEDDING_MODEL_PATH} \
     --outfile ${CONVERTED_MODEL} \
-    --outtype ${TYPE}
+    --outtype ${TYPE} \
+    ${SENTENCE_TRANSFORMERS}
 
 echo ""
 echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"

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

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

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

@@ -14,6 +14,8 @@ 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):
@@ -31,41 +33,52 @@ 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")
 
-tokenizer = AutoTokenizer.from_pretrained(model_path)
+# 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')
 
-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)
+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
 else:
-    model = AutoModel.from_pretrained(model_path, config=config)
-print(f"Model class: {type(model)}")
-print(f"Model file: {type(model).__module__}")
+    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__}")
 
 # Verify the model is using the correct sliding window
-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")
+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")
 
 model_name = os.path.basename(model_path)
 
@@ -75,34 +88,56 @@ 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():
-    outputs = model(**encoded)
-    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+    if use_sentence_transformers:
+        embeddings = model.encode(texts, convert_to_numpy=True)
+        all_embeddings = embeddings  # Shape: [batch_size, hidden_size]
 
-    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
-    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [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}'")
 
-    print(f"Hidden states shape: {hidden_states.shape}")
-    print(f"All embeddings shape: {all_embeddings.shape}")
-    print(f"Embedding dimension: {all_embeddings.shape[1]}")
+        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 embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
-    n_embd = all_embeddings.shape[1]
-    n_embd_count = all_embeddings.shape[0]
+        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()  # Empty line to match C++ output
+        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()
 
     for j in range(n_embd_count):
         embedding = all_embeddings[j]
@@ -120,29 +155,23 @@ with torch.no_grad():
 
         print()  # New line
 
-    print()  # Final empty line to match C++ output
+    print()
 
     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:
-        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")
-
+        idx = 0
         for j in range(n_embd_count):
-            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)")
+            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)")
     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

@@ -35,7 +35,11 @@ def cosine_similarity(a, b=None):
 
 def load_embeddings_from_file(filename, n_tokens, n_embd):
     embeddings = np.fromfile(filename, dtype=np.float32)
-    return embeddings.reshape(n_tokens, n_embd)
+    # 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)
 
 def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
     np.set_printoptions(suppress=True, precision=6)
@@ -48,58 +52,83 @@ 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
 
-    # 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
+    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])
         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(f"   Pooled embedding: Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
 
-    # 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}")
+        # 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}")
 
-    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}")
+        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}")
 
-    # 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}")
+        # 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}")
 
-    # 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("   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}")
 
-    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}")
+        # 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}")
 
-    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))
-    }
+        # 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))
+        }
 
 def read_prompt_from_file(file_path):
     try:

ggml/CMakeLists.txt

@@ -209,7 +209,6 @@ 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)
@@ -223,6 +222,9 @@ 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)

ggml/include/ggml-backend.h

@@ -215,6 +215,8 @@ 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-rpc.h

@@ -7,26 +7,25 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    2
+#define RPC_PROTO_MAJOR_VERSION    3
 #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);
+GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device);
 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);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, uint32_t device);
 
-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_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total);
 
-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 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, 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_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_add_server(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml.h

@@ -237,6 +237,8 @@
 #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
@@ -574,6 +576,7 @@ extern "C" {
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
         GGML_UNARY_OP_GELU_ERF,
+        GGML_UNARY_OP_XIELU,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -1148,6 +1151,18 @@ extern "C" {
             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,
@@ -1615,6 +1630,13 @@ 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

@@ -145,6 +145,9 @@ 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()

ggml/src/ggml-alloc.c

@@ -392,12 +392,8 @@ 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) {
-    size_t max_size = 0;
-    for (int i = 0; i < alloc->n_chunks; i++) {
-        max_size += alloc->chunks[i]->max_size;
-    }
-    return max_size;
+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;
 }
 
 
@@ -417,10 +413,8 @@ static void ggml_vbuffer_free(struct vbuffer * buf) {
     free(buf);
 }
 
-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_chunk_size(struct vbuffer * buf, int chunk) {
+    return buf->chunks[chunk] ? ggml_backend_buffer_get_size(buf->chunks[chunk]) : 0;
 }
 
 static size_t ggml_vbuffer_size(struct vbuffer * buf) {
@@ -885,12 +879,20 @@ 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
-        if (new_size > cur_size || galloc->buffers[i] == NULL) {
+        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) {
 #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

@@ -209,9 +209,6 @@ 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-cann/aclnn_ops.cpp

@@ -146,9 +146,7 @@ void ggml_cann_op_unary_gated(
     unary_op(ctx, acl_src0, acl_dst);
     GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst, acl_src1);
 
-    ggml_cann_release_resources(ctx, acl_src0, acl_dst);
-    if(src1)
-        ggml_cann_release_resources(ctx, acl_src1);
+    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
 /**
@@ -894,14 +892,13 @@ static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
 }
 
 /**
- * @brief Get or expand a cached float32 tensor filled with a scalar value.
+ * @brief Get or expand a cached tensor filled with a scalar value.
  *
- * This function manages cached device memory for float32 tensors. If the current
+ * This function manages cached device memory for tensors. If the current
  * cache size is insufficient for the requested tensor shape, the old memory will
- * be released and new memory will be allocated. The allocated buffer is then
- * initialized either with zeros (when @p value == 0.0f) or with the given scalar
- * value using CANN operations. Finally, an aclTensor object is created from the
- * cached memory and returned.
+ * be released and new memory will be allocated. The allocated buffer is
+ * initialized  with the given scalar value using CANN operations.
+ * Finally, an aclTensor object is created from the cached memory and returned.
  *
  * @param ctx           The CANN backend context that manages device memory.
  * @param buffer        A pointer to the cached device buffer (will be allocated
@@ -910,17 +907,19 @@ static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
  *                      updated when the cache is expanded.
  * @param ne            The tensor shape array (number of elements in each dimension).
  * @param nb            The stride size for each dimension.
+ * @param dtype         Data type of cached tensor.
  * @param dims          The number of tensor dimensions.
  * @param value         The scalar value used to fill the tensor (supports zero
  *                      initialization via memset or arbitrary values via fill_scalar).
  * @return              An aclTensor pointer created from the cached buffer.
  */
-static aclTensor* get_f32_cache_acl_tensor(
+static aclTensor* get_cache_acl_tensor(
     ggml_backend_cann_context& ctx,
     void** buffer,
     int64_t &cache_element,
     int64_t* ne,
     size_t* nb,
+    ggml_type dtype,
     int64_t dims,
     float value) {
     // Calculate total number of elements
@@ -928,7 +927,7 @@ static aclTensor* get_f32_cache_acl_tensor(
     for (int i = 0; i < dims; i++) {
         n_element *= ne[i];
     }
-    size_t size = n_element * sizeof(float);
+    size_t size = n_element * ggml_type_size(dtype);
 
     // Allocate or expand cache if needed
     if (cache_element < n_element) {
@@ -941,19 +940,17 @@ static aclTensor* get_f32_cache_acl_tensor(
         cache_element = n_element;
 
         // Initialize cache
-        if (value == 0.0f) {
-            ACL_CHECK(aclrtMemsetAsync(*buffer, size, 0, size, ctx.stream()));
-        } else {
-            int64_t pool_ne[1] = { n_element };
-            size_t pool_nb[1] = { sizeof(float) };
-            aclTensor* acl_value = ggml_cann_create_tensor(
-                *buffer, ACL_FLOAT, sizeof(float), pool_ne, pool_nb, 1);
-            aclnn_fill_scalar(ctx, 1, acl_value);
-            ggml_cann_release_resources(ctx, acl_value);
-        }
+        int64_t pool_ne[1] = { n_element };
+        size_t pool_nb[1] = { ggml_type_size(dtype) };
+        aclTensor* acl_value = ggml_cann_create_tensor(
+            *buffer, ggml_cann_type_mapping(dtype), ggml_type_size(dtype),
+            pool_ne, pool_nb, 1);
+        aclnn_fill_scalar(ctx, value, acl_value);
+        ggml_cann_release_resources(ctx, acl_value);
     }
 
-    return ggml_cann_create_tensor(*buffer, ACL_FLOAT, sizeof(float), ne, nb, dims);
+    return ggml_cann_create_tensor(*buffer, ggml_cann_type_mapping(dtype),
+            ggml_type_size(dtype), ne, nb, dims);
 }
 
 void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
@@ -965,35 +962,39 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    // build gamma, one...
+    // build gamma.
     size_t acl_gamma_nb[GGML_MAX_DIMS];
-    acl_gamma_nb[0] = sizeof(float);
+    // gamma's type is the same with dst.
+    acl_gamma_nb[0] = ggml_type_size(dst->type);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
         acl_gamma_nb[i] = acl_gamma_nb[i - 1] * src->ne[i - 1];
     }
-    aclTensor* acl_gamma = get_f32_cache_acl_tensor(
+    aclTensor* acl_gamma = get_cache_acl_tensor(
         ctx,
         &ctx.rms_norm_one_tensor_cache.cache,
         ctx.rms_norm_one_tensor_cache.size,
         src->ne,
         acl_gamma_nb,
+        dst->type,
         1,        // dims
         1.0f      // value
     );
 
-    // build rstd, zero...
+    // build rstd.
     int64_t acl_rstd_ne[] = {src->ne[1], src->ne[2], src->ne[3]};
     size_t acl_rstd_nb[GGML_MAX_DIMS - 1];
+    // rstd will always be F32.
     acl_rstd_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
         acl_rstd_nb[i] = acl_rstd_nb[i - 1] * acl_rstd_ne[i - 1];
     }
-    aclTensor* acl_rstd = get_f32_cache_acl_tensor(
+    aclTensor* acl_rstd = get_cache_acl_tensor(
         ctx,
         &ctx.rms_norm_zero_tensor_cache.cache,
         ctx.rms_norm_zero_tensor_cache.size,
         acl_rstd_ne,
         acl_rstd_nb,
+        GGML_TYPE_F32,
         GGML_MAX_DIMS - 1,
         0.0f      // value
     );
@@ -1765,33 +1766,35 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // src
     ggml_tensor* src1 = dst->src[1];  // index
 
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+
     switch (src0->type) {
-        case GGML_TYPE_F32: {
-            aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb,
-                                dst->data, dst->ne, dst->nb,
-                                src1, dst->type);
-            break;
-        }
-        case GGML_TYPE_F16: {
-            aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
-            ggml_cann_pool_alloc src_buffer_allocator(
-                ctx.pool(), ggml_nelements(src0) * sizeof(float));
-            void* src_trans_buffer = src_buffer_allocator.get();
-            size_t src_trans_nb[GGML_MAX_DIMS];
-            src_trans_nb[0] = sizeof(float);
-            for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+        case GGML_TYPE_F16:
+        case GGML_TYPE_F32:
+            if(src0->type == dst->type) {
+                aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb,
+                                    dst->data, dst->ne, dst->nb,
+                                    src1, dst->type);
+            } else {
+                aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+                ggml_cann_pool_alloc src_buffer_allocator(
+                    ctx.pool(), ggml_nelements(src0) * ggml_element_size(dst));
+                void* src_trans_buffer = src_buffer_allocator.get();
+                size_t src_trans_nb[GGML_MAX_DIMS];
+                src_trans_nb[0] = dst->nb[0];
+                for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+                }
+                aclTensor* src_trans_tensor = ggml_cann_create_tensor(
+                    src_trans_buffer, ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                    src0->ne, src_trans_nb, GGML_MAX_DIMS);
+                aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
+                aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
+                                    dst->data, dst->ne, dst->nb,
+                                    src1, dst->type);
+                ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
             }
-            aclTensor* src_trans_tensor = ggml_cann_create_tensor(
-                src_trans_buffer, ACL_FLOAT, ggml_type_size(dst->type),
-                src0->ne, src_trans_nb, GGML_MAX_DIMS);
-            aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
-            aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
-                                dst->data, dst->ne, dst->nb,
-                                src1, dst->type);
-            ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
             break;
-        }
         case GGML_TYPE_Q8_0: {
             // add 1 dim for bcast mul.
             size_t weight_nb[GGML_MAX_DIMS + 1], scale_nb[GGML_MAX_DIMS + 1],
@@ -1799,7 +1802,6 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
             int64_t weight_ne[GGML_MAX_DIMS + 1], scale_ne[GGML_MAX_DIMS + 1],
                 *dequant_ne;
             int64_t scale_offset = 0;
-
             // [3,4,5,64] -> [3,4,5,2,32]
             weight_ne[0] = QK8_0;
             weight_ne[1] = src0->ne[0] / QK8_0;
@@ -1809,7 +1811,6 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 weight_ne[i] = src0->ne[i - 1];
                 weight_nb[i] = weight_nb[i - 1] * weight_ne[i - 1];
             }
-
             // [3,4,5,64] -> [3,4,5,2,1]
             scale_ne[0] = 1;
             scale_ne[1] = src0->ne[0] / QK8_0;
@@ -1819,18 +1820,15 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 scale_ne[i] = src0->ne[i - 1];
                 scale_nb[i] = scale_nb[i - 1] * scale_ne[i - 1];
             }
-
             // [3,4,5,64] -> [3,4,5,2,32]
             dequant_ne = weight_ne;
-            dequant_nb[0] = sizeof(float);
+            dequant_nb[0] = ggml_type_size(dst->type);
             for (int i = 1; i < GGML_MAX_DIMS + 1; i++) {
                 dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1];
             }
-
             scale_offset = ggml_nelements(src0) * sizeof(int8_t);
             ggml_cann_pool_alloc dequant_buffer_allocator(
-                ctx.pool(), ggml_nelements(src0) * sizeof(float));
-
+                ctx.pool(), ggml_nelements(src0) * ggml_type_size(dst->type));
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb,
                 GGML_MAX_DIMS + 1);
@@ -1838,22 +1836,20 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 src0->data, ACL_FLOAT16, sizeof(uint16_t), scale_ne, scale_nb,
                 GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
             aclTensor* dequant_tensor = ggml_cann_create_tensor(
-                dequant_buffer_allocator.get(), ACL_FLOAT, sizeof(float),
+                dequant_buffer_allocator.get(), ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
                 dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
-
             aclnn_mul(ctx, acl_weight_tensor, acl_scale_tensor, dequant_tensor);
-            dequant_nb[0] = sizeof(float);
+            dequant_nb[0] = ggml_type_size(dst->type);
             dequant_ne = src0->ne;
             for (int i = 1; i < GGML_MAX_DIMS; i++) {
                 dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
             }
-
             aclnn_index_select_4d(ctx, dequant_buffer_allocator.get(),
                                    dequant_ne, dequant_nb,
                                    dst->data, dst->ne, dst->nb,
                                    src1, dst->type);
 
