vocab : fix ugm tokenizer precision (#13743)

ggml-ci

.devops/intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
+ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04
 
 ## Build Image
 

.devops/musa.Dockerfile

@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc3.1.1
+ARG MUSA_VERSION=rc4.0.1
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}
 
-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
@@ -21,21 +21,14 @@ RUN apt-get update && \
     libcurl4-openssl-dev \
     libgomp1
 
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
-
-RUN pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt
-
 WORKDIR /app
 
 COPY . .
 
-# Use the default MUSA archs if not specified
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.editorconfig

@@ -48,3 +48,7 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
+
+[tools/mtmd/miniaudio.h]
+trim_trailing_whitespace = unset
+insert_final_newline = unset

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

@@ -5,6 +5,10 @@ inputs:
     description: 'CURL version'
     required: false
     default: '8.6.0_6'
+  architecture:
+    description: 'Architecture of the libcurl to download'
+    required: false
+    default: 'win64'
 outputs:
   curl_path:
     description: "Path to the downloaded libcurl"
@@ -18,8 +22,9 @@ runs:
       shell: powershell
       env:
         CURL_VERSION: ${{ inputs.curl_version }}
+        ARCHITECTURE: ${{ inputs.architecture }}
       run: |
-        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
+        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-${env:ARCHITECTURE}-mingw.zip"
         mkdir $env:RUNNER_TEMP/libcurl
         tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
         echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT

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

@@ -140,3 +140,94 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
+
+  ubuntu-24-ppc64el-cpu-cross:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup PowerPC64le
+        run: |
+          sudo dpkg --add-architecture ppc64el
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-powerpc64le-linux-gnu \
+                  g++-14-powerpc64le-linux-gnu \
+                  libcurl4-openssl-dev:ppc64el
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  ubuntu-24-ppc64el-vulkan-cross:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup PowerPC64le
+        run: |
+          sudo dpkg --add-architecture ppc64el
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-powerpc64le-linux-gnu \
+                  g++-14-powerpc64le-linux-gnu \
+                  libvulkan-dev:ppc64el \
+                  libcurl4-openssl-dev:ppc64el
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -307,7 +307,7 @@ jobs:
         run: |
           cd build
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 2700
+          ctest -L main --verbose --timeout 3600
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
@@ -351,7 +351,7 @@ jobs:
 
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
-    container: mthreads/musa:rc3.1.1-devel-ubuntu22.04
+    container: mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 
     steps:
       - name: Clone
@@ -899,7 +899,7 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:

.github/workflows/docker.yml

@@ -42,8 +42,7 @@ jobs:
           - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
-          # Note: the intel images are failing due to an out of disk space error
-          # - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
           #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }

.github/workflows/release.yml

@@ -1,4 +1,4 @@
-name: Create Release
+name: Release
 
 on:
   workflow_dispatch: # allows manual triggering
@@ -227,6 +227,66 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
           name: llama-bin-ubuntu-vulkan-x64.zip
 
+  windows-cpu:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        include:
+          - arch: 'x64'
+          - arch: 'arm64'
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-cpu-${{ matrix.arch }}
+          variant: ccache
+          evict-old-files: 1d
+
+      - name: Install Ninja
+        run: |
+          choco install ninja
+
+      - name: libCURL
+        id: get_libcurl
+        uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
+
+      - name: Build
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" `
+            -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake `
+            -DGGML_NATIVE=OFF `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_OPENMP=OFF `
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
+          7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
+          name: llama-bin-win-cpu-${{ matrix.arch }}.zip
+
   windows:
     runs-on: windows-latest
 
@@ -237,47 +297,30 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
-          #- build: 'openblas-x64'
-          #  defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
-          - build: 'cpu-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF'
-          - build: 'opencl-adreno-arm64'
+          - backend: 'vulkan'
+            arch: 'x64'
+            defines: '-DGGML_VULKAN=ON'
+            target: 'ggml-vulkan'
+          - backend: 'opencl-adreno'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
+            target: 'ggml-opencl'
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
         with:
-          key: windows-latest-cmake-${{ matrix.build }}
+          key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
           variant: ccache
           evict-old-files: 1d
 
-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
-          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/openblas
-          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
-          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
-
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.backend == 'vulkan' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
@@ -291,7 +334,7 @@ jobs:
 
       - name: Install OpenCL Headers and Libs
         id: install_opencl
-        if: ${{ matrix.build == 'opencl-adreno-arm64' }}
+        if: ${{ matrix.backend == 'opencl-adreno' && matrix.arch == 'arm64' }}
         run: |
           git clone https://github.com/KhronosGroup/OpenCL-Headers
           cd OpenCL-Headers
@@ -309,44 +352,22 @@ jobs:
             -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
           cmake --build build-arm64-release --target install --config release
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -S . -B build ${{ matrix.defines }} `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Add libopenblas.dll
-        id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
-          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+          cmake -S . -B build ${{ matrix.defines }} -DGGML_NATIVE=OFF -DGGML_CPU=OFF -DGGML_BACKEND_DL=ON -DLLAMA_CURL=OFF
+          cmake --build build --config Release --target ${{ matrix.target }}
 
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
+          7z a llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
-          name: llama-bin-win-${{ matrix.build }}.zip
+          path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
+          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
     runs-on: windows-2019
@@ -359,8 +380,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
 
       - name: Install ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -379,45 +398,30 @@ jobs:
         run: |
           choco install ninja
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
         shell: cmd
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
           cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DGGML_NATIVE=OFF ^
             -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
             -DGGML_CUDA=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
-            ${{ env.CMAKE_ARGS }}
+            -DLLAMA_CURL=OFF
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+          7z a llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
         run: |
@@ -425,13 +429,13 @@ jobs:
           $dst='.\build\bin\cudart\'
           robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         uses: actions/upload-artifact@v4
         with:
-          path: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
     runs-on: windows-latest
@@ -441,15 +445,14 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -462,15 +465,18 @@ jobs:
         run:  |
           scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
-      # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
       - name: Build
         id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+        shell: cmd
+        run: |
+          call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+          cmake -G "Ninja" -B build ^
+            -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx ^
+            -DCMAKE_BUILD_TYPE=Release ^
+            -DGGML_BACKEND_DL=ON -DBUILD_SHARED_LIBS=ON ^
+            -DGGML_CPU=OFF -DGGML_SYCL=ON ^
+            -DLLAMA_CURL=OFF
+          cmake --build build --target ggml-sycl -j
 
       - name: Build the release package
         id: pack_artifacts
@@ -495,12 +501,12 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
 
           echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
+          7z a llama-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
@@ -508,14 +514,14 @@ jobs:
 
     strategy:
       matrix:
-        gpu_target: [gfx1100, gfx1101, gfx1030]
+        include:
+          - name: "radeon"
+            gpu_targets: "gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
 
       - name: Clone rocWMMA repository
         id: clone_rocwmma
@@ -525,7 +531,7 @@ jobs:
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
         with:
-          key: windows-latest-cmake-hip-release
+          key: windows-latest-cmake-hip-${{ matrix.name }}-x64
           evict-old-files: 1d
 
       - name: Install
@@ -543,50 +549,39 @@ jobs:
         run: |
           & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
           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('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
-            -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=OFF `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+            -DLLAMA_CURL=OFF
+          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
 
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+          7z a llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -648,14 +643,16 @@ jobs:
     runs-on: ubuntu-latest
 
     needs:
-      - ubuntu-22-cpu
-      - ubuntu-22-vulkan
       - windows
+      - windows-cpu
       - windows-cuda
       - windows-sycl
       - windows-hip
+      - ubuntu-22-cpu
+      - ubuntu-22-vulkan
       - macOS-arm64
       - macOS-x64
+      - ios-xcode-build
 
     steps:
       - name: Clone
@@ -673,10 +670,43 @@ jobs:
         uses: actions/download-artifact@v4
         with:
           path: ./artifact
+          merge-multiple: true
 
       - name: Move artifacts
         id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
+        run: |
+          mkdir -p release
+
+          echo "Adding CPU backend files to existing zips..."
+          for arch in x64 arm64; do
+            cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
+            temp_dir=$(mktemp -d)
+            echo "Extracting CPU backend for $arch..."
+            unzip "$cpu_zip" -d "$temp_dir"
+
+            echo "Adding CPU files to $arch zips..."
+            for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
+              if [[ "$target_zip" == "$cpu_zip" ]]; then
+                continue
+              fi
+              echo "Adding CPU backend to $(basename "$target_zip")"
+              realpath_target_zip=$(realpath "$target_zip")
+              (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
+            done
+
+            rm -rf "$temp_dir"
+          done
+
+          echo "Renaming and moving zips to release..."
+          for zip_file in artifact/llama-bin-win-*.zip; do
+            base_name=$(basename "$zip_file" .zip)
+            zip_name="llama-${{ steps.tag.outputs.name }}-${base_name#llama-}.zip"
+            echo "Moving $zip_file to release/$zip_name"
+            mv "$zip_file" "release/$zip_name"
+          done
+
+          echo "Moving other artifacts..."
+          mv -v artifact/*.zip release
 
       - name: Create release
         id: create_release
@@ -695,7 +725,7 @@ jobs:
             const path = require('path');
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact/release')) {
+            for (let file of await fs.readdirSync('./release')) {
               if (path.extname(file) === '.zip') {
                 console.log('uploadReleaseAsset', file);
                 await github.repos.uploadReleaseAsset({
@@ -703,7 +733,7 @@ jobs:
                   repo: context.repo.repo,
                   release_id: release_id,
                   name: file,
-                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                  data: await fs.readFileSync(`./release/${file}`)
                 });
               }
             }

.github/workflows/winget.yml

@@ -0,0 +1,42 @@
+name: Update Winget Package
+
+on:
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    - cron: '28 5 * * *' # Update every day at 5:28 UTC
+
+jobs:
+  update:
+    name: Update Winget Package
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Install cargo binstall
+        uses: cargo-bins/cargo-binstall@268643a6b5ea099f5718ee5cd3ff7dc89a5eb49b
+
+      - name: Install komac
+        run: |
+          cargo binstall komac@2.11.2 -y
+
+      - name: Find latest release
+        id: find_latest_release
+        uses: actions/github-script@v6
+        with:
+          script: |
+            const { data: releases } = await github.rest.repos.listReleases({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+            });
+            console.log("Latest release:", releases[0].tag_name);
+            return releases[0].tag_name;
+
+      - name: Update manifest
+        env:
+          VERSION: ${{ steps.find_latest_release.outputs.result }}
+        run: |
+          echo "Updating manifest..."
+          komac update --version ${{ env.VERSION }} \
+            --urls "https://github.com/ggml-org/llama.cpp/releases/download/${{ env.VERSION }}/llama-${{ env.VERSION }}-bin-win-vulkan-x64.zip" \
+            --token ${{ secrets.WINGET_GITHUB_TOKEN }} \
+            --submit \
+            ggml.llamacpp

README.md

@@ -16,8 +16,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
+- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
@@ -36,7 +37,7 @@ range of hardware - locally and in the cloud.
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2, AVX512 and AMX support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
-- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
@@ -236,7 +237,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
 | [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](docs/build.md#musa) | Moore Threads MTT GPU |
+| [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
@@ -571,4 +572,12 @@ automatically. For example:
 $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 ```
 
-## References
+## Dependencies
+
+- [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
+- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
+- [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
+- [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
+- [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
+- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain

build-xcframework.sh

@@ -117,6 +117,7 @@ setup_framework_structure() {
     # Copy all required headers (common for all platforms)
     cp include/llama.h             ${header_path}
     cp ggml/include/ggml.h         ${header_path}
+    cp ggml/include/ggml-opt.h     ${header_path}
     cp ggml/include/ggml-alloc.h   ${header_path}
     cp ggml/include/ggml-backend.h ${header_path}
     cp ggml/include/ggml-metal.h   ${header_path}

ci/README.md

@@ -54,7 +54,7 @@ docker run --privileged -it \
     -v $HOME/llama.cpp/ci-cache:/ci-cache \
     -v $HOME/llama.cpp/ci-results:/ci-results \
     -v $PWD:/ws -w /ws \
-    mthreads/musa:rc3.1.1-devel-ubuntu22.04
+    mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 ```
 
 Inside the container, execute the following commands:

common/CMakeLists.txt

@@ -73,6 +73,8 @@ add_library(${TARGET} STATIC
     minja/minja.hpp
     ngram-cache.cpp
     ngram-cache.h
+    regex-partial.cpp
+    regex-partial.h
     sampling.cpp
     sampling.h
     speculative.cpp
@@ -119,8 +121,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.10:
-        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
+        # v0.7.20 (+ fix to build on GCC 15):
+        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/arg.cpp

@@ -39,8 +39,8 @@
 using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
-    LLAMA_EXAMPLE_LLAVA,
-    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
+    LLAMA_EXAMPLE_MTMD,
+    LLAMA_EXAMPLE_SERVER,
 };
 
 static std::string read_file(const std::string & fname) {
@@ -1445,6 +1445,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_keep = value;
         }
     ));
+    add_opt(common_arg(
+        {"--swa-full"},
+        string_format("use full-size SWA cache (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)", params.swa_full ? "true" : "false"),
+        [](common_params & params) {
+            params.swa_full = true;
+        }
+    ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1670,7 +1678,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -2057,13 +2065,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.grp_attn_w = value;
         }
     ).set_env("LLAMA_ARG_GRP_ATTN_W").set_examples({LLAMA_EXAMPLE_MAIN}));
-    add_opt(common_arg(
-        {"-dkvc", "--dump-kv-cache"},
-        "verbose print of the KV cache",
-        [](common_params & params) {
-            params.dump_kv_cache = true;
-        }
-    ));
     add_opt(common_arg(
         {"-nkvo", "--no-kv-offload"},
         "disable KV offload",
@@ -2204,39 +2205,40 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
     add_opt(common_arg(
         {"--mmproj"}, "FILE",
-        "path to a multimodal projector file. see tools/mtmd/README.md",
+        "path to a multimodal projector file. see tools/mtmd/README.md\n"
+        "note: if -hf is used, this argument can be omitted",
         [](common_params & params, const std::string & value) {
             params.mmproj.path = value;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ"));
     add_opt(common_arg(
         {"--mmproj-url"}, "URL",
         "URL to a multimodal projector file. see tools/mtmd/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.url = value;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_URL"));
     add_opt(common_arg(
         {"--no-mmproj"},
         "explicitly disable multimodal projector, useful when using -hf",
         [](common_params & params) {
             params.no_mmproj = true;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ"));
     add_opt(common_arg(
         {"--no-mmproj-offload"},
         "do not offload multimodal projector to GPU",
         [](common_params & params) {
             params.mmproj_use_gpu = false;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
-        {"--image"}, "FILE",
-        "path to an image file. use with multimodal models. Specify multiple times for batching",
+        {"--image", "--audio"}, "FILE",
+        "path to an image or audio file. use with multimodal models, can be repeated if you have multiple files\n",
         [](common_params & params, const std::string & value) {
             params.image.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples({LLAMA_EXAMPLE_MTMD}));
     if (llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
@@ -2436,6 +2438,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
+    add_opt(common_arg(
+        {"--no-op-offload"},
+        string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"),
+        [](common_params & params) {
+            params.no_op_offload = true;
+        }
+    ));
     add_opt(common_arg(
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",
@@ -2577,7 +2586,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.n_junk = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_PASSKEY}));
+    ).set_examples({LLAMA_EXAMPLE_PASSKEY, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"--pos"}, "N",
         string_format("position of the passkey in the junk text (default: %d)", params.i_pos),
@@ -2627,13 +2636,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.i_chunk = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--parse-special"},
+        string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
+        [](common_params & params) {
+            params.parse_special = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"-pps"},
         string_format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"),
         [](common_params & params) {
             params.is_pp_shared = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_BENCH}));
+    ).set_examples({LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"-npp"}, "n0,n1,...",
         "number of prompt tokens",
@@ -2852,7 +2868,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.chat_template = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
     add_opt(common_arg(
         {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
         string_format(
@@ -2865,6 +2881,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.chat_template = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
+    add_opt(common_arg(
+        {"--no-prefill-assistant"},
+        string_format(
+            "whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)\n"
+            "when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled\n"
+        ),
+        [](common_params & params) {
+            params.prefill_assistant = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_PREFILL_ASSISTANT"));
     add_opt(common_arg(
         {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
         string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),

common/chat.cpp

@@ -6,6 +6,15 @@
 
 #include <optional>
 
+static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
+    auto time = std::chrono::system_clock::to_time_t(now);
+    auto local_time = *std::localtime(&time);
+    std::ostringstream ss;
+    ss << std::put_time(&local_time, format.c_str());
+    auto res = ss.str();
+    return res;
+}
+
 typedef minja::chat_template common_chat_template;
 
 struct common_chat_templates {
@@ -24,6 +33,7 @@ struct templates_params {
     std::string grammar;
     bool add_generation_prompt = true;
     bool extract_reasoning     = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -939,78 +949,83 @@ static void expect_tool_parameters(const std::string & name, const json & parame
     }
 }
 
-static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {
+static common_chat_params common_chat_params_init_llama_3_x(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {
     auto builtin_tools = json::array();
     common_chat_params data;
-    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-        std::vector<std::string> tool_rules;
+    if (!inputs.tools.is_null()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
 
-        auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
-            if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {
-                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
-                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
-                expect_tool_parameters(name, parameters, {"query"});
-            } else if (name == "python" || name == "code_interpreter") {
-                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
-                expect_tool_parameters(name, parameters, {"code"});
-            } else {
-                return false;
+            auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
+                if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {
+                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
+                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
+                    expect_tool_parameters(name, parameters, {"query"});
+                } else if (name == "python" || name == "code_interpreter") {
+                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
+                    expect_tool_parameters(name, parameters, {"code"});
+                } else {
+                    return false;
+                }
+
+                std::vector<std::string> kvs;
+                for (const auto & [key, value] : parameters.at("properties").items()) {
+                    kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT
+                }
+
+                tool_rules.push_back(
+                    builder.add_rule(
+                        name + "-call",
+                        "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
+                builtin_tools.push_back(name);
+
+                return true;
+            };
+
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+
+                // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
+                if (allow_python_tag_builtin_tools) {
+                    handle_builtin_tool(name, parameters);
+                }
+                tool_rules.push_back(
+                    builder.add_rule(
+                        name + "-call",
+                        "\"{\" space "
+                        "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
+                        "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
+                        "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
+                        "\"}\" space"));
+            });
+            // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+                "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
+            });
+            if (!builtin_tools.empty()) {
+                data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+                data.preserved_tokens.push_back("<|python_tag|>");
             }
-
-            std::vector<std::string> kvs;
-            for (const auto & [key, value] : parameters.at("properties").items()) {
-                kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT
-            }
-
-            tool_rules.push_back(
-                builder.add_rule(
-                    name + "-call",
-                    "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
-            builtin_tools.push_back(name);
-
-            return true;
-        };
-
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string name = function.at("name");
-            auto parameters = function.at("parameters");
-            builder.resolve_refs(parameters);
-
-            // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
-            if (allow_python_tag_builtin_tools) {
-                handle_builtin_tool(name, parameters);
-            }
-            tool_rules.push_back(
-                builder.add_rule(
-                    name + "-call",
-                    "\"{\" space "
-                    "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
-                    "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
-                    "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
-                    "\"}\" space"));
+            // Allow a few empty lines on top of the usual constrained json schema space rule.
+            builder.add_rule("root", string_join(tool_rules, " | "));
+            data.additional_stops.push_back("<|eom_id|>");
         });
-        // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
-        data.grammar_triggers.push_back({
-            COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
-            "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
-        });
-        if (!builtin_tools.empty()) {
-            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
-            data.preserved_tokens.push_back("<|python_tag|>");
-        }
-        // Allow a few empty lines on top of the usual constrained json schema space rule.
-        builder.add_rule("root", string_join(tool_rules, " | "));
-    });
-    data.additional_stops.push_back("<|eom_id|>");
+        data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
+            ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
+            : COMMON_CHAT_FORMAT_LLAMA_3_X;
+    } else {
+        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    }
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
+        {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
     });
-    data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
-        ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
-        : COMMON_CHAT_FORMAT_LLAMA_3_X;
     return data;
 }
 static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) {
@@ -1150,7 +1165,7 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
     data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
-        {"datetime", "Jan 29 2025 13:00:00 GMT"},
+        {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
@@ -1285,55 +1300,59 @@ static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & in
 static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     // https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt
     common_chat_params data;
-    json tools = inputs.tools.is_null() ? inputs.tools : json::array();
-    std::string python_code_argument_name;
-    auto has_raw_python = false;
 
-    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-        std::vector<std::string> tool_rules;
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            const auto & parameters = function.at("parameters");
-            std::string name = function.at("name");
-            if (name == "python" || name == "ipython") {
-                if (!parameters.contains("type")) {
-                    throw std::runtime_error("Missing type in python tool");
-                }
-                has_raw_python = true;
-                const auto & type = parameters.at("type");
-                if (type == "object") {
-                    auto properties = parameters.at("properties");
-                    for (auto it = properties.begin(); it != properties.end(); ++it) {
-                        if (it.value().at("type") == "string") {
-                            if (!python_code_argument_name.empty()) {
-                                throw std::runtime_error("Multiple string arguments found in python tool");
+    if (!inputs.tools.is_null()) {
+        std::string python_code_argument_name;
+        auto has_raw_python = false;
+
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                const auto & parameters = function.at("parameters");
+                std::string name = function.at("name");
+                if (name == "python" || name == "ipython") {
+                    if (!parameters.contains("type")) {
+                        throw std::runtime_error("Missing type in python tool");
+                    }
+                    has_raw_python = true;
+                    const auto & type = parameters.at("type");
+                    if (type == "object") {
+                        auto properties = parameters.at("properties");
+                        for (auto it = properties.begin(); it != properties.end(); ++it) {
+                            if (it.value().at("type") == "string") {
+                                if (!python_code_argument_name.empty()) {
+                                    throw std::runtime_error("Multiple string arguments found in python tool");
+                                }
+                                python_code_argument_name = it.key();
                             }
-                            python_code_argument_name = it.key();
                         }
+                        if (python_code_argument_name.empty()) {
+                            throw std::runtime_error("No string argument found in python tool");
+                        }
+                    } else if (type != "string") {
+                        throw std::runtime_error("Invalid type in python tool: " + type.dump());
                     }
-                    if (python_code_argument_name.empty()) {
-                        throw std::runtime_error("No string argument found in python tool");
-                    }
-                } else if (type != "string") {
-                    throw std::runtime_error("Invalid type in python tool: " + type.dump());
                 }
+                tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
+            });
+            if (has_raw_python) {
+                tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
+                data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+                data.preserved_tokens.push_back("<|python_tag|>");
             }
-            tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
+            auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
+            builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
         });
-        if (has_raw_python) {
-            tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
-            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
-            data.preserved_tokens.push_back("<|python_tag|>");
-        }
-        auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
-        builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
-        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
-    });
+        data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
+    } else {
+        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    }
 
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     // TODO: if (has_raw_python)
-    data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
     return data;
 }
 static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::string & input) {
@@ -1593,6 +1612,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.extract_reasoning = inputs.extract_reasoning;
     params.tool_choice = inputs.tool_choice;
     params.grammar = inputs.grammar;
+    params.now = inputs.now;
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }
@@ -1644,21 +1664,21 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_firefunction_v2(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);
-    }
-
     // Functionary v3.1 (w/ tools)
     if (src.find("<|start_header_id|>") != std::string::npos
         && src.find("<function=") != std::string::npos) {
         return common_chat_params_init_functionary_v3_1_llama_3_1(tmpl, params);
     }
 
-    // Llama 3.1, 3.2, 3.3 (w/ tools)
+    // Llama 3.1, 3.2, 3.3 (also requires date_string so using it even w/o tools)
     if (src.find("<|start_header_id|>ipython<|end_header_id|>") != std::string::npos) {
         auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos;
-        return common_chat_params_init_llama_3_1_tool_calls(tmpl, params, allow_python_tag_builtin_tools);
+        return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
+    }
+
+    // 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);
     }
 
     // Mistral Nemo (w/ tools)

common/chat.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include "common.h"
+#include <chrono>
 #include <string>
 #include <vector>
 