-            ggml_cann_release_resources(ctx, dequant_tensor);
+            ggml_cann_release_resources(ctx, acl_weight_tensor, acl_scale_tensor, dequant_tensor);
             break;
         }
         default:
@@ -1965,16 +1961,8 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     // Only check env once.
     static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or("on"));
     if (weight_to_nz && is_matmul_weight(weight)) {
-        int64_t acl_stride[2] = {1, transpose_ne[1]};
-
-        // Reverse ne.
-        std::reverse(transpose_ne, transpose_ne + n_dims);
-
-        std::vector<int64_t> storageDims = {transpose_ne[0], transpose_ne[1]};
-
-        acl_weight_tensor = aclCreateTensor(
-            transpose_ne, n_dims, ggml_cann_type_mapping(weight->type), acl_stride,
-            0, ACL_FORMAT_FRACTAL_NZ, storageDims.data(), 2, weight->data);
+        acl_weight_tensor =
+            ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_FRACTAL_NZ);
     } else {
         acl_weight_tensor =
             ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_ND);
@@ -3178,7 +3166,6 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         aclTensor* acl_src0_f16_tensor = nullptr;
         aclTensor* acl_src1_f16_tensor = nullptr;
         aclTensor* acl_src2_f16_tensor = nullptr;
-        aclTensor* acl_dst_f16_tensor  = nullptr;
 
         // Step 1: cast the src0 (Query) to fp16 if needed
         ggml_cann_pool_alloc src0_f16_allocator(ctx.pool());
@@ -3216,22 +3203,6 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         acl_src2_f16_tensor = ggml_cann_create_tensor(src2, src2_bsnd_ne,
             src2_bsnd_nb, GGML_MAX_DIMS);
 
-        ggml_cann_pool_alloc out_f16_allocator(ctx.pool());
-        void* out_f16_buffer = out_f16_allocator.alloc(
-                                    ggml_nelements(dst) * faElemSize);
-
-        int64_t* out_f16_ne = src0_bsnd_ne;
-        size_t out_f16_nb[GGML_MAX_DIMS];
-        out_f16_nb[0] = faElemSize;
-        for(int i = 1; i < GGML_MAX_DIMS; ++i){
-            out_f16_nb[i] = out_f16_nb[i - 1] * out_f16_ne[i - 1];
-        }
-
-        acl_dst_f16_tensor = ggml_cann_create_tensor(
-            out_f16_buffer, faDataType, faElemSize,
-            out_f16_ne, out_f16_nb, GGML_MAX_DIMS
-        );
-
         // Step 3: create the PSEShift tensor if needed
         //         this tensor is considered as mask (f16) in the llama.cpp
         aclTensor* bcast_pse_tensor = nullptr;
@@ -3317,8 +3288,8 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         aclTensor* acl_q_tensor = acl_src0_f16_tensor;
         aclTensor* acl_k_tensors[] = {acl_src1_f16_tensor};
         aclTensor* acl_v_tensors[] = {acl_src2_f16_tensor};
-        auto acl_k_tensor_list = aclCreateTensorList(acl_k_tensors, kvTensorNum);
-        auto acl_v_tensor_list = aclCreateTensorList(acl_v_tensors, kvTensorNum);
+        aclTensorList* acl_k_tensor_list = aclCreateTensorList(acl_k_tensors, kvTensorNum);
+        aclTensorList* acl_v_tensor_list = aclCreateTensorList(acl_v_tensors, kvTensorNum);
 
         int64_t numHeads = src0->ne[2]; // N
         int64_t numKeyValueHeads = src1->ne[2];
@@ -3334,8 +3305,29 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         int64_t keyAntiquantMode = 0;
         int64_t valueAntiquantMode = 0;
 
-        // Step 5: launch the FusedInferAttentionScoreV2 kernel.
-        // Refer to https://gitee.com/ascend/cann-ops-adv/blob/master/docs/FusedInferAttentionScoreV2.md
+        GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+        aclTensor * fa_dst_tensor = nullptr;
+        aclTensor * acl_dst_tensor = nullptr;
+        ggml_cann_pool_alloc out_f16_allocator(ctx.pool());
+        if (dst->type == GGML_TYPE_F32) {
+            void* out_f16_buffer = out_f16_allocator.alloc(
+                                        ggml_nelements(dst) * faElemSize);
+
+            int64_t* out_f16_ne = src0_bsnd_ne;
+            size_t out_f16_nb[GGML_MAX_DIMS];
+            out_f16_nb[0] = faElemSize;
+            for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                out_f16_nb[i] = out_f16_nb[i - 1] * out_f16_ne[i - 1];
+            }
+
+            fa_dst_tensor = ggml_cann_create_tensor(
+                out_f16_buffer, faDataType, faElemSize,
+                out_f16_ne, out_f16_nb, GGML_MAX_DIMS
+            );
+        }
+        else {
+            fa_dst_tensor = ggml_cann_create_tensor(dst);
+        }
 
         GGML_CANN_CALL_ACLNN_OP(ctx, FusedInferAttentionScoreV2,
             acl_q_tensor, acl_k_tensor_list, acl_v_tensor_list, // q, k, v
@@ -3357,23 +3349,24 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
             blockSize, antiquantMode, // blockSize, antiquantMode
             softmaxLseFlag, // softmaxLseFlag
             keyAntiquantMode, valueAntiquantMode, // keyAntiqMode, valueAntiqMode
-            acl_dst_f16_tensor, // attentionOut
+            fa_dst_tensor, // attentionOut
             nullptr // softmaxLse
         );
 
-        // Step 6: post-processing, permute and cast to f32
-        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-        // TODO: when dst is fp16, don't need cast
-        aclnn_cast(ctx, acl_dst_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
-        ggml_cann_release_resources(ctx, acl_src0_f16_tensor,
-                                         acl_src1_f16_tensor,
-                                         acl_src2_f16_tensor,
-                                         acl_dst_f16_tensor,
-                                         acl_dst_tensor);
-        if(src3 != nullptr){
-            ggml_cann_release_resources(ctx, bcast_pse_tensor);
+        if (dst->type == GGML_TYPE_F32) {
+            // Step 6: post-processing, permute and cast to f32
+            aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+            aclnn_cast(ctx, fa_dst_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
         }
-    }else{
+
+        ggml_cann_release_resources(ctx, acl_src0_f16_tensor,
+                                        acl_k_tensor_list,
+                                        acl_v_tensor_list,
+                                        fa_dst_tensor,
+                                        acl_dst_tensor,
+                                        bcast_pse_tensor);
+
+    } else {
         GGML_ABORT("Function is not implemented.");
     }
 }

ggml/src/ggml-cann/common.h

@@ -341,11 +341,18 @@ private:
 
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
+    // dst tensor
     void * node_address;
-    ggml_op node_op;
     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
+    ggml_op node_op;
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2186,7 +2186,15 @@ static void add_lru_matched_graph_node_properties(
         std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
 
         for (int src = 0; src < GGML_MAX_SRC; ++src) {
-            prop.src_address[src] = node->src[src] ? node->src[src]->data : nullptr;
+            if (node->src[src]) {
+                prop.src_address[src] = node->src[src]->data;
+                std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
+                std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
+            } else {
+                prop.src_address[src] = nullptr;
+                std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
+                std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
+            }
         }
 
         memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
@@ -2206,14 +2214,18 @@ static void add_lru_matched_graph_node_properties(
  * @param graph_node_properties The stored properties of a CANN graph node.
  * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
  */
-static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+static bool ggml_graph_node_has_matching_properties(
+        ggml_tensor * node,
+        ggml_graph_node_properties * graph_node_properties) {
     if (node->data != graph_node_properties->node_address &&
-           node->op != GGML_OP_VIEW) {
+            node->op != GGML_OP_VIEW) {
         return false;
     }
+
     if (node->op != graph_node_properties->node_op) {
         return false;
     }
+
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (node->ne[i] != graph_node_properties->ne[i]) {
             return false;
@@ -2222,17 +2234,31 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
             return false;
         }
     }
+
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (node->src[i] &&
-            node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_VIEW
-        ) {
-            return false;
+        if (node->src[i]) {
+            if (node->src[i]->data != graph_node_properties->src_address[i] &&
+                node->op != GGML_OP_VIEW) {
+                return false;
+            }
+
+            for (int d = 0; d < GGML_MAX_DIMS; d++) {
+                if (node->src[i]->ne[d] != graph_node_properties->src_ne[i][d]) {
+                    return false;
+                }
+                if (node->src[i]->nb[d] != graph_node_properties->src_nb[i][d]) {
+                    return false;
+                }
+            }
+        } else {
+            if (graph_node_properties->src_address[i] != nullptr) {
+                return false;
+            }
         }
     }
-    if (node->op == GGML_OP_SCALE &&
-        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
-        return false;
+
+    if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
+        return memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
     }
     return true;
 }

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

@@ -149,6 +149,7 @@ 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)
+#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
 #include <sys/prctl.h>
 #endif
 

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

@@ -689,8 +689,13 @@ bool ggml_is_numa(void) {
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__linux__)
     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
 }
 
@@ -2187,6 +2192,7 @@ 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;

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

@@ -29,6 +29,108 @@
 
 #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;
@@ -122,17 +224,18 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
+
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* SME GEMV */
         /* .kern_info = */ {
@@ -142,23 +245,24 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* .rhs_info = */ {
-            /* .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,
+            /* .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>,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -174,17 +278,17 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* SME GEMV */
         /* .kern_info = */ {
@@ -194,23 +298,24 @@ 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,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset_ex     = */ nullptr,
+            /* .get_rhs_packed_offset_ex = */ nullptr,
+            /* .run_kernel_ex         = */ nullptr,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* .rhs_info = */ {
-            /* .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,
+            /* .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>,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -229,17 +334,17 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* DOTPROD GEMV */
         /* .kern_info = */ {
@@ -249,23 +354,24 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* .rhs_info = */ {
-            /* .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,
+            /* .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>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -283,17 +389,17 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* i8mm GEMV */
         /* .kern_info = */ {
@@ -303,23 +409,24 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* .rhs_info = */ {
-            /* .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,
+            /* .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>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -338,17 +445,17 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* i8mm GEMV */
         /* .kern_info = */ {
@@ -358,23 +465,24 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* .rhs_info = */ {
-            /* .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,
+            /* .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>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -392,17 +500,17 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* DOTPROD GEMV */
         /* .kern_info = */ {
@@ -412,23 +520,24 @@ 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,
-            /* .run_kernel            = */ kai_run_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>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_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,
+            /* .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>,
         },
         /* .rhs_info = */ {
-            /* .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,
+            /* .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>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -443,6 +552,7 @@ 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 &&
@@ -452,6 +562,7 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
                 break;
             }
         }
+#endif
     }
 
     return kernel;
@@ -460,12 +571,14 @@ 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

@@ -4,8 +4,6 @@
 
 #pragma once
 
-#include <functional>
-#include <variant>
 #include "ggml.h"
 
 enum cpu_feature {
@@ -15,6 +13,7 @@ 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;
@@ -30,63 +29,52 @@ 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);
-    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;
+
+    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);
 };
 
 struct lhs_packing_info {
     size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
-    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;
+
+    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);
 };
 
 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);
-    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);
+
+    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);
 };
 
 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

@@ -8,6 +8,7 @@
 #include <stdexcept>
 #include <stdint.h>
 #include <string.h>
+#include <string>
 #if defined(__linux__)
 #include <asm/hwcap.h>
 #include <sys/auxv.h>
@@ -87,40 +88,6 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
     return tensor->ne[dim];
 }
 