@@ -71,6 +72,7 @@ struct common_chat_templates_inputs {
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
     bool extract_reasoning     = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
 struct common_chat_params {

common/common.cpp

@@ -443,6 +443,25 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
+    if (!str.empty() && !stop.empty()) {
+        const char text_last_char = str.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
+            if (stop[char_index] == text_last_char) {
+                const auto current_partial = stop.substr(0, char_index + 1);
+                if (string_ends_with(str, current_partial)) {
+                    return str.size() - char_index - 1;
+                }
+            }
+        }
+    }
+
+    return std::string::npos;
+}
+
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
     return std::regex_replace(s, special_chars, "\\$0");
@@ -1083,6 +1102,9 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
     }
 
+    mparams.progress_callback           = params.load_progress_callback;
+    mparams.progress_callback_user_data = params.load_progress_callback_user_data;
+
     return mparams;
 }
 
@@ -1113,6 +1135,8 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
+    cparams.op_offload        = !params.no_op_offload;
+    cparams.swa_full          = params.swa_full;
 
     if (params.reranking) {
         cparams.embeddings    = true;
@@ -1305,81 +1329,6 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
     return text;
 }
 
-//
-// KV cache utils
-//
-
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        int seq_count = 0;
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) { seq_count++; }
-        }
-        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    std::unordered_map<llama_seq_id, size_t> seqs;
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] < 0) { continue; }
-            if (seqs.find(cs_curr[j]) == seqs.end()) {
-                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                const size_t sz = seqs.size();
-                seqs[cs_curr[j]] = sz;
-            }
-        }
-        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-    }
-
-    printf("=== Sequence legend: ");
-    for (const auto & it : seqs) {
-        printf("%zu=%d, ", it.second, it.first);
-    }
-    printf("'+'=other sequence ids");
-
-    c_curr = view.cells;
-    cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) {
-                const auto & it = seqs.find(cs_curr[j]);
-                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
-            } else {
-                putchar('.');
-            }
-        }
-        putchar(' ');
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
 //
 // Embedding utils
 //
@@ -1564,3 +1513,20 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
+    const int64_t ne_datapoint = llama_n_ctx(ctx);
+    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
+    ggml_opt_dataset_t result = ggml_opt_dataset_init(
+        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
+
+    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
+    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
+        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
+    }
+
+    return result;
+}

common/common.h

@@ -6,6 +6,7 @@
 
 #include <set>
 #include <string>
+#include <string_view>
 #include <vector>
 #include <sstream>
 
@@ -75,7 +76,7 @@ enum llama_example {
     LLAMA_EXAMPLE_SERVER,
     LLAMA_EXAMPLE_CVECTOR_GENERATOR,
     LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
@@ -322,16 +323,17 @@ struct common_params {
     bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
-    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
+    bool no_op_offload     = false; // globally disable offload host tensor operations to device
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -366,6 +368,7 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;
 
@@ -409,6 +412,7 @@ struct common_params {
 
     bool process_output = false; // collect data for the output tensor
     bool compute_ppl    = true;  // whether to compute perplexity
+    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;
@@ -424,6 +428,11 @@ struct common_params {
 
     // common params
     std::string out_file; // output filename for all example programs
+    // optional callback for model loading progress and cancellation:
+    // called with a progress value between 0.0 and 1.0.
+    // return false from callback to abort model loading or true to continue
+    llama_progress_callback load_progress_callback = NULL;
+    void *                  load_progress_callback_user_data = NULL;
 };
 
 // call once at the start of a program if it uses libcommon
@@ -501,10 +510,9 @@ static bool string_starts_with(const std::string & str,
     return str.rfind(prefix, 0) == 0;
 }
 
-static bool string_ends_with(const std::string & str,
-                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
+// While we wait for C++20's std::string::ends_with...
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -613,16 +621,6 @@ std::string common_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-//
-// KV cache utils
-//
-
-// Dump the KV cache view with the number of sequences per cell.
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
-
-// Dump the KV cache view showing individual sequences in each cell (long output).
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-
 //
 // Embedding utils
 //
@@ -664,3 +662,9 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
+
+//
+// training utils
+//
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);

common/llguidance.cpp

@@ -189,6 +189,7 @@ static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
         /* .tokenize_fn                        = */ llama_sampler_llg_tokenize_fn,
         /* .use_approximate_greedy_tokenize_fn = */ false,
         /* .tokenize_user_data                 = */ vocab,
+        /* .slices                             = */ nullptr,
     };
 
     char           error_buffer[1024];

common/minja/chat-template.hpp

@@ -13,10 +13,12 @@
 #include <chrono>
 #include <cstddef>
 #include <cstdio>
+#include <ctime>
 #include <exception>
 #include <iomanip>
 #include <memory>
 #include <sstream>
+#include <stdexcept>
 #include <string>
 #include <vector>
 
@@ -393,8 +395,8 @@ class chat_template {
 
             for (const auto & message_ : adjusted_messages) {
                 auto message = message_;
-                if (!message.contains("role") || !message.contains("content")) {
-                    throw std::runtime_error("message must have 'role' and 'content' fields: " + message.dump());
+                if (!message.contains("role") || (!message.contains("content") && !message.contains("tool_calls"))) {
+                    throw std::runtime_error("message must have 'role' and one of 'content' or 'tool_calls' fields: " + message.dump());
                 }
                 std::string role = message.at("role");
 
@@ -415,7 +417,6 @@ class chat_template {
                         }
                     }
                     if (polyfill_tool_calls) {
-                        auto content = message.at("content");
                         auto tool_calls = json::array();
                         for (const auto & tool_call : message.at("tool_calls")) {
                             if (tool_call.at("type") != "function") {
@@ -434,8 +435,11 @@ class chat_template {
                         auto obj = json {
                             {"tool_calls", tool_calls},
                         };
-                        if (!content.is_null() && !content.empty()) {
-                            obj["content"] = content;
+                        if (message.contains("content")) {
+                            auto content = message.at("content");
+                            if (!content.is_null() && !content.empty()) {
+                                obj["content"] = content;
+                            }
                         }
                         message["content"] = obj.dump(2);
                         message.erase("tool_calls");

common/minja/minja.hpp

@@ -11,6 +11,7 @@
 #include <algorithm>
 #include <cctype>
 #include <cstddef>
+#include <cstdint>
 #include <cmath>
 #include <exception>
 #include <functional>
@@ -233,7 +234,7 @@ public:
       }
     } else if (is_object()) {
       if (!index.is_hashable())
-        throw std::runtime_error("Unashable type: " + index.dump());
+        throw std::runtime_error("Unhashable type: " + index.dump());
       auto it = object_->find(index.primitive_);
       if (it == object_->end())
         throw std::runtime_error("Key not found: " + index.dump());
@@ -252,7 +253,7 @@ public:
       auto index = key.get<int>();
       return array_->at(index < 0 ? array_->size() + index : index);
     } else if (object_) {
-      if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+      if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
       auto it = object_->find(key.primitive_);
       if (it == object_->end()) return Value();
       return it->second;
@@ -261,7 +262,7 @@ public:
   }
   void set(const Value& key, const Value& value) {
     if (!object_) throw std::runtime_error("Value is not an object: " + dump());
-    if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+    if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
     (*object_)[key.primitive_] = value;
   }
   Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {
@@ -398,7 +399,7 @@ public:
       }
       return false;
     } else if (object_) {
-      if (!value.is_hashable()) throw std::runtime_error("Unashable type: " + value.dump());
+      if (!value.is_hashable()) throw std::runtime_error("Unhashable type: " + value.dump());
       return object_->find(value.primitive_) != object_->end();
     } else {
       throw std::runtime_error("contains can only be called on arrays and objects: " + dump());
@@ -416,7 +417,7 @@ public:
     return const_cast<Value*>(this)->at(index);
   }
   Value& at(const Value & index) {
-    if (!index.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+    if (!index.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
     if (is_array()) return array_->at(index.get<int>());
     if (is_object()) return object_->at(index.primitive_);
     throw std::runtime_error("Value is not an array or object: " + dump());
@@ -676,8 +677,8 @@ public:
 class VariableExpr : public Expression {
     std::string name;
 public:
-    VariableExpr(const Location & location, const std::string& n)
-      : Expression(location), name(n) {}
+    VariableExpr(const Location & loc, const std::string& n)
+      : Expression(loc), name(n) {}
     std::string get_name() const { return name; }
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!context->contains(name)) {
@@ -1200,9 +1201,9 @@ public:
 
 class SliceExpr : public Expression {
 public:
-    std::shared_ptr<Expression> start, end;
-    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
-      : Expression(loc), start(std::move(s)), end(std::move(e)) {}
+    std::shared_ptr<Expression> start, end, step;
+    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e, std::shared_ptr<Expression> && st = nullptr)
+      : Expression(loc), start(std::move(s)), end(std::move(e)), step(std::move(st)) {}
     Value do_evaluate(const std::shared_ptr<Context> &) const override {
         throw std::runtime_error("SliceExpr not implemented");
     }
@@ -1219,18 +1220,35 @@ public:
         if (!index) throw std::runtime_error("SubscriptExpr.index is null");
         auto target_value = base->evaluate(context);
         if (auto slice = dynamic_cast<SliceExpr*>(index.get())) {
-          auto start = slice->start ? slice->start->evaluate(context).get<int64_t>() : 0;
-          auto end = slice->end ? slice->end->evaluate(context).get<int64_t>() : (int64_t) target_value.size();
+          auto len = target_value.size();
+          auto wrap = [len](int64_t i) -> int64_t {
+            if (i < 0) {
+              return i + len;
+            }
+            return i;
+          };
+          int64_t step = slice->step ? slice->step->evaluate(context).get<int64_t>() : 1;
+          if (!step) {
+            throw std::runtime_error("slice step cannot be zero");
+          }
+          int64_t start = slice->start ? wrap(slice->start->evaluate(context).get<int64_t>()) : (step < 0 ? len - 1 : 0);
+          int64_t end = slice->end ? wrap(slice->end->evaluate(context).get<int64_t>()) : (step < 0 ? -1 : len);
           if (target_value.is_string()) {
             std::string s = target_value.get<std::string>();
-            if (start < 0) start = s.size() + start;
-            if (end < 0) end = s.size() + end;
-            return s.substr(start, end - start);
+
+            std::string result;
+            if (start < end && step == 1) {
+              result = s.substr(start, end - start);
+            } else {
+              for (int64_t i = start; step > 0 ? i < end : i > end; i += step) {
+                result += s[i];
+              }
+            }
+            return result;
+
           } else if (target_value.is_array()) {
-            if (start < 0) start = target_value.size() + start;
-            if (end < 0) end = target_value.size() + end;
             auto result = Value::array();
-            for (auto i = start; i < end; ++i) {
+            for (int64_t i = start; step > 0 ? i < end : i > end; i += step) {
               result.push_back(target_value.at(i));
             }
             return result;
@@ -1305,6 +1323,8 @@ public:
               if (name == "iterable") return l.is_iterable();
               if (name == "sequence") return l.is_array();
               if (name == "defined") return !l.is_null();
+              if (name == "true") return l.to_bool();
+              if (name == "false") return !l.to_bool();
               throw std::runtime_error("Unknown type for 'is' operator: " + name);
             };
             auto value = eval();
@@ -1520,6 +1540,10 @@ public:
             vargs.expectArgs("endswith method", {1, 1}, {0, 0});
             auto suffix = vargs.args[0].get<std::string>();
             return suffix.length() <= str.length() && std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
+          } else if (method->get_name() == "startswith") {
+            vargs.expectArgs("startswith method", {1, 1}, {0, 0});
+            auto prefix = vargs.args[0].get<std::string>();
+            return prefix.length() <= str.length() && std::equal(prefix.begin(), prefix.end(), str.begin());
           } else if (method->get_name() == "title") {
             vargs.expectArgs("title method", {0, 0}, {0, 0});
             auto res = str;
@@ -2082,28 +2106,37 @@ private:
 
       while (it != end && consumeSpaces() && peekSymbols({ "[", "." })) {
         if (!consumeToken("[").empty()) {
-            std::shared_ptr<Expression> index;
+          std::shared_ptr<Expression> index;
+          auto slice_loc = get_location();
+          std::shared_ptr<Expression> start, end, step;
+          bool has_first_colon = false, has_second_colon = false;
+
+          if (!peekSymbols({ ":" })) {
+            start = parseExpression();
+          }
+
+          if (!consumeToken(":").empty()) {
+            has_first_colon = true;
+            if (!peekSymbols({ ":", "]" })) {
+              end = parseExpression();
+            }
             if (!consumeToken(":").empty()) {
-              auto slice_end = parseExpression();
-              index = std::make_shared<SliceExpr>(slice_end->location, nullptr, std::move(slice_end));
-            } else {
-              auto slice_start = parseExpression();
-              if (!consumeToken(":").empty()) {
-                consumeSpaces();
-                if (peekSymbols({ "]" })) {
-                  index = std::make_shared<SliceExpr>(slice_start->location, std::move(slice_start), nullptr);
-                } else {
-                  auto slice_end = parseExpression();
-                  index = std::make_shared<SliceExpr>(slice_start->location, std::move(slice_start), std::move(slice_end));
-                }
-              } else {
-                index = std::move(slice_start);
+              has_second_colon = true;
+              if (!peekSymbols({ "]" })) {
+                step = parseExpression();
               }
             }
-            if (!index) throw std::runtime_error("Empty index in subscript");
-            if (consumeToken("]").empty()) throw std::runtime_error("Expected closing bracket in subscript");
+          }
 
-            value = std::make_shared<SubscriptExpr>(value->location, std::move(value), std::move(index));
+          if ((has_first_colon || has_second_colon) && (start || end || step)) {
+            index = std::make_shared<SliceExpr>(slice_loc, std::move(start), std::move(end), std::move(step));
+          } else {
+            index = std::move(start);
+          }
+          if (!index) throw std::runtime_error("Empty index in subscript");
+          if (consumeToken("]").empty()) throw std::runtime_error("Expected closing bracket in subscript");
+
+          value = std::make_shared<SubscriptExpr>(value->location, std::move(value), std::move(index));
         } else if (!consumeToken(".").empty()) {
             auto identifier = parseIdentifier();
             if (!identifier) throw std::runtime_error("Expected identifier in subscript");

common/regex-partial.cpp

@@ -0,0 +1,204 @@
+#include "regex-partial.h"
+#include "common.h"
+#include <functional>
+#include <optional>
+
+common_regex::common_regex(const std::string & pattern) :
+    pattern(pattern),
+    rx(pattern),
+    rx_reversed_partial(regex_to_reversed_partial_regex(pattern)) {}
+
+common_regex_match common_regex::search(const std::string & input, size_t pos, bool as_match) const {
+    std::smatch match;
+    if (pos > input.size()) {
+        throw std::runtime_error("Position out of bounds");
+    }
+    auto start = input.begin() + pos;
+    auto found = as_match
+        ? std::regex_match(start, input.end(), match, rx)
+        : std::regex_search(start, input.end(), match, rx);
+    if (found) {
+        common_regex_match res;
+        res.type = COMMON_REGEX_MATCH_TYPE_FULL;
+        for (size_t i = 0; i < match.size(); ++i) {
+            auto begin = pos + match.position(i);
+            res.groups.emplace_back(begin, begin + match.length(i));
+        }
+        return res;
+    }
+    std::match_results<std::string::const_reverse_iterator> srmatch;
+    if (std::regex_match(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial)) {
+        auto group = srmatch[1].str();
+        if (group.length() != 0) {
+            auto it = srmatch[1].second.base();
+            // auto position = static_cast<size_t>(std::distance(input.begin(), it));
+            if ((!as_match) || it == input.begin()) {
+                common_regex_match res;
+                res.type = COMMON_REGEX_MATCH_TYPE_PARTIAL;
+                const size_t begin = std::distance(input.begin(), it);
+                const size_t end = input.size();
+                if (begin == std::string::npos || end == std::string::npos || begin > end) {
+                    throw std::runtime_error("Invalid range");
+                }
+                res.groups.push_back({begin, end});
+                return res;
+            }
+        }
+    }
+    return {};
+}
+
+/*
+  Transforms a regex pattern to a partial match pattern that operates on a reversed input string to find partial final matches of the original pattern.
+
+  Ideally we'd like to use boost::match_partial (https://beta.boost.org/doc/libs/1_59_0/libs/regex/doc/html/boost_regex/partial_matches.html)
+  to see if a string ends with a partial regex match, but but it's not in std::regex yet.
+  Instead, we'll the regex into a partial match regex operating as a full match on the reverse iterators of the input.
+
+  - /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:(?:d)?c)?b)?a).*
+  - /a|b/ -> (a|b).*
+  - /a*?/ -> error, could match ""
+  - /a*b/ -> ((?:b)?a*+).* (final repetitions become eager)
+  - /.*?ab/ -> ((?:b)?a).* (merge .*)
+  - /a.*?b/ -> ((?:b)?.*?a).* (keep reluctant matches)
+  - /a(bc)d/ -> ((?:(?:d)?(?:(?:c)?b))?a).*
+  - /a(bc|de)/ -> ((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a).*
+  - /ab{2,4}c/ -> abbb?b?c -> ((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a).*
+
+  The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern
+  (i.e. just where the final .* starts in the inverted pattern; all other groups are turned into non-capturing groups, and reluctant quantifiers are ignored)
+*/
+std::string regex_to_reversed_partial_regex(const std::string & pattern) {
+    auto it = pattern.begin();
+    const auto end = pattern.end();
+
+    std::function<std::string()> process = [&]() {
+        std::vector<std::vector<std::string>> alternatives(1);
+        std::vector<std::string> * sequence = &alternatives.back();
+
+        while (it != end) {
+            if (*it == '[') {
+                auto start = it;
+                ++it;
+                while (it != end) {
+                    if ((*it == '\\') && (++it != end)) {
+                        ++it;
+                    } else if ((it != end) && (*it == ']')) {
+                        break;
+                    } else {
+                        ++it;
+                    }
+                }
+                if (it == end) {
+                    throw std::runtime_error("Unmatched '[' in pattern");
+                }
+                ++it;
+                sequence->push_back(std::string(start, it));
+            } else if (*it == '*' || *it == '?' || *it == '+') {
+                if (sequence->empty()) {
+                    throw std::runtime_error("Quantifier without preceding element");
+                }
+                sequence->back() += *it;
+                auto is_star = *it == '*';
+                ++it;
+                if (is_star) {
+                    if (*it == '?') {
+                        ++it;
+                    }
+                }
+            } else if (*it == '{') {
+                if (sequence->empty()) {
+                    throw std::runtime_error("Repetition without preceding element");
+                }
+                ++it;
+                auto start = it;
+                while (it != end && *it != '}') {
+                    ++it;
+                }
+                if (it == end) {
+                    throw std::runtime_error("Unmatched '{' in pattern");
+                }
+                auto parts = string_split(std::string(start, it), ",");
+                ++it;
+                if (parts.size() > 2) {
+                    throw std::runtime_error("Invalid repetition range in pattern");
+                }
+
+                auto parseOptInt = [&](const std::string & s, const std::optional<int> & def = std::nullopt) -> std::optional<int> {
+                    if (s.empty()) {
+                        return def;
+                    }
+                    return std::stoi(s);
+                };
+                auto min = parseOptInt(parts[0], 0);
+                auto max = parts.size() == 1 ? min : parseOptInt(parts[1]);
+                if (min && max && *max < *min) {
+                    throw std::runtime_error("Invalid repetition range in pattern");
+                }
+                // Brutal but... let's repeat at least min times, then ? for the delta between min & max (or * for unbounded)
+                auto part = sequence->back();
+                sequence->pop_back();
+                for (int i = 0; i < *min; i++) {
+                    sequence->push_back(part);
+                }
+                if (max) {
+                    for (int i = *min; i < *max; i++) {
+                        sequence->push_back(part + "?");
+                    }
+                } else {
+                    sequence->push_back(part + "*");
+                }
+            } else if (*it == '(') {
+                ++it;
+                if (it != end && *it == '?' && (it + 1 != end) && *(it + 1) == ':') {
+                    it += 2;
+                }
+                auto sub = process();
+                if (*it != ')') {
+                    throw std::runtime_error("Unmatched '(' in pattern");
+                }
+                ++it;
+                auto & part = sequence->emplace_back("(?:");
+                part += sub;
+                part += ")";
+            } else if (*it == ')') {
+                break;
+            } else if (*it == '|') {
+                ++it;
+                alternatives.emplace_back();
+                sequence = &alternatives.back();
+            } else if (*it == '\\' && (++it != end)) {
+                auto str = std::string("\\") + *it;
+                sequence->push_back(str);
+                ++it;
+            } else if (it != end) {
+                sequence->push_back(std::string(1, *it));
+                ++it;
+            }
+        }
+
+        // /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:d)?c)?b)?a).*
+        // if n(=4) parts, opening n-1(=3) non-capturing groups after the 1 capturing group
+        // We'll do the outermost capturing group and final .* in the enclosing function.
+        std::vector<std::string> res_alts;
+        for (const auto & parts : alternatives) {
+            auto & res = res_alts.emplace_back();
+            for (size_t i = 0; i < parts.size() - 1; i++) {
+                res += "(?:";
+            }
+            for (auto it = parts.rbegin(); it != parts.rend(); ++it) {
+                res += *it;
+                if (it != parts.rend() - 1) {
+                    res += ")?";
+                }
+            }
+        }
+        return string_join(res_alts, "|");
+    };
+    auto res = process();
+    if (it != end) {
+        throw std::runtime_error("Unmatched '(' in pattern");
+    }
+
+    return "(" + res + ")[\\s\\S]*";
+}

common/regex-partial.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include <regex>
+#include <string>
+
+enum common_regex_match_type {
+    COMMON_REGEX_MATCH_TYPE_NONE,
+    COMMON_REGEX_MATCH_TYPE_PARTIAL,
+    COMMON_REGEX_MATCH_TYPE_FULL,
+};
+
+struct common_string_range {
+    size_t begin;
+    size_t end;
+    common_string_range(size_t begin, size_t end) : begin(begin), end(end) {
+        if (begin > end) {
+            throw std::runtime_error("Invalid range");
+        }
+    }
+    // prevent default ctor
+    common_string_range() = delete;
+    bool empty() const {
+        return begin == end;
+    }
+    bool operator==(const common_string_range & other) const {
+        return begin == other.begin && end == other.end;
+    }
+};
+
+struct common_regex_match {
+    common_regex_match_type type = COMMON_REGEX_MATCH_TYPE_NONE;
+    std::vector<common_string_range> groups;
+
+    bool operator==(const common_regex_match & other) const {
+        return type == other.type && groups == other.groups;
+    }
+    bool operator!=(const common_regex_match & other) const {
+        return !(*this == other);
+    }
+};
+
+class common_regex {
+    std::string pattern;
+    std::regex rx;
+    std::regex rx_reversed_partial;
+
+  public:
+    explicit common_regex(const std::string & pattern);
+
+    common_regex_match search(const std::string & input, size_t pos, bool as_match = false) const;
+
+    const std::string & str() const { return pattern; }
+};
+
+// For testing only (pretty print of failures).
+std::string regex_to_reversed_partial_regex(const std::string & pattern);

convert_hf_to_gguf.py

@@ -45,7 +45,7 @@ class SentencePieceTokenTypes(IntEnum):
 
 class ModelType(IntEnum):
     TEXT = 1
-    VISION = 2
+    MMPROJ = 2
 
 
 AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")
@@ -54,7 +54,7 @@ AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")
 class ModelBase:
     _model_classes: dict[ModelType, dict[str, type[ModelBase]]] = {
         ModelType.TEXT: {},
-        ModelType.VISION: {},
+        ModelType.MMPROJ: {},
     }
 
     dir_model: Path
@@ -88,7 +88,7 @@ class ModelBase:
                  small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
-                type(self) is VisionModel:
+                type(self) is MmprojModel:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
         self.dir_model = dir_model
@@ -308,6 +308,8 @@ class ModelBase:
                             gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
                             gguf.MODEL_TENSOR.POSNET_NORM1,
                             gguf.MODEL_TENSOR.POSNET_NORM2,
+                            gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
+                            gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -426,14 +428,18 @@ class ModelBase:
             logger.warning(f"Failed to load model config from {dir_model}: {e}")
             logger.warning("Trying to load config.json instead")
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-                return json.load(f)
+                config = json.load(f)
+                if "llm_config" in config:
+                    # rename for InternVL
+                    config["text_config"] = config["llm_config"]
+                return config
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
         assert names
 
         def func(modelcls: AnyModel) -> AnyModel:
-            model_type = ModelType.VISION if modelcls.model_arch == gguf.MODEL_ARCH.CLIP_VISION else ModelType.TEXT
+            model_type = ModelType.MMPROJ if modelcls.model_arch == gguf.MODEL_ARCH.MMPROJ else ModelType.TEXT
             for name in names:
                 cls._model_classes[model_type][name] = modelcls
             return modelcls
@@ -794,6 +800,9 @@ class TextModel(ModelBase):
         if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
             # ref: https://huggingface.co/mistral-community/pixtral-12b
             res = "pixtral"
+        if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
+            # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
+            res = "seed-coder"
 
         if res is None:
             logger.warning("\n")
@@ -1106,60 +1115,87 @@ class TextModel(ModelBase):
             self.gguf_writer.add_pooling_type(pooling_type)
 