-template <typename Variant, typename Ret, typename... Args, std::size_t... Is>
-constexpr bool variant_any_invocable_impl(std::index_sequence<Is...>) {
-    using V = std::remove_reference_t<Variant>;
-    return (std::is_invocable_r_v<
-                Ret,
-                std::variant_alternative_t<Is, V>,
-                Args...> || ...);
-}
-
-template <typename Variant, typename Ret, typename... Args>
-constexpr bool variant_any_invocable_v =
-    variant_any_invocable_impl<Variant, Ret, Args...>(
-        std::make_index_sequence<
-            std::variant_size_v<std::remove_reference_t<Variant>>>{});
-
-template<typename Ret, typename Variant, typename... Args>
-static inline Ret variant_call(Variant && var, Args&&... args) {
-    static_assert(variant_any_invocable_v<std::remove_reference_t<Variant>, Ret, Args...>,
-                  "No alternative in Variant is invocable with the provided arguments and return type.");
-
-    return std::visit(
-        [&](auto && f) -> Ret {
-            using F = std::decay_t<decltype(f)>;
-            if constexpr (std::is_invocable_r_v<Ret, F, Args...>) {
-                return std::invoke(std::forward<decltype(f)>(f), std::forward<Args>(args)...);
-            } else {
-                GGML_ABORT("Invalid function type in variant_call");
-                GGML_UNREACHABLE();
-            }
-        },
-        std::forward<Variant>(var)
-    );
-}
-
 namespace ggml::cpu::kleidiai {
 
 static size_t round_down(size_t x, size_t y) {
@@ -145,7 +112,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             return false;
         }
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
-        GGML_ASSERT(kernels);
+        if (!kernels) {
+            return false;
+        }
         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;
@@ -159,16 +128,18 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         size_t sr = kernel->get_sr();
 
         if (kernels->rhs_type == GGML_TYPE_Q4_0) {
-            size = variant_call<size_t>(lhs_info->packed_size, m, k, QK4_0, mr, kr, sr);
+            if (!lhs_info->packed_size_ex) return false;
+            size = lhs_info->packed_size_ex(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 = variant_call<size_t>(lhs_info->packed_size, m * r, k, mr, kr, sr) +
-                   variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
+            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) +
                    k * n * sizeof(float) + n * sizeof(float);
         } else {
-            GGML_ASSERT(false);
+            return false;
         }
 
         return true;
@@ -196,12 +167,18 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_TENSOR_BINARY_OP_LOCALS
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
-        GGML_ASSERT(kernels);
+        if (!kernels) {
+            return false;
+        }
 
         const 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;
@@ -228,10 +205,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const int64_t kr = (int64_t) kernel->get_kr();
         const int64_t sr = (int64_t) kernel->get_sr();
 
-        const size_t lhs_packed_size = variant_call<size_t>(lhs_info->packed_size, (size_t)m, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
-        const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, (size_t)n, (size_t)k);
-        const size_t kxn_size        = (size_t)k * (size_t)n * sizeof(float);
-        const size_t bias_size       = (size_t)n * sizeof(float);
+        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 kxn_size        = k * n * sizeof(float);
+        const size_t bias_size       = n * sizeof(float);
 
         const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
         GGML_ASSERT(wsize_required <= params->wsize);
@@ -259,10 +236,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                     const int64_t m_count = (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 = variant_call<size_t>(
-                        lhs_info->get_packed_offset, (size_t)m_start, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
-                    const size_t next_block_off = variant_call<size_t>(
-                        lhs_info->get_packed_offset, (size_t)(m_start + mr), (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
+                    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;
 
                     int64_t remaining = m_count;
@@ -278,9 +253,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                         const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
                         void * dst_ptr       = lhs_packed + dst_off;
 
-                        variant_call<void>(lhs_info->pack_func,
-                                        (size_t)take, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr,
-                                        /*m_idx_start*/ 0, src_ptr, lhs_stride, dst_ptr);
+                        lhs_info->pack_func_ex(take, k, 0, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
 
                         cur       += take;
                         remaining -= take;
@@ -296,10 +269,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                                         reinterpret_cast<const uint16_t *>(rhs_batch_base),
                                         rhs_stride);
 
-                variant_call<void>(kernels->rhs_info.pack_func,
-                                   /*num_groups*/ 1, (size_t)n, (size_t)k, (size_t)nr, (size_t)kr, (size_t)sr,
-                                   /*rhs_stride (bytes)*/ (size_t)(n * sizeof(float)),
-                                   rhs_kxn, bias, nullptr, rhs_packed, /*extra_bytes*/ 0, /*params*/ nullptr);
+                kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, 0, n * sizeof(float),
+                             rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
             }
 
             ggml_barrier(params->threadpool);
@@ -320,20 +291,15 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                     const int64_t n_to_process = (ith == num_threads_n - 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 = variant_call<size_t>(
-                        lhs_info->get_packed_offset, (size_t)0, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
-                    const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, (size_t)n_start, (size_t)k);
-                    const size_t dst_offset        = kernel->get_dst_offset((size_t)0, (size_t)n_start, dst_stride);
+                    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 void * lhs_ptr = lhs_packed + lhs_packed_offset0;
                     const void * rhs_ptr = rhs_packed + rhs_packed_offset;
                     float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);
 
-                    variant_call<void>(kernel->run_kernel,
-                                       (size_t)m, (size_t)n_to_process, (size_t)k,
-                                       lhs_ptr, rhs_ptr,
-                                       dst_ptr, dst_stride, sizeof(float),
-                                       -FLT_MAX, FLT_MAX);
+                    kernel->run_kernel_ex(m, n_to_process, k, 0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
                 }
             }
 
@@ -354,13 +320,19 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_TENSOR_BINARY_OP_LOCALS
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
-        GGML_ASSERT(kernels);
+        if (!kernels) {
+            return false;
+        }
 
         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;
@@ -402,25 +374,26 @@ 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 = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, QK4_0, mr, kr, sr);
+            const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(m_start, k, QK4_0, mr, kr, sr);
             void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
 
-            variant_call<void>(lhs_info->pack_func, m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+            // 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);
         }
 
         ggml_barrier(params->threadpool);
 
         // Perform the operation
         const size_t dst_stride        = dst->nb[1];
-        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 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 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) {
-            variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+            kernel->run_kernel_ex(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
                                sizeof(float), -FLT_MAX, FLT_MAX);
         }
 
@@ -429,7 +402,9 @@ 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);
-        GGML_ASSERT(ctx.kernels);
+        if (!ctx.kernels) {
+            return false;
+        }
 
         const ggml_tensor * src0 = dst->src[0];
         const ggml_tensor * src1 = dst->src[1];
@@ -438,6 +413,9 @@ 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);
@@ -480,7 +458,7 @@ public:
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
-        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);
+        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);
 
         return 0;
         GGML_UNUSED(data_size);
@@ -548,7 +526,7 @@ 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 variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
+    return ctx.kernels->rhs_info.packed_size_ex(n, k, nr, kr, QK4_0);
 
     GGML_UNUSED(buft);
 }

ggml/src/ggml-cpu/ops.cpp

@@ -3467,31 +3467,27 @@ 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);
 
-                ggml_float sum = 0.0;
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)x[i00];
-                }
-
+                float sum = 0.0;
+                ggml_vec_sum_f32(ne00, &sum, x);
                 float mean = sum/ne00;
 
                 float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
+                float variance = 0;
 
-                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);
-                }
+#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
 
-                float variance = sum2/ne00;
                 const float scale = 1.0f/sqrtf(variance + eps);
-
                 ggml_vec_scale_f32(ne00, y, scale);
             }
         }
@@ -8135,7 +8131,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         }
 
         // V /= S
-        const float S_inv = 1.0f/S;
+        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
         ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
         // dst indices
@@ -8637,7 +8633,7 @@ 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 = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+                    const float dt_soft_plus = ggml_softplus(dt[h]);
                     const float dA = expf(dt_soft_plus * A[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
@@ -8734,7 +8730,7 @@ 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 = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+                    const float dt_soft_plus = ggml_softplus(dt[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
                     // dim
@@ -8997,6 +8993,10 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_exp(params, dst);
             } break;
+        case GGML_UNARY_OP_XIELU:
+            {
+                ggml_compute_forward_xielu(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");

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

@@ -52,6 +52,15 @@ 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);
 }
@@ -121,6 +130,86 @@ 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);
 }
@@ -184,3 +273,17 @@ 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_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

@@ -22,6 +22,7 @@ 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_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-cpu/vec.cpp

@@ -404,6 +404,72 @@ 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

@@ -44,6 +44,7 @@ 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);
 
@@ -143,14 +144,14 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         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 elemnst
+            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 ekements
+            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);
@@ -159,7 +160,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 
             ax3 = GGML_F16x_VEC_LOAD(x[0] + i + 2*ggml_f16_epr, 2);
             sum_02 = GGML_F16x_VEC_FMA(sum_02, ax3, ay3);
-            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
+            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);
@@ -654,11 +655,11 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
         }
         // leftovers
         // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
-        if (np < n) {
-            svbool_t pg = svwhilelt_b32(np, n);
-            ay1 = svld1_f32(pg, y + np);
+        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 + np, ay1);
+            svst1_f32(pg, y + i, ay1);
         }
     #elif defined(__riscv_v_intrinsic)
         for (int i = 0, avl; i < n; i += avl) {
@@ -819,7 +820,8 @@ 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) {
-        y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i])));
+        const float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : expm1f(v));
     }
 }
 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

@@ -44,6 +44,8 @@ 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/common.cuh

@@ -220,14 +220,6 @@ 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)
@@ -253,7 +245,8 @@ static bool fp16_available(const int cc) {
 }
 
 static bool fast_fp16_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
+    return GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -262,27 +255,6 @@ 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) ||
@@ -600,6 +572,10 @@ 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) {
@@ -968,13 +944,6 @@ 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
 };
 

ggml/src/ggml-cuda/cpy.cu

@@ -8,18 +8,16 @@
 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_direct, const int ne,
+static __global__ void cpy_flt(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, char ** cdst_indirect, int graph_cpynode_index) {
+                               const int nb12, const int nb13) {
     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);
@@ -63,18 +61,16 @@ 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_direct, const int ne,
+static __global__ void cpy_f32_q(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, char ** cdst_indirect, int graph_cpynode_index) {
+                                 const int nb12, const int nb13) {
     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;
@@ -91,18 +87,16 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_q_f32(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, char ** cdst_indirect, int graph_cpynode_index) {
+                                 const int nb12, const int nb13) {
     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;
@@ -118,67 +112,47 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
-// Copy destination pointers to GPU to be available when pointer indirection is in use
-
-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;
-    }
-    // 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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 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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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) {
 
     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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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 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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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) {
 
     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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q4_0_f32_cuda(
@@ -187,22 +161,22 @@ 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     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, cdst_indirect, graph_cpynode_index++);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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) {
 
     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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q4_1_f32_cuda(
@@ -211,22 +185,22 @@ 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     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, cdst_indirect, graph_cpynode_index++);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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) {
 
     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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q5_0_f32_cuda(
@@ -235,22 +209,22 @@ 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     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, cdst_indirect, graph_cpynode_index++);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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) {
 
     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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q5_1_f32_cuda(
@@ -259,25 +233,25 @@ 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     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, cdst_indirect, graph_cpynode_index++);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 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, char ** cdst_indirect, int & graph_cpynode_index) {
+    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) {
 
     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, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -311,16 +285,6 @@ 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;
 
-    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)
@@ -329,134 +293,62 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         } else
 #endif // GGML_USE_MUSA && GGML_MUSA_MUDNN_COPY
         {
-            if (src0->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, dest_ptrs_d, graph_cpynode_index);
-            } else {
-                CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
-            }
+            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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        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);
+        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);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        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);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        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);
+        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);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        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);
+        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);
     } 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, dest_ptrs_d, graph_cpynode_index);
+        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);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        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);
+        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);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        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);
+        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);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        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);
+        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);
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        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);
+        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);
     } 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)) {
-        // Prioritize CUDA graph compatibility over direct memory copy optimization.
-        // Using copy kernels here maintains graph indirection support, preventing performance regression from disabled CUDA graphs.
-        if (src0->type == GGML_TYPE_F32) {
-            return (void*) cpy_flt<cpy_1_flt<float, float>>;
-        } else {
-            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));
-    }
+    ggml_cuda_cpy(ctx, src0, dst);
 }

ggml/src/ggml-cuda/cpy.cuh

@@ -2,10 +2,6 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
 
 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

@@ -793,8 +793,6 @@ 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();
@@ -878,7 +876,7 @@ 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 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+    if (mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
         const int s31 = mask->nb[1] / sizeof(half2);
         const int s33 = mask->nb[3] / sizeof(half2);
 
@@ -916,8 +914,7 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
-        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.
+        const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 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);
@@ -946,7 +943,7 @@ void launch_fattn(
 
         blocks_num.x = ntiles_x;
         blocks_num.y = parallel_blocks;
-        blocks_num.z = Q->ne[2]*Q->ne[3];
+        blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));

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

@@ -1,755 +1,45 @@
 #include "common.cuh"
-#include "fattn-common.cuh"
 #include "fattn-tile.cuh"
-
-// 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]) {
+#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;
         case  64: {
-            launch_fattn_tile_switch_ncols< 64, use_logit_softcap>(ctx, dst);
+            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);
         } break;
         case 128: {
-            launch_fattn_tile_switch_ncols<128, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<128, 128>(ctx, dst);
         } break;
         case 256: {
-            launch_fattn_tile_switch_ncols<256, use_logit_softcap>(ctx, dst);
+            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);
         } 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);
-    }
-}

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

@@ -535,8 +535,6 @@ void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_ten
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-
     if (Q->ne[1] == 1) {
         constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {

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

@@ -6,19 +6,19 @@
 #include "fattn-common.cuh"
 #include "fattn-wmma-f16.cuh"
 
-#ifdef FP16_MMA_AVAILABLE
+#ifdef GGML_USE_WMMA_FATTN
 #if !defined(GGML_USE_HIP)
 #include <mma.h>
-#ifdef GGML_USE_MUSA
+#if defined(GGML_USE_MUSA)
 namespace wmma = mtmusa::wmma;
 #else // GGML_USE_MUSA
 namespace wmma = nvcuda::wmma;
 #endif // GGML_USE_MUSA
-#elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
+#elif defined(GGML_USE_HIP)
 #include <rocwmma/rocwmma.hpp>
 namespace wmma = rocwmma;
 #endif // !defined(GGML_USE_HIP)
-#endif // FP16_MMA_AVAILABLE
+#endif // GGML_USE_WMMA_FATTN
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, typename KQ_acc_t, bool use_logit_softcap>
@@ -45,7 +45,7 @@ static __global__ void flash_attn_ext_f16(
                             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) {
-#if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
+#if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(D == 128 || D == 256)) {
         NO_DEVICE_CODE;
@@ -481,7 +481,7 @@ static __global__ void flash_attn_ext_f16(
               ne31, ne32, ne33,
               nb31, nb32, nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
+#endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
 }
 
 constexpr int get_max_power_of_2(int x) {

ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -1,3 +1,51 @@
+#pragma once
+
 #include "common.cuh"
 