 
-class VisionModel(ModelBase):
-    model_type = ModelType.VISION
-    model_arch = gguf.MODEL_ARCH.CLIP_VISION
+class MmprojModel(ModelBase):
+    model_type = ModelType.MMPROJ
+    model_arch = gguf.MODEL_ARCH.MMPROJ
     preprocessor_config: dict[str, Any]
     global_config: dict[str, Any]
 
+    has_vision_encoder: bool = True # by default
+    has_audio_encoder: bool = False
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
 
-        if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION:
-            raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")
+        if self.model_arch != gguf.MODEL_ARCH.MMPROJ:
+            raise TypeError("MmprojModel must be subclassed with model_arch = gguf.MODEL_ARCH.MMPROJ")
+
+        if self.has_vision_encoder and self.has_audio_encoder:
+            raise NotImplementedError("both vision + audio not supported yet")
 
         # get n_embd of the text model
         if "text_config" not in self.hparams:
             self.hparams["text_config"] = {}
+        if "audio_config" not in self.hparams:
+            self.hparams["audio_config"] = {}
         text_config = {**self.hparams, **self.hparams["text_config"]}
         self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
         assert self.n_embd_text > 0, "n_embd not found in hparams"
 
-        if "vision_config" not in self.hparams:
-            raise ValueError("vision_config not found in hparams")
         # move vision config to the top level, while preserving the original hparams in global_config
         self.global_config = self.hparams
-        self.hparams = self.hparams["vision_config"]
+
+        if "vision_config" in self.hparams:
+            self.hparams = self.hparams["vision_config"]
+        elif "audio_config" in self.hparams:
+            self.hparams = self.hparams["audio_config"]
+        else:
+            raise ValueError("vision_config / audio_config not found in hparams")
 
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"])
-        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count)
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
 
         # load preprocessor config
         with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
             self.preprocessor_config = json.load(f)
 
     def set_type(self):
-        self.gguf_writer.add_type(gguf.GGUFType.CLIP_VISION)
+        self.gguf_writer.add_type(gguf.GGUFType.MMPROJ)
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
-        self.gguf_writer.add_vision_has_vision_encoder(True)
 
-        # vision config
-        self.gguf_writer.add_vision_image_size(self.find_hparam(["image_size"]))
-        self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
-        self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
-        self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_vision_block_count(self.block_count)
-        self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
+        if self.has_vision_encoder:
+            self.gguf_writer.add_clip_has_vision_encoder(True)
+            self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
 
-        # preprocessor config
-        self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
-        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
+            # vision config
+            self.gguf_writer.add_vision_image_size(self.find_hparam(["image_size"]))
+            self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
+            self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
+            self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
+            self.gguf_writer.add_vision_block_count(self.block_count)
+            self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
+
+            # preprocessor config
+            self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
+            self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
+
+        elif self.has_audio_encoder:
+            self.gguf_writer.add_clip_has_audio_encoder(True)
+            self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+
+            # audio config
+            self.gguf_writer.add_audio_embedding_length(self.find_hparam(["hidden_size"]))
+            self.gguf_writer.add_audio_feed_forward_length(self.find_hparam(["intermediate_size"]))
+            self.gguf_writer.add_audio_block_count(self.block_count)
+            self.gguf_writer.add_audio_head_count(self.find_hparam(["num_attention_heads"]))
+
+        else:
+            raise ValueError("MmprojModel must have either vision or audio encoder")
 
     def write_vocab(self):
-        raise ValueError("VisionModel does not support vocab writing")
+        raise ValueError("MmprojModel does not support vocab writing")
 
 
 @ModelBase.register("GPTNeoXForCausalLM")
@@ -1943,7 +1979,7 @@ class LlamaModel(TextModel):
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
 )
-class LlavaVisionModel(VisionModel):
+class LlavaVisionModel(MmprojModel):
     img_break_tok_id = -1
 
     def __init__(self, *args, **kwargs):
@@ -1969,7 +2005,7 @@ class LlavaVisionModel(VisionModel):
         super().set_gguf_parameters()
         hparams = self.hparams
         if hparams["model_type"] == "pixtral":
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PIXTRAL)
             self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
 
             # hidden_act
@@ -2008,7 +2044,7 @@ class LlavaVisionModel(VisionModel):
 
 
 @ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")
-class SmolVLMModel(VisionModel):
+class SmolVLMModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         if self.hparams["model_type"] == "smolvlm_vision":
@@ -2020,7 +2056,7 @@ class SmolVLMModel(VisionModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.IDEFICS3)
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.IDEFICS3)
         self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
         self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("scale_factor", 2))
         self.gguf_writer.add_vision_use_gelu(True)
@@ -2062,6 +2098,9 @@ class Llama4Model(LlamaModel):
         self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+
         # split the gate_up into gate and up
         if "gate_up_proj" in name:
             name_up = name.replace("gate_up_proj", "up_proj.weight")
@@ -2082,6 +2121,26 @@ class Llama4Model(LlamaModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Llama4ForConditionalGeneration")
+class Llama4VisionModel(MmprojModel):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LLAMA4)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams["norm_eps"])
+        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / self.hparams["pixel_shuffle_ratio"]))
+        assert self.hparams["hidden_act"] == "gelu"
+        self.gguf_writer.add_vision_use_gelu(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid # unused
+        if "multi_modal_projector" in name or "vision_model" in name:
+            # process vision tensors
+            if "positional_embedding_vlm" in name and ".weight" not in name:
+                name += ".weight"
+            return [(self.map_tensor_name(name), data_torch)]
+        return []
+
+
 @ModelBase.register("Mistral3ForConditionalGeneration")
 class Mistral3Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
@@ -2606,10 +2665,15 @@ class Qwen2Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if self.hf_arch == "Qwen2Model":
             name = f"model.{name}"  # map to Qwen2ForCausalLM tensors
+        if "language_model." in name:
+            name = name.replace("language_model.", "") # for InternVL
+        if name.startswith("mlp") or name.startswith("vision_model"):
+            # skip visual tensors
+            return []
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
+@ModelBase.register("Qwen2VLModel", "Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2VL
 
@@ -2633,8 +2697,8 @@ class Qwen2VLModel(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
-class Qwen2VLVisionModel(VisionModel):
+@ModelBase.register("Qwen2VLModel", "Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
+class Qwen2VLVisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.hparams["image_size"] = self.hparams.get("image_size", 560)
@@ -2649,9 +2713,9 @@ class Qwen2VLVisionModel(VisionModel):
         super().set_gguf_parameters()
         hparams = self.hparams
         if self.global_config['model_type'] == 'qwen2_vl':
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN2VL)
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN2VL)
         elif self.global_config['model_type'] == 'qwen2_5_vl':
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL)
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN25VL)
             self.gguf_writer.add_vision_use_silu(True)
             # find n_wa_pattern (window attention pattern)
             fullatt_block_indexes = hparams.get("fullatt_block_indexes")
@@ -2709,6 +2773,62 @@ class Qwen2VLVisionModel(VisionModel):
         return [] # skip other tensors
 
 
+@ModelBase.register("InternVisionModel")
+class InternVisionModel(MmprojModel):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.INTERNVL)
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
+        # hidden_act
+        if hparams["hidden_act"] == "silu":
+            self.gguf_writer.add_vision_use_silu(True)
+        elif hparams["hidden_act"] == "gelu":
+            self.gguf_writer.add_vision_use_gelu(True)
+        else:
+            raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+        # downsample_ratio
+        downsample_ratio = self.global_config.get("downsample_ratio")
+        assert downsample_ratio is not None
+        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, name, n_dims  # unused
+        if ".patch_embd." in new_name:
+            return gguf.GGMLQuantizationType.F16
+        if ".position_embd." in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        if name.startswith("vision_model") or name.startswith("mlp"):
+            # process visual tensors
+            # correct name
+            if name.startswith("vision_model"):
+                name = "vision_tower." + name
+            if (".ls" in name or "position_embedding" in name) and not name.endswith(".weight"):
+                name += ".weight"
+            # split QKV tensors if needed
+            if ".qkv." in name:
+                if data_torch.ndim == 2: # weight
+                    c3, _ = data_torch.shape
+                else: # bias
+                    c3 = data_torch.shape[0]
+                assert c3 % 3 == 0
+                c = c3 // 3
+                wq = data_torch[:c]
+                wk = data_torch[c: c * 2]
+                wv = data_torch[c * 2:]
+                return [
+                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.q_proj")), wq),
+                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.k_proj")), wk),
+                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.v_proj")), wv),
+                ]
+            return [(self.map_tensor_name(name), data_torch)]
+        return [] # skip other tensors
+
+
 @ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -3360,6 +3480,11 @@ class InternLM2Model(TextModel):
         head_dim = n_embd // num_heads
         num_groups = num_heads // q_per_kv
 
+        name = name.replace("language_model.", "") # InternVL
+        if name.startswith("mlp") or name.startswith("vision_model"):
+            # skip visual tensors
+            return []
+
         if bid is not None and f"model.layers.{bid}.attention.wqkv" in name:
             qkv = data_torch
 
@@ -3433,6 +3558,10 @@ class InternLM3Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
+        name = name.replace("language_model.", "") # InternVL
+        if name.startswith("mlp") or name.startswith("vision_model"):
+            # skip visual tensors
+            return []
         if name.endswith(("q_proj.weight", "q_proj.bias")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_head)
         if name.endswith(("k_proj.weight", "k_proj.bias")):
@@ -3907,11 +4036,11 @@ class Gemma3Model(TextModel):
 
 
 @ModelBase.register("Gemma3ForConditionalGeneration")
-class Gemma3VisionModel(VisionModel):
+class Gemma3VisionModel(MmprojModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.GEMMA3)
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GEMMA3)
         # default values below are taken from HF tranformers code
         self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
         self.gguf_writer.add_vision_use_gelu(True)
@@ -5669,11 +5798,20 @@ class GraniteModel(LlamaModel):
             logger.info("gguf: (granite) logits_scale = %s", logits_scale)
 
 
-@ModelBase.register("GraniteMoeForCausalLM")
+@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
 class GraniteMoeModel(GraniteModel):
     """Conversion for IBM's GraniteMoeForCausalLM"""
     model_arch = gguf.MODEL_ARCH.GRANITE_MOE
 
+    def set_gguf_parameters(self):
+        """GraniteMoeShared uses GraniteMoe parameters plus the following:
+        - shared_intermediate_size
+        """
+        super().set_gguf_parameters()
+        if shared_feed_forward_length := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_feed_forward_length)
+            logger.info("gguf: (granitemoeshared) shared_feed_forward_length = %s", shared_feed_forward_length)
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         """In modeling_granitemoe, the JetMoe implementation of parallel experts
         is used. This essentially merges w1 and w3 into a single tensor with 2x
@@ -5684,12 +5822,21 @@ class GraniteMoeModel(GraniteModel):
         if name.endswith("block_sparse_moe.input_linear.weight"):
             ffn_dim = self.hparams["intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
-            gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
+            gate, up = data_torch.split(ffn_dim, dim=-2)
             return [
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
             ]
 