+#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+#define GGML_USE_WMMA_FATTN
+#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+
+#if defined(GGML_HIP_ROCWMMA_FATTN)
+#if defined(CDNA) && (ROCWMMA_VERSION_MAJOR < 2 || ROCWMMA_VERSION_MINOR > 0 || ROCWMMA_VERSION_PATCH > 0)
+#define GGML_USE_WMMA_FATTN
+#elif defined(CDNA)
+#warning "rocwmma fattn on CDNA is broken on rocwmma v2.0.0, expect degraded performance"
+#endif // defined(CDNA) && (ROCWMMA_VERSION_MAJOR < 2 || ROCWMMA_VERSION_MINOR > 0 || ROCWMMA_VERSION_PATCH > 0)
+#if defined(RDNA3)
+#define GGML_USE_WMMA_FATTN
+#endif // defined(RDNA3)
+#if defined(RDNA4) && ROCWMMA_VERSION_MAJOR > 1
+#define GGML_USE_WMMA_FATTN
+#elif defined(RDNA4)
+#warning "rocwmma fattn is not suported on RDNA4 on rocwmma < v2.0.0, expect degraded performance"
+#endif // defined(RDNA4) && ROCWMMA_VERSION_MAJOR > 1
+#endif // defined(GGML_HIP_ROCWMMA_FATTN)
+
+// WMMA flash attention requires FP16 matrix instructions to be available for ggml code.
+static bool ggml_cuda_should_use_wmma_fattn(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_RDNA3(cc) || GGML_CUDA_CC_IS_MTHREADS(cc)) {
+        return true;
+    } else if (GGML_CUDA_CC_IS_CDNA(cc)){
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (ROCWMMA_VERSION_MAJOR < 2 || ROCWMMA_VERSION_MINOR > 0 || ROCWMMA_VERSION_PATCH > 0)
+        return true;
+#else
+        return false;
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) (ROCWMMA_VERSION_MAJOR < 2 || ROCWMMA_VERSION_MINOR > 0 || ROCWMMA_VERSION_PATCH > 0)
+    } else if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+#if defined(GGML_HIP_ROCWMMA_FATTN) && ROCWMMA_VERSION_MAJOR > 1
+        return true;
+#else
+        return false;
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && ROCWMMA_VERSION_MAJOR > 1
+    } else {
+        return false;
+    }
+#endif // defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
+}
+
 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/fattn.cu

@@ -198,6 +198,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     return BEST_FATTN_KERNEL_NONE;
 #endif// FLASH_ATTN_AVAILABLE
 
+    const ggml_tensor * KQV   = dst;
     const ggml_tensor * Q     = dst->src[0];
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
@@ -206,31 +207,32 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     const int gqa_ratio = Q->ne[2] / K->ne[2];
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
 
+    float max_bias = 0.0f;
+    memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+    // The effective batch size for the kernel can be increased by gqa_ratio.
+    // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
+    const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+
     const int cc = ggml_cuda_info().devices[device].cc;
 
     switch (K->ne[0]) {
+        case  40:
         case  64:
-        case 128:
-        case 256:
-            if (V->ne[0] != K->ne[0]) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
-            break;
         case  80:
         case  96:
+        case 128:
         case 112:
+        case 256:
             if (V->ne[0] != K->ne[0]) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!fp16_mma_available(cc) && !turing_mma_available(cc)) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
             break;
         case 576:
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!turing_mma_available(cc) || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;
@@ -264,47 +266,57 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_NONE;
     }
 
-    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0;
-
-    // If Turing tensor cores available, use them except for some cases with batch size 1:
-    if (turing_mma_available(cc)) {
-        best_fattn_kernel best = BEST_FATTN_KERNEL_MMA_F16;
+    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
+    // If Turing tensor cores available, use them:
+    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40) {
         if (can_use_vector_kernel) {
             if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
-                    best = BEST_FATTN_KERNEL_VEC;
+                    return BEST_FATTN_KERNEL_VEC;
                 }
             } else {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
                     if (Q->ne[1] <= 2) {
-                        best = BEST_FATTN_KERNEL_VEC;
+                        return BEST_FATTN_KERNEL_VEC;
                     }
                 } else {
                     if (Q->ne[1] == 1) {
-                        best = BEST_FATTN_KERNEL_VEC;
+                        return BEST_FATTN_KERNEL_VEC;
                     }
                 }
             }
-            if ((gqa_ratio % 2 != 0 || !mask) && Q->ne[1] == 1) {
-                best = BEST_FATTN_KERNEL_VEC; // GQA-specific optimizations in the mma kernel do not apply.
+            if (!gqa_opt_applies && Q->ne[1] == 1) {
+                return BEST_FATTN_KERNEL_VEC;
             }
         }
 
-        return best;
+        return BEST_FATTN_KERNEL_MMA_F16;
     }
 
-    // Use kernels specialized for small batch sizes if possible:
-    if (Q->ne[1] <= 8 && can_use_vector_kernel) {
-        return BEST_FATTN_KERNEL_VEC;
-    }
-
-    // For large batch sizes, use the WMMA kernel if possible:
-    if (fp16_mma_available(cc)) {
+    // Use the WMMA kernel if possible:
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 576) {
+        if (can_use_vector_kernel && Q->ne[1] <= 2) {
+            return BEST_FATTN_KERNEL_VEC;
+        }
         return BEST_FATTN_KERNEL_WMMA_F16;
     }
 
-    // If there is no suitable kernel for tensor cores or small batch sizes, use the generic kernel for large batch sizes:
+    // If there are no tensor cores available, use the generic tile kernel:
+    if (can_use_vector_kernel) {
+        if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+            if (Q->ne[1] == 1) {
+                if (!gqa_opt_applies) {
+                    return BEST_FATTN_KERNEL_VEC;
+                }
+            }
+        } else {
+            if (Q->ne[1] <= 2) {
+                return BEST_FATTN_KERNEL_VEC;
+            }
+        }
+    }
     return BEST_FATTN_KERNEL_TILE;
 }
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -231,7 +231,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
-        info.devices[id].integrated = prop.integrated;
+        info.devices[id].integrated = false; // Temporarily disabled due to issues with corrupted output (e.g. #15034)
         info.devices[id].nsm        = prop.multiProcessorCount;
         info.devices[id].smpb       = prop.sharedMemPerBlock;
         info.devices[id].warp_size  = prop.warpSize;
@@ -2334,6 +2334,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_ELU:
                     ggml_cuda_op_elu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_XIELU:
+                    ggml_cuda_op_xielu(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2630,11 +2633,10 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
 }
 
 #ifdef USE_CUDA_GRAPH
-static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
+static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
     bool use_cuda_graph) {
 
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
-    cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
     const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
     const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
@@ -2685,33 +2687,11 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_CPY) {
-
-            // Store the pointers which are updated for each token, such that these can be sent
-            // to the device and accessed using indirection from CUDA graph
-            cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data);
-
-            // store a pointer to each copy op CUDA kernel to identify it later
-            void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
-            if (!ptr) {
-                use_cuda_graph = false;
-#ifndef NDEBUG
-                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
-#endif
-            }
-        }
-
         if (!use_cuda_graph) {
             break;
         }
     }
 
-    if (use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = true;
-        // copy pointers to GPU so they can be accessed via indirection within CUDA graph
-        ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream());
-    }
-
     return use_cuda_graph;
 }
 
@@ -2730,7 +2710,6 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
 
 static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
     if (node->data != graph_node_properties->node_address &&
-          node->op != GGML_OP_CPY &&
           node->op != GGML_OP_VIEW) {
         return false;
     }
@@ -2751,7 +2730,6 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (node->src[i] &&
             node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_CPY &&
             node->op != GGML_OP_VIEW
         ) {
             return false;
@@ -3117,7 +3095,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     if (use_cuda_graph) {
         cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
 
-        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph);
+        use_cuda_graph = check_node_graph_compatibility(cgraph, use_cuda_graph);
 
         // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
         if (use_cuda_graph && cuda_graph_update_required) {
@@ -3144,10 +3122,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    if (!use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = false;
-    }
-
 #else
     bool use_cuda_graph = false;
     bool cuda_graph_update_required = false;
@@ -3864,7 +3838,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->device = i;
                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
 
-                ggml_cuda_set_device(i);
                 cudaDeviceProp prop;
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;

ggml/src/ggml-cuda/mmf.cu

@@ -1,5 +1,7 @@
 #include "ggml.h"
 #include "mmf.cuh"
+#include "mmid.cuh"
+
 
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
@@ -37,6 +39,12 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     const int64_t ids_s0 = ids ? ids->nb[0] / ggml_type_size(ids->type) : 0;
     const int64_t ids_s1 = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
 
+    mmf_ids_data ids_info{};
+    mmf_ids_data * ids_info_ptr = nullptr;
+    ggml_cuda_pool_alloc<int32_t> ids_src_compact_dev;
+    ggml_cuda_pool_alloc<int32_t> ids_dst_compact_dev;
+    ggml_cuda_pool_alloc<int32_t> expert_bounds_dev;
+
     // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
     const int64_t ncols_dst          = ids ? ne2  : ne1;
     const int64_t nchannels_dst      = ids ? ne1 : ne2;
@@ -54,6 +62,33 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
         nchannels_y      = ids->ne[0];
     }
 
+    if (ids && ncols_dst > 16) {
+        const int64_t n_expert_used = ids->ne[0];
+        const int64_t n_experts     = ne02;
+        const int64_t n_tokens      = ne12;
+        const int64_t ne_get_rows   = n_tokens * n_expert_used;
+
+        ids_src_compact_dev.alloc(ctx.pool(), ne_get_rows);
+        ids_dst_compact_dev.alloc(ctx.pool(), ne_get_rows);
+        expert_bounds_dev.alloc(ctx.pool(), n_experts + 1);
+
+        const int si1  = static_cast<int>(ids_s1);
+        const int sis1 = static_cast<int>(src1->nb[2] / src1->nb[1]);
+
+        GGML_ASSERT(sis1 > 0);
+
+        ggml_cuda_launch_mm_ids_helper(ids_d, ids_src_compact_dev.get(), ids_dst_compact_dev.get(), expert_bounds_dev.get(),
+            static_cast<int>(n_experts), static_cast<int>(n_tokens), static_cast<int>(n_expert_used), static_cast<int>(ne11), si1, sis1, ctx.stream());
+        CUDA_CHECK(cudaGetLastError());
+
+        ids_info.ids_src_compact   = ids_src_compact_dev.get();
+        ids_info.ids_dst_compact   = ids_dst_compact_dev.get();
+        ids_info.expert_bounds_dev = expert_bounds_dev.get();
+        ids_info.n_experts         = static_cast<int>(n_experts);
+        ids_info.sis1              = sis1;
+        ids_info_ptr = &ids_info;
+    }
+
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
@@ -61,7 +96,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_F16: {
             const half2 * src0_d = (const half2 *) src0->data;
@@ -69,7 +104,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data;
@@ -77,7 +112,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -98,10 +133,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
     }
 
     if (mul_mat_id) {
-        if (type == GGML_TYPE_F32 && src1_ncols > 32) {
+        if (src0_ne[1] <= 1024 && src1_ncols > 512) {
             return false;
-        }
-        if ((type == GGML_TYPE_F16 || type == GGML_TYPE_BF16) && src1_ncols > 64) {
+        } else if(src0_ne[1] > 1024 && src1_ncols > 128) {
             return false;
         }
     } else {

ggml/src/ggml-cuda/mmf.cuh

@@ -7,6 +7,14 @@ using namespace ggml_cuda_mma;
 
 #define MMF_ROWS_PER_BLOCK 32
 
+struct mmf_ids_data {
+    const int32_t * ids_src_compact = nullptr;
+    const int32_t * ids_dst_compact = nullptr;
+    const int32_t * expert_bounds_dev = nullptr;
+    int n_experts = 0;
+    int sis1 = 0;
+};
+
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
 bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols, bool mul_mat_id);
@@ -224,6 +232,250 @@ static __global__ void mul_mat_f(
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
+
+//This kernel is for larger batch sizes of mul_mat_id
+template <typename T, int rows_per_block, int cols_per_block, int nwarps>
+__launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
+static __global__ void mul_mat_f_ids(
+        const T * __restrict__ x, const float * __restrict__ y,
+        const int32_t * __restrict__ ids_src_compact, const int32_t * __restrict__ ids_dst_compact,
+        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst,
+        const int ncols, const int ncols_dst_total, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const uint3 sis1_fd, const uint3 nch_fd) {
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile< 8, 8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int ntA = rows_per_block / tile_A::I;
+    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
+
+    const int row0        = blockIdx.x * rows_per_block;
+
+    const int expert_idx = blockIdx.y;
+    const int expert_start = expert_bounds[expert_idx];
+    const int expert_end   = expert_bounds[expert_idx + 1];
+    const int ncols_expert = expert_end - expert_start;
+
+    const int tiles_for_expert = (ncols_expert + cols_per_block - 1) / cols_per_block;
+    const int tile_idx = blockIdx.z;
+    if (tile_idx >= tiles_for_expert) {
+        return;
+    }
+
+    const int col_base = tile_idx * cols_per_block;
+
+    GGML_UNUSED(channel_ratio);
+
+    const int channel_x   = expert_idx;
+    const int sample_dst  = 0;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x  + row0*stride_row;
+    y   += int64_t(sample_y)  *stride_sample_y;
+    dst += int64_t(sample_dst)*stride_sample_dst;
+
+    const int32_t * ids_src_expert = ids_src_compact + expert_start;
+    const int32_t * ids_dst_expert = ids_dst_compact + expert_start;
+
+    extern __shared__ char data_mmv[];
+    char * compute_base = data_mmv;
+
+    //const float2 * y2 = (const float2 *) y;
+
+    tile_C C[ntA][ntB];
+
+    T * tile_xy = (T *) compute_base + threadIdx.y*(tile_A::I * tile_k_padded);
+
+    for (int col = threadIdx.y*warp_size + threadIdx.x; col < ncols; col += nwarps*warp_size) {
+        tile_A A[ntA][warp_size / tile_A::J];
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int i = 0; i < tile_A::I; ++i) {
+                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
+                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
+            }
+        }
+
+        if constexpr (std::is_same_v<T, float>) {
+            float vals_buf[2][tile_B::I];
+            auto gather_tile = [&](int tile_idx_local, float *vals) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + tile_idx_local*tile_B::I;
+                    const int global_j = col_base + j;
+                    float val = 0.0f;
+                    if (j < cols_per_block && global_j < ncols_expert) {
+                        const int src_entry = ids_src_expert[global_j];
+                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
+                        const int token   = (int) qrm.x;
+                        const int channel = (int) qrm.y;
+                        if (token < ncols_dst_total) {
+                            val = y[channel*stride_channel_y + token*stride_col_y + col];
+                        }
+                    }
+                    vals[j0] = val;
+                }
+            };
+
+            gather_tile(0, vals_buf[0]);
+
+            int curr_buf = 0;
+            int next_buf = 1;
+#pragma unroll
+            for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = vals_buf[curr_buf][j0];
+                }
+
+                if (itB + 1 < ntB) {
+                    gather_tile(itB + 1, vals_buf[next_buf]);
+                }
+
+#pragma unroll
+                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                    tile_B B;
+                    load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                    for (int itA = 0; itA < ntA; ++itA) {
+                        mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                    }
+                }
+
+                if (itB + 1 < ntB) {
+                    curr_buf ^= 1;
+                    next_buf ^= 1;
+                }
+            }
+        } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+            float2 vals_buf[2][tile_B::I];
+            auto gather_tile = [&](int tile_idx_local, float2 *vals) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + tile_idx_local*tile_B::I;
+                    const int global_j = col_base + j;
+                    float2 tmp = make_float2(0.0f, 0.0f);
+                    if (j < cols_per_block && global_j < ncols_expert) {
+                        const int src_entry = ids_src_expert[global_j];
+                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
+                        const int token   = (int) qrm.x;
+                        const int channel = (int) qrm.y;
+                        if (token < ncols_dst_total) {
+                            tmp = *(const float2*) &y[channel*stride_channel_y + 2*(token*stride_col_y + col)];
+                        }
+                    }
+                    vals[j0] = tmp;
+                }
+            };
+
+            if (ntB > 0) {
+                gather_tile(0, vals_buf[0]);
+            }
+
+            int curr_buf = 0;
+            int next_buf = 1;
+#pragma unroll
+            for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const float2 tmp = vals_buf[curr_buf][j0];
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                }
+
+                if (itB + 1 < ntB) {
+                    gather_tile(itB + 1, vals_buf[next_buf]);
+                }
+
+#pragma unroll
+                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                    tile_B B;
+                    load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                    for (int itA = 0; itA < ntA; ++itA) {
+                        mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                    }
+                }
+
+                if (itB + 1 < ntB) {
+                    curr_buf ^= 1;
+                    next_buf ^= 1;
+                }
+            }
+        } else {
+            static_assert(std::is_same_v<T, void>, "unsupported type");
+        }
+    }
+
+    float * buf_iw = (float *) compute_base;
+    constexpr int kiw = nwarps*rows_per_block + 4;
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+#pragma unroll
+    for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int l = 0; l < tile_C::ne; ++l) {
+                const int i = threadIdx.y*rows_per_block + itA*tile_C::I + tile_C::get_i(l);
+                const int j = itB*tile_C::J + tile_C::get_j(l);
+                buf_iw[j*kiw + i] = C[itA][itB].x[l];
+            }
+        }
+    }
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j0 + nwarps > cols_per_block && j >= cols_per_block) {
+            return;
+        }
+
+        float sum = 0.0f;
+        static_assert(rows_per_block == warp_size, "need loop/check");
+#pragma unroll
+        for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
+            const int i = i0 + threadIdx.x;
+
+            sum += buf_iw[j*kiw + i];
+        }
+
+        const int global_j = col_base + j;
+        if (j < cols_per_block && global_j < ncols_expert && nchannels_dst > 0) {
+            const int dst_entry = ids_dst_expert[global_j];
+            const uint2 qrm = fast_div_modulo((uint32_t) dst_entry, nch_fd);
+            const int token = (int) qrm.x;
+            if (token < ncols_dst_total) {
+                const int slot = (int) qrm.y;
+                dst[slot*stride_channel_dst + token*stride_col_dst + row0 + threadIdx.x] = sum;
+            }
+        }
+    }
+#else
+    GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
+        ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd);
+    NO_DEVICE_CODE;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
 template<typename T, int cols_per_block, int nwarps>
 static inline void mul_mat_f_switch_ids(
         const T * x, const float * y, const int32_t * ids, float * dst,
@@ -232,13 +484,35 @@ static inline void mul_mat_f_switch_ids(
         const int64_t stride_col_id, const int64_t stride_row_id,
         const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
         const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream) {
-    if (ids) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream,
+        const mmf_ids_data * ids_data) {
+    const bool has_ids_data = ids_data && ids_data->ids_src_compact;
+
+    // Use the compact-ids kernel only for larger tiles; for small ncols_dst (< 16)
+    // we prefer the normal mul_mat_f path with has_ids=true.
+    if (has_ids_data && ncols_dst > 16) {
+        const int max_tiles = (int) ((ncols_dst + cols_per_block - 1) / cols_per_block);
+        if (max_tiles == 0) {
+            return;
+        }
+        dim3 block_nums_ids(block_nums.x, ids_data->n_experts, max_tiles);
+
+        const uint3 sis1_fd = ids_data->sis1 > 0 ? init_fastdiv_values((uint32_t) ids_data->sis1) : make_uint3(0, 0, 1);
+        const uint3 nch_fd  = init_fastdiv_values((uint32_t) nchannels_dst);
+
+        mul_mat_f_ids<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
+            (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->expert_bounds_dev, dst,
+            ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+            sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst,
+            sis1_fd, nch_fd);
+    } else if (ids) {
         const int64_t col_tiles = (ncols_dst + cols_per_block - 1) / cols_per_block;
         dim3 block_nums_ids = block_nums;
         block_nums_ids.y *= col_tiles;
+
         mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
-             (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
              stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
              sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     } else {
@@ -258,7 +532,7 @@ void mul_mat_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
     typedef tile<16, 8, T>     tile_A;
     typedef tile< 8, 8, T>     tile_B;
 
@@ -290,7 +564,7 @@ void mul_mat_f_cuda(
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
     const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;
-    const int64_t grid_y = ids ? nchannels_x : nchannels_dst; // per expert when ids present
+    const int64_t grid_y = ids ? nchannels_x : nchannels_dst;
 
     const dim3 block_nums(nrows_x/rows_per_block, grid_y, nsamples_dst);
     const dim3 block_dims(warp_size, nwarps_best, 1);
@@ -300,49 +574,57 @@ void mul_mat_f_cuda(
             mul_mat_f_switch_ids<T, cols_per_block, 1>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 2: {
             mul_mat_f_switch_ids<T, cols_per_block, 2>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 3: {
             mul_mat_f_switch_ids<T, cols_per_block, 3>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 4: {
             mul_mat_f_switch_ids<T, cols_per_block, 4>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 5: {
             mul_mat_f_switch_ids<T, cols_per_block, 5>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 6: {
             mul_mat_f_switch_ids<T, cols_per_block, 6>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 7: {
             mul_mat_f_switch_ids<T, cols_per_block, 7>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 8: {
             mul_mat_f_switch_ids<T, cols_per_block, 8>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -361,7 +643,7 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
 
     const int ncols_case = (ids && ncols_dst > 16) ? 16 : ncols_dst;
 
@@ -371,82 +653,82 @@ static void mul_mat_f_switch_cols_per_block(
         case  1: {
             mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  2: {
             mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  3: {
             mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  4: {
             mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  5: {
             mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream, ids_data);
         } break;
         case  6: {
             mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  7: {
             mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  8: {
             mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  9: {
             mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 10: {
             mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 11: {
             mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 12: {
             mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 13: {
             mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 14: {
             mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 15: {
             mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 16: {
             mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -462,7 +744,7 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, \
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,\
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, \
-        cudaStream_t stream);
+        cudaStream_t stream, const mmf_ids_data * ids_data);
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 #define DECL_MMF_CASE_EXTERN(ncols_dst) \

ggml/src/ggml-cuda/mmid.cu

@@ -0,0 +1,164 @@
+#include "common.cuh"
+#include "mmid.cuh"
+
+// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
+struct mm_ids_helper_store {
+    uint32_t data;
+
+    __device__ mm_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
+        data = (it & 0x003FFFFF) | (iex_used << 22);
+    }
+
+    __device__ uint32_t it() const {
+        return data & 0x003FFFFF;
+    }
+
+    __device__ uint32_t iex_used() const {
+        return data >> 22;
+    }
+};
+static_assert(sizeof(mm_ids_helper_store) == 4, "unexpected size for mm_ids_helper_store");
+
+// Helper function for mul_mat_id, converts ids to a more convenient format.
+// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
+// ids_dst describes the same mapping but for the dst tensor.
+// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
+template <int n_expert_used_template>
+__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
+    const int expert = blockIdx.x;
+
+    extern __shared__ char data_mm_ids_helper[];
+    mm_ids_helper_store * store = (mm_ids_helper_store *) data_mm_ids_helper;
+
+    int nex_prev   = 0; // Number of columns for experts with a lower index.
+    int it_compact = 0; // Running index for the compact slice of this expert.
+
+    if constexpr (n_expert_used_template == 0) {
+        // Generic implementation:
+        for (int it = 0; it < n_tokens; ++it) {
+            int iex_used = -1; // The index at which the expert is used, if any.
+            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
+                const int expert_used = ids[it*si1 + iex];
+                nex_prev += expert_used < expert;
+                if (expert_used == expert) {
+                    iex_used = iex;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact] = mm_ids_helper_store(it, iex_used);
+            }
+
+            if (warp_reduce_any<warp_size>(iex_used != -1)) {
+                it_compact++;
+            }
+        }
+    } else {
+        // Implementation optimized for specific numbers of experts used:
+        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
+        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
+        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
+            const int it = it0 + threadIdx.x / neu_padded;
+
+            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
+            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
+                ids[it*si1 + iex] : INT_MAX;
+            const int iex_used = expert_used == expert ? iex : -1;
+            nex_prev += expert_used < expert;
+
+            // Whether the threads at this token position have used the expert:
+            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
+
+            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
+            int it_compact_add_lower = 0;
+#pragma unroll
+            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
+                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
+                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
+                    it_compact_add_lower += tmp;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact + it_compact_add_lower] = mm_ids_helper_store(it, iex_used);
+            }
+
+            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
+            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
+        }
+    }
+    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
+
+    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
+        const mm_ids_helper_store store_it = store[itc];
+        const int it       = store_it.it();
+        const int iex_used = store_it.iex_used();
+        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
+        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    expert_bounds[expert] = nex_prev;
+
+    if (expert < static_cast<int>(gridDim.x) - 1) {
+        return;
+    }
+
+    expert_bounds[gridDim.x] = nex_prev + it_compact;
+}
+
+template <int n_expert_used_template>
+static void launch_mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mm_ids_helper_store");
+    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mm_ids_helper_store");
+
+    const int id = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[id].warp_size;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+    CUDA_SET_SHARED_MEMORY_LIMIT(mm_ids_helper<n_expert_used_template>, smpbo);
+
+    const dim3 num_blocks(n_experts, 1, 1);
+    const dim3 block_size(warp_size, 1, 1);
+    const size_t nbytes_shared = n_tokens*sizeof(mm_ids_helper_store);
+    GGML_ASSERT(nbytes_shared <= smpbo);
+    mm_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
+        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
+}
+
+void ggml_cuda_launch_mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    switch (n_expert_used) {
+        case  2:
+            launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  4:
+            launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  6:
+            launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  8:
+            launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case 16:
+            launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case 32:
+            launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        default:
+            launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+    }
+}

ggml/src/ggml-cuda/mmid.cuh

@@ -0,0 +1,5 @@
+#pragma once
+
+void ggml_cuda_launch_mm_ids_helper(
+        const int32_t * ids, int32_t * ids_src1, int32_t * ids_dst, int32_t * expert_bounds,
+        int n_experts, int n_tokens, int n_expert_used, int nchannels_y, int si1, int sis1, cudaStream_t stream);