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams["shared_intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            gate, up = data_torch.split(ffn_dim, dim=-2)
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+            ]
+
         return super().modify_tensors(data_torch, name, bid)
 
 
@@ -5840,6 +5987,52 @@ class ChameleonModel(TextModel):
         return data_torch
 
 
+@ModelBase.register("UltravoxModel")
+class UltravoxModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA # dummy
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        raise NotImplementedError("Ultravox does not have text decoder. Please use --mmproj argument")
+
+
+@ModelBase.register("UltravoxModel")
+class UltravoxAudioModel(MmprojModel):
+    has_vision_encoder = False # no vision encoder
+    has_audio_encoder = True
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["hidden_size"] = self.hparams["d_model"]
+        self.hparams["intermediate_size"] = self.hparams["encoder_ffn_dim"]
+        self.hparams["num_attention_heads"] = self.hparams["encoder_attention_heads"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.ULTRAVOX)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams["num_mel_bins"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
+        self.gguf_writer.add_audio_stack_factor(self.global_config["stack_factor"])
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, new_name, n_dims  # unused
+        if ".conv" in name and ".weight" in name:
+            return gguf.GGMLQuantizationType.F16
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # prevent clash naming with vision tensors
+        if name.startswith("multi_modal_projector"):
+            name = "audio." + name
+
+        if "conv1.bias" in name or "conv2.bias" in name:
+            # transpose conv1 and conv2 bias
+            data_torch = data_torch.unsqueeze(-1)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
 ###### CONVERSION LOGIC ######
 
 
@@ -6015,13 +6208,15 @@ def split_str_to_n_bytes(split_str: str) -> int:
 
 
 def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str:
+    # TODO @ngxson : this won't work correctly if the model has both audio & vision encoders
+    # maybe we should fallback to text model's arch in that case, since not many models have both
     text_config = hparams.get("text_config", {})
     vision_config = hparams.get("vision_config", {})
     arch = hparams["architectures"][0]
     # if "architectures" is found in the sub-config, use that instead
     if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
         arch = text_config["architectures"][0]
-    elif model_type == ModelType.VISION and vision_config.get("architectures") is not None:
+    elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
         arch = vision_config["architectures"][0]
     return arch
 
@@ -6084,7 +6279,7 @@ def main() -> None:
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
-        model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
+        model_type = ModelType.MMPROJ if args.mmproj else ModelType.TEXT
         hparams = ModelBase.load_hparams(dir_model)
         model_architecture = get_model_architecture(hparams, model_type)
         logger.info(f"Model architecture: {model_architecture}")

convert_hf_to_gguf_update.py

@@ -116,6 +116,7 @@ models = [
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
+    {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
 ]
 
 

docs/backend/CANN.md

@@ -56,60 +56,82 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
 
 ## Model Supports
 
-| Model Name                  | FP16  | Q8_0 | Q4_0 |
+| Model Name                  | FP16  | Q4_0 | Q8_0 |
 |:----------------------------|:-----:|:----:|:----:|
-| AquilaChat2-7B              |   √   |   √  |   √  |
-| Baichuan-7b                 |   √   |   √  |   √  |
-| Baichuan2-7B-Chat           |   √   |   √  |   √  |
-| bitnet_b1_58-large          |   √   |   √  |   √  |
-| bloom-560m                  |   √   |   x  |   √  |
-| bloomz-alpaca-560m          |   √   |   x  |   √  |
-| c4ai-command-r-35B-v01      |   x   |   x  |   x  |
-| chatglm3-6B                 |   x   |   x  |   x  |
-| chinese-alpaca-2-1.3b       |   √   |   √  |   √  |
-| CodeShell-7B                |   √   |   √  |   √  |
-| deepseek-ai_deepseek-coder-1.3B-base | x |   x  |   x  |
-| deepseek-ai_DeepSeek-V2-Lite | x   |   x  |   x   |
-| deepseek-coder-6.7B-instruct | x   |   x  |   x   |
-| DeepSeek-V2-Lite-64x1.5B    |   x   |   x  |   x  |
-| falcon-7b-instruct          |   √   |   √  |   √  |
-| flan-t5-large               |   √   |   √  |   √  |
-| gemma-2-9b-it               |   √   |   √  |   √  |
-| glm-4-9B                    |   x   |   x  |   x  |
-| gpt2                        |   √   |   √  |   √  |
-| Gpt2-163M                   |   √   |   √  |   √  |
-| granite-3B-code-instruct    |   √   |   √  |   √  |
+| Llama-2                     |   √   |   √  |   √  |
+| Llama-3                     |   √   |   √  |   √  |
+| Mistral-7B                  |   √   |   √  |   √  |
+| Mistral MOE                 |   √   |   √  |   √  |
+| DBRX                        |   -   |   -  |   -  |
+| Falcon                      |   √   |   √  |   √  |
+| Chinese LLaMA/Alpaca        |   √   |   √  |   √  |
+| Vigogne(French)             |   √   |   √  |   √  |
+| BERT                        |   x   |   x  |   x  |
+| Koala                       |   √   |   √  |   √  |
+| Baichuan                    |   √   |   √  |   √  |
+| Aquila 1 & 2                |   √   |   √  |   √  |
+| Starcoder models            |   √   |   √  |   √  |
+| Refact                      |   √   |   √  |   √  |
+| MPT                         |   √   |   √  |   √  |
+| Bloom                       |   √   |   √  |   √  |
+| Yi models                   |   √   |   √  |   √  |
+| stablelm models             |   √   |   √  |   √  |
+| DeepSeek models             |   x   |   x  |   x  |
+| Qwen models                 |   √   |   √  |   √  |
+| PLaMo-13B                   |   √   |   √  |   √  |
+| Phi models                  |   √   |   √  |   √  |
+| PhiMoE                      |   √   |   √  |   √  |
+| GPT-2                       |   √   |   √  |   √  |
+| Orion                       |   √   |   √  |   √  |
+| InternlLM2                  |   √   |   √  |   √  |
+| CodeShell                   |   √   |   √  |   √  |
+| Gemma                       |   √   |   √  |   √  |
+| Mamba                       |   √   |   √  |   √  |
+| Xverse                      |   √   |   √  |   √  |
+| command-r models            |   √   |   √  |   √  |
+| Grok-1                      |   -   |   -  |   -  |
+| SEA-LION                    |   √   |   √  |   √  |
 | GritLM-7B                   |   √   |   √  |   √  |
-| internlm2_5-7b-chat         |   √   |   √  |   √  |
-| koala-7B-HF                 |   √   |   √  |   √  |
-| Llama-2-7b-chat-hf          |   √   |   √  |   √  |
-| Llama-3-Smaug-8B            |   √   |   √  |   √  |
-| Llama2-Chinese-7b-Chat      |   √   |   √  |   √  |
-| Llama3-8B                   |   √   |   √  |   √  |
-| Llama3-8b-chinese           |   √   |   √  |   √  |
-| mamba-130m-hf               |   √   |   √  |   √  |
-| Mistral-7B-Instruct-v0.2    |   √   |   √  |   √  |
-| Mixtral-8x7B-Instruct-v0.1  |   x   |   √  |   √  |
-| mpt-7B                      |   √   |   √  |   √  |
-| OLMo-1B-hf                  |   √   |   √  |   √  |
-| OpenELM-3B-Instruct         |   √   |   √  |   √  |
-| Orion-14b-base              |   √   |   √  |   √  |
-| phi1                        |   x   |   x  |   x  |
-| phi2                        |   x   |   x  |   x  |
-| Phi-3-mini-4k-instruct      |   √   |   √  |   √  |
-| plamo-13b                   |   √   |   √  |   √  |
-| pythia-70M                  |   x   |   x  |   x  |
-| Qwen-7B                     |   √   |   √  |   √  |
-| Qwen2-1.5B-Instruct         |   √   |   x  |   √  |
-| Refact-1_6B-fim             |   √   |   √  |   √  |
-| SmolLM-135M                 |   √   |   √  |   √  |
-| stablelm-zephyr             |   x   |   x  |   x  |
-| stablelm-2-zephyr-1_6b      |   x   |   x  |   x  |
-| starcoderbase-1b            |   √   |   √  |   √  |
-| starcoder2-3b               |   √   |   √  |   √  |
-| vigogne-7b-chat             |   √   |   √  |   √  |
-| xverse-7b-chat              |   √   |   √  |   √  |
-| Yi-6b-Chat                  |   √   |   √  |   √  |
+| OLMo                        |   √   |   √  |   √  |
+| OLMo 2                      |   √   |   √  |   √  |
+| OLMoE                       |   √   |   √  |   √  |
+| Granite models              |   √   |   √  |   √  |
+| GPT-NeoX                    |   √   |   √  |   √  |
+| Pythia                      |   √   |   √  |   √  |
+| Snowflake-Arctic MoE        |   -   |   -  |   -  |
+| Smaug                       |   √   |   √  |   √  |
+| Poro 34B                    |   √   |   √  |   √  |
+| Bitnet b1.58 models         |   √   |   x  |   x  |
+| Flan-T5                     |   √   |   √  |   √  |
+| Open Elm models             |   x   |   √  |   √  |
+| chatGLM3-6B + ChatGLM4-9b +  GLMEdge-1.5b + GLMEdge-4b    |   √   |   √  |   √  |
+| GLM-4-0414                  |   √   |   √  |   √  |
+| SmolLM                      |   √   |   √  |   √  |
+| EXAONE-3.0-7.8B-Instruct    |   √   |   √  |   √  |
+| FalconMamba Models          |   √   |   √  |   √  |
+| Jais Models                 |   -   |   x  |   x  |
+| Bielik-11B-v2.3             |   √   |   √  |   √  |
+| RWKV-6                      |   -   |   √  |   √  |
+| QRWKV-6                     |   √   |   √  |   √  |
+| GigaChat-20B-A3B            |   x   |   x  |   x  |
+| Trillion-7B-preview         |   √   |   √  |   √  |
+| Ling models                 |   √   |   √  |   √  |
+
+
+**Multimodal**
+| Model Name                  | FP16  | Q4_0 | Q8_0 |
+|:----------------------------|:-----:|:----:|:----:|
+| LLaVA 1.5 models, LLaVA 1.6 models      |   x   |   x  |   x  |
+|  BakLLaVA                   |   √   |   √  |   √  |
+|  Obsidian                   |   √   |   -  |   -  |
+|  ShareGPT4V                 |   x   |   -  |   -  |
+|  MobileVLM 1.7B/3B models   |   -   |   -  |   -  |
+|  Yi-VL                      |   -   |   -  |   -  |
+|  Mini CPM                   |   √   |   √  |   √  |
+|  Moondream                  |   √   |   √  |   √  |
+|  Bunny                      |   √   |   -  |   -  |
+|  GLM-EDGE                   |   √   |   √  |   √  |
+|  Qwen2-VL                   |   √   |   √  |   √  |
 
 
 

docs/backend/SYCL.md

@@ -17,25 +17,25 @@
 
 **SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.
 
-**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
+**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to Intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
-- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
+- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
 ### Llama.cpp + SYCL
 
-The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.
+The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
+SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.
 
 ## Recommended Release
 
-The SYCL backend would be broken by some PRs due to no online CI.
-
-The following release is verified with good quality:
+The following releases are verified and recommended:
 
 |Commit ID|Tag|Release|Verified  Platform| Update date|
 |-|-|-|-|-|
+|24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8|b5377 |[llama-b5377-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b5377/llama-b5377-bin-win-sycl-x64.zip) |ArcB580/Linux/oneAPI 2025.1<br>LNL Arc GPU/Windows 11/oneAPI 2025.1.1|2025-05-15|
 |3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
 |fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
@@ -106,15 +106,14 @@ SYCL backend supports Intel GPU Family:
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
-| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake    |
-| Intel iGPU                    | Support | iGPU in 13700k,iGPU in 13400, i5-1250P, i7-1260P, i7-1165G7 |
+| Intel Arc Series              | Support | Arc 770, 730M, Arc A750, B580         |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
+| Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |
 
 *Notes:*
 
 - **Memory**
   - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
-
   - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
 
 - **Execution Unit (EU)**
@@ -138,9 +137,11 @@ Note: AMD GPU support is highly experimental and is incompatible with F16.
 Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
 
 ## Docker
-The docker build option is currently limited to *intel GPU* targets.
+
+The docker build option is currently limited to *Intel GPU* targets.
 