ggml/src/ggml-cuda/mmq.cu

@@ -1,141 +1,6 @@
 #include "mmq.cuh"
 #include "quantize.cuh"
-
-#include <vector>
-
-// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
-struct mmq_ids_helper_store {
-    uint32_t data;
-
-    __device__ mmq_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
-        data = (it & 0x003FFFFF) | (iex_used << 22);
-    }
-
-    __device__ uint32_t it() const {
-        return data & 0x003FFFFF;
-    }
-
-    __device__ uint32_t iex_used() const {
-        return data >> 22;
-    }
-};
-static_assert(sizeof(mmq_ids_helper_store) == 4, "unexpected size for mmq_ids_helper_store");
-
-// Helper function for mul_mat_id, converts ids to a more convenient format.
-// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
-// ids_dst describes the same mapping but for the dst tensor.
-// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
-template <int n_expert_used_template>
-__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
-static __global__ void mmq_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
-    const int expert = blockIdx.x;
-
-    extern __shared__ char data_mmq_ids_helper[];
-    mmq_ids_helper_store * store = (mmq_ids_helper_store *) data_mmq_ids_helper;
-
-    int nex_prev   = 0; // Number of columns for experts with a lower index.
-    int it_compact = 0; // Running index for the compact slice of this expert.
-
-    if constexpr (n_expert_used_template == 0) {
-        // Generic implementation:
-        for (int it = 0; it < n_tokens; ++it) {
-            int iex_used = -1; // The index at which the expert is used, if any.
-            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
-                const int expert_used = ids[it*si1 + iex];
-                nex_prev += expert_used < expert;
-                if (expert_used == expert) {
-                    iex_used = iex;
-                }
-            }
-
-            if (iex_used != -1) {
-                store[it_compact] = mmq_ids_helper_store(it, iex_used);
-            }
-
-            if (warp_reduce_any<warp_size>(iex_used != -1)) {
-                it_compact++;
-            }
-        }
-    } else {
-        // Implementation optimized for specific numbers of experts used:
-        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
-        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
-        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
-            const int it = it0 + threadIdx.x / neu_padded;
-
-            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
-            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
-                ids[it*si1 + iex] : INT_MAX;
-            const int iex_used = expert_used == expert ? iex : -1;
-            nex_prev += expert_used < expert;
-
-            // Whether the threads at this token position have used the expert:
-            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
-
-            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
-            int it_compact_add_lower = 0;
-#pragma unroll
-            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
-                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
-                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
-                    it_compact_add_lower += tmp;
-                }
-            }
-
-            if (iex_used != -1) {
-                store[it_compact + it_compact_add_lower] = mmq_ids_helper_store(it, iex_used);
-            }
-
-            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
-            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
-        }
-    }
-    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
-
-    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
-        const mmq_ids_helper_store store_it = store[itc];
-        const int it       = store_it.it();
-        const int iex_used = store_it.iex_used();
-        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
-        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
-    }
-
-    if (threadIdx.x != 0) {
-        return;
-    }
-
-    expert_bounds[expert] = nex_prev;
-
-    if (expert < static_cast<int>(gridDim.x) - 1) {
-        return;
-    }
-
-    expert_bounds[gridDim.x] = nex_prev + it_compact;
-}
-
-template <int n_expert_used_template>
-static void launch_mmq_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
-    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mmq_ids_helper_store");
-    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mmq_ids_helper_store");
-
-    const int id = ggml_cuda_get_device();
-    const int warp_size = ggml_cuda_info().devices[id].warp_size;
-    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
-    CUDA_SET_SHARED_MEMORY_LIMIT(mmq_ids_helper<n_expert_used_template>, smpbo);
-
-    const dim3 num_blocks(n_experts, 1, 1);
-    const dim3 block_size(warp_size, 1, 1);
-    const size_t nbytes_shared = n_tokens*sizeof(mmq_ids_helper_store);
-    GGML_ASSERT(nbytes_shared <= smpbo);
-    mmq_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
-        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
-}
+#include "mmid.cuh"
 
 static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     switch (args.type_x) {
@@ -293,36 +158,8 @@ void ggml_cuda_mul_mat_q(
         const int si1  = ids->nb[1] / ggml_element_size(ids);
         const int sis1 = nb12 / nb11;
 
-        switch (n_expert_used) {
-            case  2:
-                launch_mmq_ids_helper< 2> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  4:
-                launch_mmq_ids_helper< 4> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  6:
-                launch_mmq_ids_helper< 6> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  8:
-                launch_mmq_ids_helper< 8> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case 16:
-                launch_mmq_ids_helper<16> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case 32:
-                launch_mmq_ids_helper<32> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            default:
-                launch_mmq_ids_helper< 0> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-        }
+        ggml_cuda_launch_mm_ids_helper((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+            ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
         CUDA_CHECK(cudaGetLastError());
     }
 

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(112, 112);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(128, 128);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(256, 256);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(40, 40);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(576, 512);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(64, 64);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(80, 80);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(96, 96);

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -3,8 +3,17 @@
 from glob import glob
 import os
 
+HEAD_SIZES_KQ = [40, 64, 80, 96, 112, 128, 256, 576]
+
 TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0"]
 
+SOURCE_FATTN_TILE = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE({head_size_kq}, {head_size_v});
+"""
+
 SOURCE_FATTN_VEC = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
 
 #include "../fattn-vec.cuh"
@@ -51,6 +60,11 @@ def get_short_name(long_quant_name):
 for filename in glob("*.cu"):
     os.remove(filename)
 
+for head_size_kq in HEAD_SIZES_KQ:
+    head_size_v = head_size_kq if head_size_kq != 576 else 512
+    with open(f"fattn-tile-instance-dkq{head_size_kq}-dv{head_size_v}.cu", "w") as f:
+        f.write(SOURCE_FATTN_TILE.format(head_size_kq=head_size_kq, head_size_v=head_size_v))
+
 for type_k in TYPES_KV:
     for type_v in TYPES_KV:
         with open(f"fattn-vec-instance-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
@@ -64,7 +78,9 @@ for ncols in [8, 16, 32, 64]:
         with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f:
             f.write(SOURCE_FATTN_MMA_START)
 
-            for head_size_kq in [64, 80, 96, 112, 128, 256, 576]:
+            for head_size_kq in HEAD_SIZES_KQ:
+                if head_size_kq == 40:
+                    continue
                 if head_size_kq != 576 and ncols2 == 16:
                     continue
                 if head_size_kq == 576 and ncols2 != 16:

ggml/src/ggml-cuda/topk-moe.cu

@@ -13,7 +13,7 @@
 
     It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models
 */
-template <size_t n_experts, bool with_norm>
+template <int n_experts, bool with_norm>
 __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * logits,
                                                                   float *       weights,
                                                                   int32_t *     ids,
@@ -204,8 +204,6 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
 
     GGML_ASSERT(ids->nb[1] / ggml_type_size(ids->type) == (size_t) n_experts);
 
-    cudaStream_t stream = ctx.stream();
-
     const int n_expert_used = weights->ne[1];
 
     if (with_norm) {

ggml/src/ggml-cuda/unary.cu

@@ -1,4 +1,5 @@
 #include "unary.cuh"
+#include "convert.cuh"
 
 static __device__ __forceinline__ float op_abs(float x) {
     return fabsf(x);
@@ -375,6 +376,59 @@ void ggml_cuda_op_swiglu_oai(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     swiglu_oai_cuda(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), alpha, limit, stream);
 }
 
+/* CUDA kernel + launcher for xIELU */
+
+template <typename T>
+static __global__ void xielu_kernel(const T * x, T * dst, const int k, float alpha_n, float alpha_p, float beta, float eps) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    const float xi = ggml_cuda_cast<float>(x[i]);
+
+    const float gate_pos = (xi > 0.0f);
+    const float y_pos = alpha_p * xi * xi + beta * xi;
+    const float min_v_eps = fminf(xi, eps);
+    const float y_neg = (expm1f(min_v_eps) - xi) * alpha_n + beta * xi;
+    const float out = gate_pos * y_pos + (1.0f - gate_pos) * y_neg;
+
+    dst[i] = ggml_cuda_cast<T>(out);
+}
+
+template <typename T>
+static void xielu_cuda(const T * x, T * dst, const int k, float alpha_n, float alpha_p, float beta, float eps, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_XIELU_BLOCK_SIZE) / CUDA_XIELU_BLOCK_SIZE;
+    xielu_kernel<<<num_blocks, CUDA_XIELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, alpha_n, alpha_p, beta, eps);
+}
+
+void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const void * src0_d = src0->data;
+    void * dst_d = dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    const 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);
+
+    if (src0->type == GGML_TYPE_F16) {
+        xielu_cuda((const half *)src0_d, (half *)dst_d, ggml_nelements(src0), alpha_n, alpha_p, beta, eps, stream);
+    } else {
+        xielu_cuda((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), alpha_n, alpha_p, beta, eps, stream);
+    }
+}
+
+
+
 /* silu_back */
 
 static __device__ __forceinline__ float op_silu_back(float grad, float x) {

ggml/src/ggml-cuda/unary.cuh

@@ -16,6 +16,7 @@
 #define CUDA_SIN_BLOCK_SIZE 256
 #define CUDA_COS_BLOCK_SIZE 256
 #define CUDA_GLU_BLOCK_SIZE 256
+#define CUDA_XIELU_BLOCK_SIZE 256
 
 void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
@@ -72,3 +73,5 @@ void ggml_cuda_op_swiglu_oai(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/vendors/hip.h

@@ -6,6 +6,10 @@
 #include <hip/hip_fp16.h>
 #include <hip/hip_bf16.h>
 
+#if defined(GGML_HIP_ROCWMMA_FATTN)
+#include <rocwmma/rocwmma-version.hpp>
+#endif // defined(GGML_HIP_ROCWMMA_FATTN)
+
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N

ggml/src/ggml-hip/CMakeLists.txt

@@ -39,12 +39,6 @@ endif()
 find_package(hip     REQUIRED)
 find_package(hipblas REQUIRED)
 find_package(rocblas REQUIRED)
-if (GGML_HIP_ROCWMMA_FATTN)
-    CHECK_INCLUDE_FILE_CXX("rocwmma/rocwmma.hpp" FOUND_ROCWMMA)
-    if (NOT ${FOUND_ROCWMMA})
-        message(FATAL_ERROR "rocwmma has not been found")
-    endif()
-endif()
 
 if (${hip_VERSION} VERSION_LESS 6.1)
     message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
@@ -59,6 +53,8 @@ file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
 list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
 
 file(GLOB   GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
+file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-tile*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
 list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
@@ -117,10 +113,6 @@ if (NOT GGML_HIP_MMQ_MFMA)
     add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
 endif()
 
-if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
-    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
-endif()
-
 if (GGML_HIP_EXPORT_METRICS)
     set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
 endif()

ggml/src/ggml-impl.h

@@ -102,6 +102,9 @@ static bool ggml_op_is_empty(enum ggml_op op) {
     }
 }
 
+static inline float ggml_softplus(float input) {
+    return (input > 20.0f) ? input : logf(1 + expf(input));
+}
 //
 // logging
 //

ggml/src/ggml-metal/ggml-metal-common.cpp

@@ -112,7 +112,7 @@ static bool ggml_mem_ranges_add_dst(ggml_mem_ranges_t mrs, const ggml_tensor * t
 }
 
 bool ggml_mem_ranges_add(ggml_mem_ranges_t mrs, const ggml_tensor * tensor) {
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (tensor->src[i]) {
             ggml_mem_ranges_add_src(mrs, tensor->src[i]);
         }
@@ -173,7 +173,7 @@ static bool ggml_mem_ranges_check_dst(ggml_mem_ranges_t mrs, const ggml_tensor *
 }
 
 bool ggml_mem_ranges_check(ggml_mem_ranges_t mrs, const ggml_tensor * tensor) {
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (tensor->src[i]) {
             if (!ggml_mem_ranges_check_src(mrs, tensor->src[i])) {
                 return false;

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -268,6 +268,25 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_glu(ggml_metal_library_t l
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_SUM);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_op_sum_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows(ggml_metal_library_t lib, const ggml_tensor * op) {
     GGML_ASSERT(op->src[0]->nb[0] == ggml_type_size(op->src[0]->type));
 
@@ -338,7 +357,13 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv(ggml_metal_librar
     char base[256];
     char name[256];
 
-    snprintf(base, 256, "kernel_ssm_conv_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type));
+    const char * suffix = "";
+
+    if (op->src[1]->ne[0] % 4 == 0) {
+        suffix = "_4";
+    }
+
+    snprintf(base, 256, "kernel_ssm_conv_%s_%s%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type), suffix);
     snprintf(name, 256, "%s", base);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
@@ -352,15 +377,15 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv(ggml_metal_librar
 }
 
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan(ggml_metal_library_t lib, const ggml_tensor * op)  {
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+
     char base[256];
     char name[256];
 
-    if (op->src[3]->ne[0] == 1) {
-        snprintf(base, 256, "kernel_ssm_scan_group_%s", ggml_type_name(op->src[0]->type));
-    } else {
-        snprintf(base, 256, "kernel_ssm_scan_%s", ggml_type_name(op->src[0]->type));
-    }
-    snprintf(name, 256, "%s", base);
+    const int nsg = (ne00 + 31)/32;
+
+    snprintf(base, 256, "kernel_ssm_scan_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s_nsg=%d", base, nsg);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {
@@ -369,7 +394,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan(ggml_metal_librar
 
     res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
 
-    ggml_metal_pipeline_set_smem(res, 32*sizeof(float));
+    ggml_metal_pipeline_set_smem(res, 32*sizeof(float)*nsg);
 
     return res;
 }
@@ -918,6 +943,96 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort(ggml_metal_library
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        bool    has_mask,
+        int32_t ncpsg) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+    GGML_UNUSED(op);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_%s",
+            "flash_attn_ext_pad");
+
+    snprintf(name, 256, "%s_mask=%d_ncpsg=%d",
+            base,
+            has_mask,
+            ncpsg);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_bool(cv, has_mask,  FC_FLASH_ATTN_EXT_PAD + 0);
+  //ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_PAD + 1);
+  //ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_PAD + 2);
+  //ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_PAD + 3);
+
+  //ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_PAD + 20);
+  //ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_PAD + 21);
+  //ggml_metal_cv_set_int32(cv, nsg,  FC_FLASH_ATTN_EXT_PAD + 22);
+  //ggml_metal_cv_set_int32(cv, nwg,  FC_FLASH_ATTN_EXT_PAD + 23);
+  //ggml_metal_cv_set_int32(cv, nqptg, FC_FLASH_ATTN_EXT_PAD + 24);
+    ggml_metal_cv_set_int32(cv, ncpsg, FC_FLASH_ATTN_EXT_PAD + 25);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_blk(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        int32_t nqptg,
+        int32_t ncpsg) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+    GGML_UNUSED(op);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_%s",
+            "flash_attn_ext_blk");
+
+    snprintf(name, 256, "%s_nqptg=%d_ncpsg=%d",
+            base,
+            nqptg,
+            ncpsg);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+  //ggml_metal_cv_set_bool(cv, has_mask,  FC_FLASH_ATTN_EXT_BLK + 0);
+  //ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_BLK + 1);
+  //ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_BLK + 2);
+  //ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_BLK + 3);
+
+  //ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_BLK + 20);
+  //ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_BLK + 21);
+  //ggml_metal_cv_set_int32(cv, nsg,  FC_FLASH_ATTN_EXT_BLK + 22);
+  //ggml_metal_cv_set_int32(cv, nwg,  FC_FLASH_ATTN_EXT_BLK + 23);
+    ggml_metal_cv_set_int32(cv, nqptg, FC_FLASH_ATTN_EXT_BLK + 24);
+    ggml_metal_cv_set_int32(cv, ncpsg, FC_FLASH_ATTN_EXT_BLK + 25);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         ggml_metal_library_t lib,
         const ggml_tensor * op,
@@ -925,6 +1040,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
@@ -937,18 +1053,23 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
     const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
     const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
 