 ### Build image
+
 ```sh
 # Using FP16
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
@@ -148,9 +149,10 @@ docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f
 
 *Notes*:
 
-To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="GGML_SYCL_F16=ON"` argument from the previous command.
+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.
 
 ### Run container
 
@@ -250,7 +252,7 @@ sycl-ls
 
 - **Intel GPU**
 
-When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:
+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]
@@ -282,7 +284,7 @@ For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
 
 #### Intel GPU
 
-```
+```sh
 ./examples/sycl/build.sh
 ```
 
@@ -351,7 +353,7 @@ cmake --build build --config Release -j -v
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -398,11 +400,15 @@ Choose one of following methods to run.
 
 ```sh
 ./examples/sycl/run-llama2.sh 0
+# OR
+./examples/sycl/run-llama3.sh 0
 ```
 - Use multiple devices:
 
 ```sh
 ./examples/sycl/run-llama2.sh
+# OR
+./examples/sycl/run-llama3.sh
 ```
 
 2. Command line
@@ -425,13 +431,13 @@ Examples:
 - Use device 0:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm layer
 ```
 
 *Notes:*
@@ -452,7 +458,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 1. Install GPU driver
 
-Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+Intel GPU drivers instructions guide and download page can be found here: [Get Intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
 
 2. Install Visual Studio
 
@@ -629,7 +635,7 @@ Once it is completed, final results will be in **build/Release/bin**
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -648,7 +654,7 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow
 build\bin\llama-ls-sycl-device.exe
 ```
 
-This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *Intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -658,13 +664,14 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+
 #### Choose level-zero devices
 
 |Chosen Device ID|Setting|
 |-|-|
-|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|0|Default option. You may also want to `set ONEAPI_DEVICE_SELECTOR="level_zero:0"`|
 |1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
-|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"` or `set ONEAPI_DEVICE_SELECTOR="level_zero:*"`|
 
 #### Execute
 
@@ -673,7 +680,13 @@ Choose one of following methods to run.
 1. Script
 
 ```
-examples\sycl\win-run-llama2.bat
+examples\sycl\win-run-llama-2.bat
+```
+
+or
+
+```
+examples\sycl\win-run-llama-3.bat
 ```
 