+    // do bounds checks for the mask?
+    const bool bc_mask = op->src[3] && (op->src[3]->ne[1] % 8 != 0);
+
     snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
             "flash_attn_ext",
             ggml_type_name(op->src[1]->type),
             dk,
             dv);
 
-    snprintf(name, 256, "%s_mask=%d_sinks=%d_bias=%d_scap=%d_ns10=%d_ns20=%d_nsg=%d",
+    snprintf(name, 256, "%s_mask=%d_sinks=%d_bias=%d_scap=%d_kvpad=%d_bcm=%d_ns10=%d_ns20=%d_nsg=%d",
             base,
             has_mask,
             has_sinks,
             has_bias,
             has_scap,
+            has_kvpad,
+            bc_mask,
             ns10,
             ns20,
             nsg);
@@ -964,6 +1085,9 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
     ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT + 1);
     ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT + 2);
     ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT + 3);
+    ggml_metal_cv_set_bool(cv, has_kvpad, FC_FLASH_ATTN_EXT + 4);
+
+    ggml_metal_cv_set_bool(cv, bc_mask, FC_FLASH_ATTN_EXT + 10);
 
     ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT + 20);
     ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT + 21);
@@ -983,6 +1107,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg,
         int32_t nwg) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
@@ -1002,12 +1127,13 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
             dk,
             dv);
 
-    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_softcap=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
+    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
             base,
             has_mask,
             has_sinks,
             has_bias,
             has_scap,
+            has_kvpad,
             ns10,
             ns20,
             nsg, nwg);
@@ -1023,6 +1149,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
     ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_VEC + 1);
     ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_VEC + 2);
     ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_VEC + 3);
+    ggml_metal_cv_set_bool(cv, has_kvpad, FC_FLASH_ATTN_EXT_VEC + 4);
 
     ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_VEC + 20);
     ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_VEC + 21);
@@ -1374,3 +1501,40 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_timestep_embedding(ggml_me
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_opt_step_adamw(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_OPT_STEP_ADAMW);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_opt_step_adamw_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_opt_step_sgd(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_OPT_STEP_SGD);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_opt_step_sgd_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}

ggml/src/ggml-metal/ggml-metal-device.h

@@ -109,6 +109,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_set_rows          (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -134,6 +135,20 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad               (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_arange            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_timestep_embedding(ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_opt_step_adamw    (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_opt_step_sgd      (ggml_metal_library_t lib, const struct ggml_tensor * op);
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        bool    has_mask,
+        int32_t ncpsg);
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_blk(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        int32_t nqptg,
+        int32_t ncpsg);
 
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         ggml_metal_library_t lib,
@@ -142,6 +157,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg);
 
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
@@ -151,6 +167,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg,
         int32_t nwg);
 

ggml/src/ggml-metal/ggml-metal-device.m

@@ -656,6 +656,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_COS:
         case GGML_OP_LOG:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
@@ -692,7 +693,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
             // for new head sizes, add checks here
-            if (op->src[0]->ne[0] != 40 &&
+            if (op->src[0]->ne[0] != 32 &&
+                op->src[0]->ne[0] != 40 &&
                 op->src[0]->ne[0] != 64 &&
                 op->src[0]->ne[0] != 80 &&
                 op->src[0]->ne[0] != 96 &&
@@ -776,9 +778,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 };
             }
         case GGML_OP_GET_ROWS:
-            {
-                return op->ne[3] == 1;
-            }
+            return true;
         case GGML_OP_SET_ROWS:
             {
                 if (op->src[0]->type != GGML_TYPE_F32) {
@@ -800,6 +800,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                         return false;
                 };
             }
+        case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
+            return has_simdgroup_reduction;
         default:
             return false;
     }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -69,11 +69,20 @@
 #define N_SG_IQ4_XS 2
 
 // function constants offsets
-#define FC_FLASH_ATTN_EXT              100
-#define FC_FLASH_ATTN_EXT_VEC          200
-#define FC_FLASH_ATTN_EXT_VEC_REDUCE   300
-#define FC_MUL_MV                      400
-#define FC_MUL_MM                      500
+#define FC_FLASH_ATTN_EXT_PAD          100
+#define FC_FLASH_ATTN_EXT_BLK          200
+#define FC_FLASH_ATTN_EXT              300
+#define FC_FLASH_ATTN_EXT_VEC          400
+#define FC_FLASH_ATTN_EXT_VEC_REDUCE   500
+#define FC_MUL_MV                      600
+#define FC_MUL_MM                      700
+
+// op-specific constants
+#define OP_FLASH_ATTN_EXT_NQPTG 8
+#define OP_FLASH_ATTN_EXT_NCPSG 64
+
+#define OP_FLASH_ATTN_EXT_VEC_NQPTG 1
+#define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
 // kernel argument structs
 //
@@ -178,6 +187,7 @@ typedef struct {
 } ggml_metal_kargs_clamp;
 
 typedef struct {
+    int64_t  nk0;
     int64_t  ne00;
     int64_t  ne01;
     int64_t  ne02;
@@ -243,6 +253,35 @@ typedef struct {
     int32_t  sect_3;
 } ggml_metal_kargs_rope;
 
+typedef struct {
+    int32_t  ne11;
+    int32_t  ne_12_2; // assume K and V are same shape
+    int32_t  ne_12_3;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
+    int32_t  ne31;
+    int32_t  ne32;
+    int32_t  ne33;
+    uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
+} ggml_metal_kargs_flash_attn_ext_pad;
+
+typedef struct {
+    int32_t  ne01;
+    int32_t  ne30;
+    int32_t  ne31;
+    int32_t  ne32;
+    int32_t  ne33;
+    uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
+} ggml_metal_kargs_flash_attn_ext_blk;
+
 typedef struct {
     int32_t  ne01;
     int32_t  ne02;
@@ -261,6 +300,7 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne31;
     int32_t  ne32;
     int32_t  ne33;
     uint64_t nb31;
@@ -295,6 +335,7 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne31;
     int32_t  ne32;
     int32_t  ne33;
     uint64_t nb31;
@@ -503,6 +544,10 @@ typedef struct{
     float    limit;
 } ggml_metal_kargs_glu;
 
+typedef struct {
+    uint64_t np;
+} ggml_metal_kargs_sum;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
@@ -572,32 +617,45 @@ typedef struct {
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
     uint64_t s_off;
+    uint64_t nb00;
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;
+    uint64_t nb10;
     uint64_t nb11;
     uint64_t nb12;
+    uint64_t ns12;
     uint64_t nb13;
+    uint64_t nb20;
     uint64_t nb21;
+    uint64_t ns21;
     uint64_t nb22;
+    int64_t  ne30;
     uint64_t nb31;
     uint64_t nb41;
     uint64_t nb42;
+    uint64_t ns42;
     uint64_t nb43;
     uint64_t nb51;
     uint64_t nb52;
+    uint64_t ns52;
     uint64_t nb53;
+    uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;
 
 typedef struct {
-    int64_t  ne00;
+    int32_t  ne00t;
+    int32_t  ne00;
     uint64_t nb01;
     uint64_t nb02;
-    int64_t  ne10;
+    uint64_t nb03;
+    int32_t  ne10;
     uint64_t nb10;
     uint64_t nb11;
+    uint64_t nb12;
     uint64_t nb1;
     uint64_t nb2;
+    uint64_t nb3;
 } ggml_metal_kargs_get_rows;
 
 typedef struct {
@@ -719,4 +777,12 @@ typedef struct {
     uint64_t nb01;
 } ggml_metal_kargs_argmax;
 
+typedef struct {
+    int64_t  np;
+} ggml_metal_kargs_opt_step_adamw;
+
+typedef struct {
+    int64_t  np;
+} ggml_metal_kargs_opt_step_sgd;
+
 #endif // GGML_METAL_IMPL

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -226,6 +226,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             GGML_TENSOR_LOCALS(uint64_t, nb0, node->src[0], nb);
             GGML_TENSOR_LOCALS( int64_t, ne1, node->src[1], ne);
             GGML_TENSOR_LOCALS(uint64_t, nb1, node->src[1], nb);
+            GGML_TENSOR_LOCALS( int64_t, ne2, node->src[2], ne);
+            GGML_TENSOR_LOCALS(uint64_t, nb2, node->src[2], nb);
+            GGML_TENSOR_LOCALS( int64_t, ne3, node->src[3], ne);
+            GGML_TENSOR_LOCALS(uint64_t, nb3, node->src[3], nb);
             GGML_TENSOR_LOCALS( int64_t, ne,  node,         ne);
             GGML_TENSOR_LOCALS(uint64_t, nb,  node,         nb);
 
@@ -237,6 +241,14 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
                 GGML_LOG_DEBUG("%s: src1 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[1]->type), ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
                         ggml_is_contiguous(node->src[1]), node->src[1]->name);
             }
+            if (node->src[2]) {
+                GGML_LOG_DEBUG("%s: src2 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[2]->type), ne20, ne21, ne22, ne23, nb20, nb21, nb22, nb23,
+                        ggml_is_contiguous(node->src[2]), node->src[2]->name);
+            }
+            if (node->src[3]) {
+                GGML_LOG_DEBUG("%s: src3 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[3]->type), ne30, ne31, ne32, ne33, nb30, nb31, nb32, nb33,
+                        ggml_is_contiguous(node->src[3]), node->src[3]->name);
+            }
             if (node) {
                 GGML_LOG_DEBUG("%s: node  - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(node->type), ne0, ne1, ne2, ne3, nb0, nb1, nb2, nb3,
                         node->name);
@@ -289,6 +301,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_glu(ctx, idx);
             } break;
+        case GGML_OP_SUM:
+            {
+                n_fuse = ggml_metal_op_sum(ctx, idx);
+            } break;
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
             {
@@ -398,6 +414,14 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_argmax(ctx, idx);
             } break;
+        case GGML_OP_OPT_STEP_ADAMW:
+            {
+                n_fuse = ggml_metal_op_opt_step_adamw(ctx, idx);
+            } break;
+        case GGML_OP_OPT_STEP_SGD:
+            {
+                n_fuse = ggml_metal_op_opt_step_sgd(ctx, idx);
+            } break;
        default:
             {
                 GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(node->op));
@@ -577,6 +601,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
         ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
 
         ggml_metal_kargs_cpy args = {
+            /*.nk0  =*/ ne00,
             /*.ne00 =*/ ne00,
             /*.ne01 =*/ ne01,
             /*.ne02 =*/ ne02,
@@ -827,6 +852,30 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_sum(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op  = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    const uint64_t n = (uint64_t) ggml_nelements(op->src[0]);
+
+    ggml_metal_kargs_sum args = {
+        /*.np =*/ n,
+    };
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_sum(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, 1, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -906,23 +955,31 @@ int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_get_rows(lib, op->src[0]->type);
 
     ggml_metal_kargs_get_rows args = {
-        /*.ne00 =*/ ne00,
-        /*.nb01 =*/ nb01,
-        /*.nb02 =*/ nb02,
-        /*.ne10 =*/ ne10,
-        /*.nb10 =*/ nb10,
-        /*.nb11 =*/ nb11,
-        /*.nb1  =*/ nb1,
-        /*.nb2  =*/ nb2,
+        /*.ne00t =*/ ggml_is_quantized(op->src[0]->type) ? ne00/16 : ne00,
+        /*.ne00  =*/ ne00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne10  =*/ ne10,
+        /*.nb10  =*/ nb10,
+        /*.nb11  =*/ nb11,
+        /*.nb12  =*/ nb12,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
     };
 
+    const int nth = std::min(args.ne00t, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    const int nw0 = (args.ne00t + nth - 1)/nth;
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, ne10, ne11, ne12, 32, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, nw0*ne10, ne11, ne12, nth, 1, 1);
 
     return 1;
 }
@@ -1117,7 +1174,7 @@ int ggml_metal_op_ssm_conv(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
     ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 3);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
 
     ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
 
@@ -1172,25 +1229,36 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.s_off        =*/ ggml_nelements(op->src[1]) * sizeof(float),
+        /*.nb00         =*/ nb00,
         /*.nb01         =*/ nb01,
         /*.nb02         =*/ nb02,
         /*.nb03         =*/ nb03,
+        /*.nb10         =*/ nb10,
         /*.nb11         =*/ nb11,
         /*.nb12         =*/ nb12,
+        /*.ns12         =*/ nb12/nb10,
         /*.nb13         =*/ nb13,
+        /*.nb20         =*/ nb20,
         /*.nb21         =*/ nb21,
+        /*.ns21         =*/ nb21/nb20,
         /*.nb22         =*/ nb22,
+        /*.ne30         =*/ ne30,
         /*.nb31         =*/ nb31,
         /*.nb41         =*/ nb41,
         /*.nb42         =*/ nb42,
+        /*.ns42         =*/ nb42/nb40,
         /*.nb43         =*/ nb43,
         /*.nb51         =*/ nb51,
         /*.nb52         =*/ nb52,
+        /*.ns52         =*/ nb52/nb50,
         /*.nb53         =*/ nb53,
+        /*.nb0          =*/ nb0,
     };
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_ssm_scan(lib, op);
 
+    GGML_ASSERT(d_state <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
     const size_t sms = ggml_metal_pipeline_get_smem(pipeline);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
@@ -1206,13 +1274,7 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, sms, 0);
 
-    if (ne30 == 1) {
-        // Mamba-2
-        ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
-    } else {
-        GGML_ASSERT(d_inner == 1);
-        ggml_metal_encoder_dispatch_threadgroups(enc, n_head, n_seqs, 1, d_state, 1, 1);
-    }
+    ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
 
     return 1;
 }
@@ -1273,26 +1335,23 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
 
     GGML_ASSERT(ne00 % ggml_blck_size(op->src[0]->type) == 0);
 
-    // TODO: support
-    //const int32_t nk00 = ne00/ggml_blck_size(op->type);
-    const int32_t nk00 = ne00;
-
-    int nth = 32; // SIMD width
-
-    while (nth < nk00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
-        nth *= 2;
+    int64_t nk0 = ne00;
+    if (ggml_is_quantized(op->src[0]->type)) {
+        nk0 = ne00/16;
+    } else if (ggml_is_quantized(op->type)) {
+        nk0 = ne00/ggml_blck_size(op->type);
     }
 
-    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    int nth = std::min<int>(nk0, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
     // when rows are small, we can batch them together in a single threadgroup
     int nrptg = 1;
 
     // TODO: relax this constraint in the future
     if (ggml_blck_size(op->src[0]->type) == 1 && ggml_blck_size(op->type) == 1) {
-        if (nth > nk00) {
-            nrptg = (nth + nk00 - 1)/nk00;
-            nth   = nk00;
+        if (nth > nk0) {
+            nrptg = (nth + nk0 - 1)/nk0;
+            nth   = nk0;
 
             if (nrptg*nth > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
                 nrptg--;
@@ -1300,10 +1359,11 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
         }
     }
 
-    nth = std::min(nth, nk00);
+    nth = std::min<int>(nth, nk0);
 
     ggml_metal_kargs_cpy args = {
-        /*.ne00 =*/ nk00,
+        /*.nk0  =*/ nk0,
+        /*.ne00 =*/ ne00,
         /*.ne01 =*/ ne01,
         /*.ne02 =*/ ne02,
         /*.ne03 =*/ ne03,
@@ -1321,12 +1381,14 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
         /*.nb3  =*/ nb3,
     };
 
+    const int nw0 = nrptg == 1 ? (nk0 + nth - 1)/nth : 1;
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, nrptg, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, nw0*(ne01 + nrptg - 1)/nrptg, ne02, ne03, nth, nrptg, 1);
 
     return 1;
 }
@@ -1520,9 +1582,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         !ggml_is_transposed(op->src[1]) &&
         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-        props_dev->has_simdgroup_mm && ne00 >= 64 &&
-        (ne11 > ne11_mm_min || (ggml_is_quantized(op->src[0]->type) && ne12 > 1))) {
-        //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+        props_dev->has_simdgroup_mm && ne00 >= 64 && ne11 > ne11_mm_min) {
+        //GGML_LOG_INFO("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
         // some Metal matrix data types require aligned pointers
         // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
@@ -1875,20 +1936,107 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const ggml_tensor * op) {
     return (ne01 < 20) && (ne00 % 32 == 0);
 }
 
+size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    } else {
+        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    }
+
+    return res;
+}
+
+size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    if (!has_mask) {
+        return res;
+    }
+
+    const bool is_vec = ggml_metal_op_flash_attn_ext_use_vec(op);
+
+    // this optimization is not useful for the vector kernels
+    if (is_vec) {
+        return res;
+    }
+
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
+    const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
+
+    const int64_t ne1 = (ne01 + nqptg - 1)/nqptg;
+    const int64_t ne0 = (ne30 + ncpsg - 1)/ncpsg;
+
+    res += GGML_PAD(ggml_type_size(GGML_TYPE_I8)*ne0*ne1*ne32*ne33, 32);
+
+    return res;
+}
+
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
-    const int64_t nwg = 32;
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
 
-    const int64_t ne01 = op->src[0]->ne[1];
-    const int64_t ne02 = op->src[0]->ne[2];
-    const int64_t ne03 = op->src[0]->ne[3];
-    const int64_t ne20 = op->src[2]->ne[0];
+    size_t res = 0;
 
-    // temp buffer for writing the results from each workgroup
-    // - ne20: the size of the Value head
-    // -  + 2: the S and M values for each intermediate result
-    return ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+    if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        const int64_t nwg = 32;
+
+        // temp buffer for writing the results from each workgroup
+        // - ne20: the size of the Value head
+        // -  + 2: the S and M values for each intermediate result
+        res += ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+    }
+
+    return res;
 }
 
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
@@ -1910,8 +2058,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS( int32_t, nb,  op,         nb);
 
-    GGML_ASSERT(ne00 % 4  == 0);
-    GGML_ASSERT(ne11 % 32 == 0);
+    GGML_ASSERT(ne00 % 4 == 0);
 
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == op->src[2]->type);
@@ -1921,8 +2068,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     GGML_ASSERT(ne12 == ne22);
 
     GGML_ASSERT(!op->src[3] || op->src[3]->type == GGML_TYPE_F16);
-    GGML_ASSERT(!op->src[3] || op->src[3]->ne[1] >= GGML_PAD(op->src[0]->ne[1], 8) &&
-            "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+    GGML_ASSERT(!op->src[3] || op->src[3]->ne[1] >= op->src[0]->ne[1] &&
+            "the Flash-Attention Metal kernel requires the mask to be at least n_queries big");
 
     float scale;
     float max_bias;
@@ -1949,15 +2096,111 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
     GGML_ASSERT(ne01 < 65536);
 
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
+    ggml_metal_buffer_id bid_src2 = ggml_metal_get_buffer_id(op->src[2]);
+    ggml_metal_buffer_id bid_src3 = has_mask  ? ggml_metal_get_buffer_id(op->src[3]) : bid_src0;
+    ggml_metal_buffer_id bid_src4 = has_sinks ? ggml_metal_get_buffer_id(op->src[4]) : bid_src0;
+
+    ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_pad = bid_dst;
+    bid_pad.offs += ggml_nbytes(op);
+
+    ggml_metal_buffer_id bid_blk = bid_pad;
+    bid_blk.offs += ggml_metal_op_flash_attn_ext_extra_pad(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_blk;
+    bid_tmp.offs += ggml_metal_op_flash_attn_ext_extra_blk(op);
+
     if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
         // half8x8 kernel
-        const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
-        const int64_t ncpsg = 64; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPTG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 8  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_pad args0 = {
+                /*.ne11    =*/ne11,
+                /*.ne_12_2 =*/ne12,
+                /*.ne_12_3 =*/ne13,
+                /*.nb11    =*/nb11,
+                /*.nb12    =*/nb12,
+                /*.nb13    =*/nb13,
+                /*.nb21    =*/nb21,
+                /*.nb22    =*/nb22,
+                /*.nb23    =*/nb23,
+                /*.ne31    =*/ne31,
+                /*.ne32    =*/ne32,
+                /*.ne33    =*/ne33,
+                /*.nb31    =*/nb31,
+                /*.nb32    =*/nb32,
+                /*.nb33    =*/nb33,
+            };
+
+            ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
+        }
+
+        if (has_mask) {
+            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_blk args0 = {
+                /*.ne01 =*/ ne01,
+                /*.ne30 =*/ ne30,
+                /*.ne31 =*/ ne31,
+                /*.ne32 =*/ ne32,
+                /*.ne33 =*/ ne33,
+                /*.nb31 =*/ nb31,
+                /*.nb32 =*/ nb32,
+                /*.nb33 =*/ nb33,
+            };
+
+            ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_blk(lib, op, nqptg, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_blk,  2);
+
+            const int32_t nblk1 = ((ne01 + nqptg - 1)/nqptg);
+            const int32_t nblk0 = ((ne30 + ncpsg - 1)/ncpsg);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, nblk0, nblk1, ne32*ne33, 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) == 0);
+        }
+
+        if (need_sync) {
+            ggml_metal_op_concurrency_reset(ctx);
+        }
+
         const int is_q = ggml_is_quantized(op->src[1]->type) ? 1 : 0;
 
         // 2*(2*ncpsg)
@@ -2007,6 +2250,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.nb21          =*/ nb21,
             /*.nb22          =*/ nb22,
             /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
             /*.nb31          =*/ nb31,
@@ -2023,24 +2267,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, nsg);
+        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        if (op->src[3]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 4);
-        }
-        if (op->src[4]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 5);
-        }
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         6);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
+        ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
+        ggml_metal_encoder_set_buffer  (enc, bid_blk,  7);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  8);
 
         ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
@@ -2048,14 +2286,62 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        const int64_t nqptg = 1;  // queries per threadgroup    !! sync with kernel template arguments !!
-        const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
-        const int64_t nkpsg = 1*ncpsg;
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPTG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nkpsg = 1*ncpsg;
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 1  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_pad args0 = {
+                /*.ne11    =*/ne11,
+                /*.ne_12_2 =*/ne12,
+                /*.ne_12_3 =*/ne13,
+                /*.nb11    =*/nb11,
+                /*.nb12    =*/nb12,
+                /*.nb13    =*/nb13,
+                /*.nb21    =*/nb21,
+                /*.nb22    =*/nb22,
+                /*.nb23    =*/nb23,
+                /*.ne31    =*/ne31,
+                /*.ne32    =*/ne32,
+                /*.ne33    =*/ne33,
+                /*.nb31    =*/nb31,
+                /*.nb32    =*/nb32,
+                /*.nb33    =*/nb33,
+            };
+
+            ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
+        }
+
+        if (need_sync) {
+            ggml_metal_op_concurrency_reset(ctx);
+        }
+
         // ne00 + 2*ncpsg*(nsg)
         // for each query, we load it as f16 in shared memory (ne00)
         // and store the soft_max values and the mask
@@ -2120,6 +2406,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.nb21          =*/ nb21,
             /*.nb22          =*/ nb22,
             /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
             /*.nb31          =*/ nb31,
@@ -2136,25 +2423,17 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, nsg, nwg);
+        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        if (op->src[3]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 4);
-        }
-        if (op->src[4]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 5);
-        }
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
 
         const size_t smem = FATTN_SMEM(nsg);
 
@@ -2162,23 +2441,25 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         GGML_ASSERT(smem <= props_dev->max_theadgroup_memory_size);
 
         if (nwg == 1) {
+            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) == 0);
+
             // using 1 workgroup -> write the result directly into dst
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 6);
+            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_dst, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
             ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
         } else {
             // sanity checks
+            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
+
             GGML_ASSERT(ne01*ne02*ne03 == ne1*ne2*ne3);
             GGML_ASSERT((uint64_t)ne1*ne2*ne3 <= (1u << 31));
 
-            ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
-
             // write the results from each workgroup into a temp buffer
-            ggml_metal_buffer_id bid_tmp = bid_dst;
-            bid_tmp.offs += ggml_nbytes(op);
-            ggml_metal_encoder_set_buffer(enc, bid_tmp, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
             ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
@@ -3156,3 +3437,73 @@ int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
 
     return 1;
 }
+
+int ggml_metal_op_opt_step_adamw(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_opt_step_adamw(lib, op);
+
+    const int64_t np = ggml_nelements(op->src[0]);
+    ggml_metal_kargs_opt_step_adamw args = {
+        /*.np =*/ np,
+    };
+
+    int ida = 0;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), ida++);
+
+    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+    const int64_t n = (np + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_opt_step_sgd(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_opt_step_sgd(lib, op);
+
+    const int64_t np = ggml_nelements(op->src[0]);
+    ggml_metal_kargs_opt_step_sgd args = {
+        /*.np =*/ np,
+    };
+
+    int ida = 0;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++);
+
+    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+    const int64_t n = (np + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -39,6 +39,8 @@ size_t ggml_metal_op_mul_mat_id_extra_ids(const struct ggml_tensor * op);
 // return true if we should use the FA vector kernel for this op
 bool ggml_metal_op_flash_attn_ext_use_vec(const struct ggml_tensor * op);
 
+size_t ggml_metal_op_flash_attn_ext_extra_pad(const struct ggml_tensor * op);
+size_t ggml_metal_op_flash_attn_ext_extra_blk(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
 
 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);
@@ -48,6 +50,7 @@ int ggml_metal_op_scale             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_clamp             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_unary             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_glu               (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_sum               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_sum_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
@@ -76,6 +79,8 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argmax            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argsort           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_leaky_relu        (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_opt_step_adamw    (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_opt_step_sgd      (ggml_metal_op_t ctx, int idx);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-metal/ggml-metal.cpp

@@ -193,9 +193,9 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
             } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
-                if (ggml_metal_op_flash_attn_ext_use_vec(tensor)) {
-                    res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
-                }
+                res += ggml_metal_op_flash_attn_ext_extra_pad(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
             } break;
         default:
             break;