 2. Command line
@@ -697,13 +710,13 @@ Examples:
 - Use device 0:
 
 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm layer
 ```
 
 
@@ -714,7 +727,9 @@ Note:
 ```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
+
 Or
+
 ```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
@@ -726,14 +741,17 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
 | GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
-| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
 | GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
+| GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
+1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
+
 #### Runtime
 
 | Name              | Value            | Function                                                                                                                  |
@@ -741,6 +759,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
+| GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
 
@@ -750,7 +769,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Q&A
 
-- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
+- Error:  `error while loading shared libraries: libsycl.so: cannot open shared object file: No such file or directory`.
 
   - Potential cause: Unavailable oneAPI installation or not set ENV variables.
   - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
@@ -779,18 +798,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   It's same for other projects including llama.cpp SYCL backend.
 
-- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+- `Native API failed. Native API returns: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)`, `ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 3503030272 Bytes of memory on device`, or `failed to allocate SYCL0 buffer`
 
-  Device Memory is not enough.
+  You are running out of Device Memory.
 
   |Reason|Solution|
   |-|-|
-  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
-  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
+  | The default context is too big. It leads to excessive memory usage.|Set `-c 8192` or a smaller value.|
+  | The model is too big and requires more memory than what is available.|Choose a smaller model or change to a smaller quantization, like Q5 -> Q4;<br>Alternatively, use more than one device to load model.|
 
 ### **GitHub contribution**:
-Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
+Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
 
 ## TODO
 
-- NA
+- Review ZES_ENABLE_SYSMAN: https://github.com/intel/compute-runtime/blob/master/programmers-guide/SYSMAN.md#support-and-limitations

docs/docker.md

@@ -22,6 +22,9 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 
@@ -104,7 +107,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
 
 The defaults are:
 
-- `MUSA_VERSION` set to `rc3.1.1`
+- `MUSA_VERSION` set to `rc4.0.1`
 
 The resulting images, are essentially the same as the non-MUSA images:
 

docs/multimodal.md

@@ -0,0 +1,92 @@
+# Multimodal
+
+llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature:
+- [llama-mtmd-cli](../tools/mtmd/README.md)
+- [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
+
+Currently, we support **image** and **audio** input. Audio is highly experimental and may have reduced quality.
+
+To enable it, you can use one of the 2 methods below:
+
+- Use `-hf` option with a supported model (see a list of pre-quantized model below)
+    - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
+    - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf`
+- Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively
+
+By default, multimodal projector will be offloaded to GPU. To disable this, add `--no-mmproj-offload`
+
+For example:
+
+```sh
+# simple usage with CLI
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+
+# simple usage with server
+llama-server -hf ggml-org/gemma-3-4b-it-GGUF
+
+# using local file
+llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.gguf
+
+# no GPU offload
+llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
+```
+
+## Pre-quantized models
+
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/ggml-org
+
+Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
+
+NOTE: some models may require large context window, for example: `-c 8192`
+
+**Vision models**:
+
+```sh
+# Gemma 3
+(tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
+(tool_name) -hf ggml-org/gemma-3-12b-it-GGUF
+(tool_name) -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+(tool_name) -hf ggml-org/SmolVLM-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+(tool_name) -hf ggml-org/pixtral-12b-GGUF
+
+# Qwen 2 VL
+(tool_name) -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
+
+# Qwen 2.5 VL
+(tool_name) -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+(tool_name) -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF
+
+# InternVL 2.5 and 3
+(tool_name) -hf ggml-org/InternVL2_5-1B-GGUF
+(tool_name) -hf ggml-org/InternVL2_5-4B-GGUF
+(tool_name) -hf ggml-org/InternVL3-1B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF
+
+# Llama 4 Scout
+(tool_name) -hf ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF
+```
+
+**Audio models**:
+
+```sh
+# Ultravox 0.5
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF
+```

examples/CMakeLists.txt

@@ -32,6 +32,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
     add_subdirectory(gen-docs)
+    add_subdirectory(training)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/lookahead/lookahead.cpp

@@ -50,8 +50,6 @@ int main(int argc, char ** argv) {
     const int N = 5;  // n-gram size
     const int G = 15; // max verification n-grams
 
-    const bool dump_kv_cache = params.dump_kv_cache;
-
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -152,9 +150,6 @@ int main(int argc, char ** argv) {
     // here we keep adding new n-grams as we go
     ngram_container ngrams_observed(llama_vocab_n_tokens(vocab), N, G);
 
-    // debug
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
-
     const auto t_dec_start = ggml_time_us();
 
     // sample first token
@@ -172,12 +167,6 @@ int main(int argc, char ** argv) {
     }
 
     while (true) {
-        // debug
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
         // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
         //
         // Example for W = 5, N = 4, G = 2:
@@ -473,8 +462,6 @@ int main(int argc, char ** argv) {
 
     common_sampler_free(smpl);
 
-    llama_kv_cache_view_free(&kvc_view);
-
     llama_batch_free(batch);
 
     llama_backend_free();

examples/lookup/lookup.cpp

@@ -24,8 +24,6 @@ int main(int argc, char ** argv){
     // max. number of additional tokens to draft if match is found
     const int n_draft = params.speculative.n_max;
 
-    const bool dump_kv_cache = params.dump_kv_cache;
-
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -110,18 +108,9 @@ int main(int argc, char ** argv){
 
     llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
 
-    // debug
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
-
     const auto t_dec_start = ggml_time_us();
 
     while (true) {
-        // debug
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
         // print current draft sequence
         LOG_DBG("drafted %s\n", string_from(ctx, draft).c_str());
 

examples/parallel/README.md

@@ -1,3 +1,14 @@
 # llama.cpp/example/parallel
 
 Simplified simulation of serving incoming requests in parallel
+
+## Example
+
+Generate 128 client requests (`-ns 128`), simulating 8 concurrent clients (`-np 8`). The system prompt is shared (`-pps`), meaning that it is computed once at the start. The client requests consist of 10 junk questions (`-j 10`) followed by the actual question.
+
+```bash
+llama-parallel -m model.gguf -np 8 -ns 128 --top-k 1 -pps --junk 10 -c 16384
+```
+
+> [!NOTE]
+> It's recommended to use base models with this example. Instruction tuned models might not be able to properly follow the custom chat template specified here, so the results might not be as expected.

examples/parallel/parallel.cpp

@@ -34,11 +34,61 @@ static std::string k_system =
 R"(Transcript of a never ending dialog, where the User interacts with an Assistant.
 The Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.
 
-User: Recommend a nice restaurant in the area.
-Assistant: I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
-User: Who is Richard Feynman?
-Assistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
-User:)";
+User:
+Recommend a nice restaurant in the area.
+Assistant:
+I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
+User:
+Who is Richard Feynman?
+Assistant:
+Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
+)";
+
+static std::vector<std::string> k_questions = {
+    "What is the tallest mountain in the world?",
+    "Who was the first person to win two Nobel Prizes?",
+    "Which country invented paper?",
+    "What organ is primarily responsible for pumping blood throughout the body?",
+    "Which planet is known for its prominent ring system?",
+    "Who directed the movie 'Inception'?",
+    "What is the freezing point of water in Fahrenheit?",
+    "Which animal is known to have the longest lifespan?",
+    "What language has the most native speakers worldwide?",
+    "What is the capital city of Canada?",
+    "Who is credited with inventing the World Wide Web?",
+    "Which metal is liquid at room temperature?",
+    "What is the term for an animal that eats both plants and meat?",
+    "Who painted 'The Starry Night'?",
+    "What gas do humans exhale that plants use for photosynthesis?",
+    "What year did World War II end?",
+    "Which continent has the most countries?",
+    "Who wrote the novel 'Frankenstein'?",
+    "What does DNA stand for?",
+    "What is the main ingredient in traditional Japanese miso soup?"
+};
+
+static std::vector<std::string> k_answers = {
+    "The tallest mountain in the world is Mount Everest.",
+    "Marie Curie was the first person to win two Nobel Prizes.",
+    "Paper was invented in China.",
+    "The heart is the organ responsible for pumping blood.",
+    "Saturn is known for its prominent ring system.",
+    "Christopher Nolan directed the movie 'Inception'.",
+    "The freezing point of water in Fahrenheit is 32°F.",
+    "The bowhead whale is known to have the longest lifespan among mammals.",
+    "Mandarin Chinese has the most native speakers in the world.",
+    "The capital city of Canada is Ottawa.",
+    "Tim Berners-Lee is credited with inventing the World Wide Web.",
+    "Mercury is the metal that is liquid at room temperature.",
+    "An animal that eats both plants and meat is called an omnivore.",
+    "'The Starry Night' was painted by Vincent van Gogh.",
+    "Humans exhale carbon dioxide, which plants use in photosynthesis.",
+    "World War II ended in 1945.",
+    "Africa is the continent with the most countries.",
+    "The novel 'Frankenstein' was written by Mary Shelley.",
+    "DNA stands for Deoxyribonucleic Acid.",
+    "The main ingredient in traditional Japanese miso soup is fermented soybean paste."
+};
 
 static std::vector<std::string> k_prompts = {
     "What is the meaning of life?",
@@ -49,7 +99,7 @@ static std::vector<std::string> k_prompts = {
     "What is the best way to learn a new language?",
     "How to get a job at Google?",
     "If you could have any superpower, what would it be?",
-    "I want to learn how to play the piano.",
+    "I want to learn how to play the piano. What would be the best way to do it?",
 };
 
 struct client {
@@ -68,6 +118,7 @@ struct client {
     int64_t t_start_prompt;
     int64_t t_start_gen;
 
+    int32_t n_past    = 0;
     int32_t n_prompt  = 0;
     int32_t n_decoded = 0;
     int32_t i_batch   = -1;
@@ -107,6 +158,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     params.n_predict = 128;
+    params.n_junk = 0;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
@@ -126,7 +178,11 @@ int main(int argc, char ** argv) {
     // insert new requests as soon as the previous one is done
     const bool cont_batching = params.cont_batching;
 
-    const bool dump_kv_cache = params.dump_kv_cache;
+    // is the system prompt shared in the cache
+    const bool is_sp_shared = params.is_pp_shared;
+
+    // extra text to insert in each client's prompt in order to make it larger
+    const int32_t n_junk = params.n_junk;
 
     // init llama.cpp
     llama_backend_init();
@@ -169,6 +225,7 @@ int main(int argc, char ** argv) {
     }
 
     std::vector<llama_token> tokens_system;
+
     tokens_system = common_tokenize(ctx, k_system, true);
     const int32_t n_tokens_system = tokens_system.size();
 
@@ -182,15 +239,13 @@ int main(int argc, char ** argv) {
     int32_t n_total_gen    = 0;
     int32_t n_cache_miss   = 0;
 
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients);
-
     const auto t_main_start = ggml_time_us();
 
     LOG_INF("%s: Simulating parallel requests from clients:\n", __func__);
     LOG_INF("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
     LOG_INF("\n");
 
-    {
+    if (is_sp_shared) {
         LOG_INF("%s: Evaluating the system prompt ...\n", __func__);
 
         for (int32_t i = 0; i < n_tokens_system; ++i) {
@@ -213,11 +268,6 @@ int main(int argc, char ** argv) {
     LOG_INF("Processing requests ...\n\n");
 
     while (true) {
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
         common_batch_clear(batch);
 
         // decode any currently ongoing sequences
@@ -228,7 +278,7 @@ int main(int argc, char ** argv) {
 
             client.i_batch = batch.n_tokens;
 
-            common_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
+            common_batch_add(batch, client.sampled, client.n_past++, { client.id + 1 }, true);
 
             client.n_decoded += 1;
         }
@@ -254,9 +304,23 @@ int main(int argc, char ** argv) {
                     client.t_start_gen    = 0;
 
                     client.input    = k_prompts[rand() % k_prompts.size()];
-                    client.prompt   = client.input + "\nAssistant:";
                     client.response = "";
 
+                    // construct the prompt:
+                    // [system prompt] + [junk] + [user prompt]
+                    client.n_past = 0;
+                    client.prompt = "";
+                    if (is_sp_shared) {
+                        client.n_past = n_tokens_system;
+                    } else {
+                        client.prompt += k_system;
+                    }
+                    for (int i = 0; i < n_junk; ++i) {
+                        const int r = rand() % k_questions.size();
+                        client.prompt += "User:\n" + k_questions[r] + "\nAssistant:\n " + k_answers[r] + "\n";
+                    }
+                    client.prompt += "User:\n" + client.input + "\nAssistant:\n";
+
                     common_sampler_reset(client.smpl);
 
                     // do not prepend BOS because we have a system prompt!
@@ -264,7 +328,7 @@ int main(int argc, char ** argv) {
                     tokens_prompt = common_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        common_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
+                        common_batch_add(batch, tokens_prompt[i], client.n_past++, { client.id + 1 }, false);
                     }
 
                     // extract the logits only for the last token
@@ -363,10 +427,9 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (llama_vocab_is_eog(vocab, id) ||
-                         (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
-                         client.response.find("User:") != std::string::npos ||
-                         client.response.find('\n') != std::string::npos)) {
+                    (llama_vocab_is_eog(vocab, id) ||
+                     (params.n_predict > 0 && client.n_decoded >= params.n_predict) ||
+                     client.response.find("User:") != std::string::npos)) {
                     // basic reverse prompt
                     const size_t pos = client.response.find("User:");
                     if (pos != std::string::npos) {

examples/retrieval/retrieval.cpp

@@ -81,14 +81,14 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
     }
 }
 
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+static void batch_encode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        LOG_ERR("%s : failed to decode\n", __func__);
+    if (llama_encode(ctx, batch) < 0) {
+        LOG_ERR("%s : failed to encode\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -233,7 +233,7 @@ int main(int argc, char ** argv) {
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
             float * out = emb + p * n_embd;
-            batch_decode(ctx, batch, out, s, n_embd);
+            batch_encode(ctx, batch, out, s, n_embd);
             common_batch_clear(batch);
             p += s;
             s = 0;
@@ -246,7 +246,7 @@ int main(int argc, char ** argv) {
 
     // final batch
     float * out = emb + p * n_embd;
-    batch_decode(ctx, batch, out, s, n_embd);
+    batch_encode(ctx, batch, out, s, n_embd);
 
     // save embeddings to chunks
     for (int i = 0; i < n_chunks; i++) {
@@ -267,7 +267,7 @@ int main(int argc, char ** argv) {
         batch_add_seq(query_batch, query_tokens, 0);
 
         std::vector<float> query_emb(n_embd, 0);
-        batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
+        batch_encode(ctx, query_batch, query_emb.data(), 1, n_embd);
 
         common_batch_clear(query_batch);
 

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_kv_self_used_cells(ctx) == 0;
+        const bool is_first = llama_kv_self_seq_pos_max(ctx, 0) == 0;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_kv_self_used_cells(ctx);
+            int n_ctx_used = llama_kv_self_seq_pos_max(ctx, 0);
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");

examples/simple/simple.cpp

@@ -84,13 +84,13 @@ int main(int argc, char ** argv) {
     model_params.n_gpu_layers = ngl;
 
     llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
+    const llama_vocab * vocab = llama_model_get_vocab(model);
     // tokenize the prompt
 
     // find the number of tokens in the prompt

examples/sycl/run-llama2.sh

@@ -12,16 +12,16 @@ source /opt/intel/oneapi/setvars.sh
 
 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
 MODEL_FILE=models/llama-2-7b.Q4_0.gguf
-NGL=33
-CONEXT=4096
+NGL=99
+CONTEXT=4096
 
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi

examples/sycl/run-llama3.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2025 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+# If you want more control, DPC++ Allows selecting a specific device through the
+# following environment variable
+#export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+source /opt/intel/oneapi/setvars.sh
+
+#export GGML_SYCL_DEBUG=1
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
+MODEL_FILE=models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
+NGL=99 # Layers offloaded to the GPU. If the device runs out of memory, reduce this value according to the model you are using.
+CONTEXT=4096
+
+if [ $# -gt 0 ]; then
+    GGML_SYCL_DEVICE=$1
+    echo "Using $GGML_SYCL_DEVICE as the main GPU"
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+else
+    #use multiple GPUs with same max compute units
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT}
+fi

examples/sycl/win-run-llama2.bat

@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 99 -s 0

examples/sycl/win-run-llama3.bat

@@ -0,0 +1,9 @@
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+
+.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -e -ngl 99

examples/training/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-finetune)
+add_executable(${TARGET} finetune.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/training/README.md

@@ -0,0 +1,17 @@
+# llama.cpp/examples/training
+
+This directory contains examples related to language model training using llama.cpp/GGML.
+So far finetuning is technically functional (for FP32 models and limited hardware setups) but the code is very much WIP.
+Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory.
+**For CPU training, compile llama.cpp without any additional backends such as CUDA.**
+**For CUDA training, use the maximum number of GPU layers.**
+
+Proof of concept:
+
+``` sh
+export model_name=llama_3.2-1b && export quantization=f32
+./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
+./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
+```
+
+The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.

examples/training/finetune.cpp

@@ -0,0 +1,96 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.escape = false;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+        return 1;
+    }
+
+    if (params.use_mmap) {
+        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
+        params.use_mmap = false;
+    }
+    if (params.cache_type_k != GGML_TYPE_F32) {
+        LOG_INF("%s: force changing k cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_k = GGML_TYPE_F32;
+    }
+    if (params.cache_type_v != GGML_TYPE_F32) {
+        LOG_INF("%s: force changing v cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_v = GGML_TYPE_F32;
+    }
+
+    common_init();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model and apply lora adapter, if any
+    common_init_result llama_init = common_init_from_params(params);
+    llama_model_ptr   & model = llama_init.model;
+    llama_context_ptr & ctx   = llama_init.context;
+
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n", __func__);
+        return 1;
+    }
+
+    // print system information
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+    }
+
+    constexpr float val_split = 0.05f;
+
+    std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
+    ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2);
+
+    struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
+    optimizer_params.adamw.alpha = 1e-7f; // learning rate
+
+    struct llama_opt_params lopt_params {
+        /*n_ctx_train     =*/ 0,
+        /*param_filter    =*/ llama_opt_param_filter_all,
+        /*param_filter_ud =*/ nullptr,
+        /*get_opt_pars    =*/ ggml_opt_get_constant_optimizer_params,
+        /*get_opt_pars_ud =*/ &optimizer_params,
+    };
+    llama_opt_init(ctx.get(), model.get(), lopt_params);
+
+    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_eval  = ggml_opt_result_init();
+
+    for (int epoch = 0; epoch < 2; ++epoch) {
+        llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
+            ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+        fprintf(stderr, "\n");
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_eval);
+    }
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_eval);
+
+    llama_model_save_to_file(model.get(), "finetuned-model.gguf");
+
+    llama_backend_free();
+
+    return 0;
+}

ggml/CMakeLists.txt

@@ -193,6 +193,7 @@ option(GGML_RPC                             "ggml: use RPC"
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 option(GGML_SYCL_GRAPH                      "ggml: enable graphs in the SYCL backend"         ON)
+option(GGML_SYCL_DNN                        "ggml: enable oneDNN in the SYCL backend"         ON)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
 set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING

ggml/include/ggml-backend.h

@@ -248,7 +248,7 @@ extern "C" {
         // preferrably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         // initialize buffers from a max size graph (optional)
         reserve_graph = build_graph(sched, max_batch_size);
@@ -289,7 +289,7 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph

ggml/include/ggml-opt.h

@@ -37,13 +37,16 @@ extern "C" {
     // ====== Dataset ======
 
     GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
-            int64_t ne_datapoint, // number of elements per datapoint
-            int64_t ne_label,     // number of elements per label
-            int64_t ndata,        // total number of datapoints/labels
-            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+            enum ggml_type type_data,    // the type for the internal data tensor
+            enum ggml_type type_label,   // the type for the internal labels tensor
+            int64_t        ne_datapoint, // number of elements per datapoint
+            int64_t        ne_label,     // number of elements per label
+            int64_t        ndata,        // total number of datapoints/labels
+            int64_t        ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
     GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
 
     // get underlying tensors that store the data
+    GGML_API int64_t              ggml_opt_dataset_ndata (ggml_opt_dataset_t dataset);
     GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
     GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
 
@@ -56,13 +59,19 @@ extern "C" {
             struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
             struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
             int64_t              ibatch);
+    GGML_API void ggml_opt_dataset_get_batch_host(
+            ggml_opt_dataset_t   dataset,
+            void               * data_batch,
+            size_t               nb_data_batch,
+            void               * labels_batch,
+            int64_t              ibatch);
 
     // ====== Model / Context ======
 
     enum ggml_opt_build_type {
-        GGML_OPT_BUILD_TYPE_FORWARD,
-        GGML_OPT_BUILD_TYPE_GRAD,
-        GGML_OPT_BUILD_TYPE_OPT,
+        GGML_OPT_BUILD_TYPE_FORWARD = 10,
+        GGML_OPT_BUILD_TYPE_GRAD    = 20,
+        GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
@@ -81,20 +90,22 @@ extern "C" {
     // userdata can be used to pass arbitrary data
     typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
 
-    // returns the default optimizer params (constant)
+    // returns the default optimizer params (constant, hard-coded values)
     // userdata is not used
     GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
 
+    // casts userdata to ggml_opt_optimizer_params and returns it
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata);
+
     // parameters for initializing a new optimization context
     struct ggml_opt_params {
         ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
 
-        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
-
-        // the forward graph is defined by inputs and outputs
-        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
-        struct ggml_tensor * inputs;
-        struct ggml_tensor * outputs;
+        // by default the forward graph needs to be reconstructed for each eval
+        // if ctx_compute, inputs, and outputs are set the graphs are instead allocated statically
+        struct ggml_context * ctx_compute;
+        struct ggml_tensor  * inputs;
+        struct ggml_tensor  * outputs;
 
         enum ggml_opt_loss_type  loss_type;
         enum ggml_opt_build_type build_type;
@@ -107,12 +118,9 @@ extern "C" {
 
     // get parameters for an optimization context with defaults set where possible
     // parameters for which no sensible defaults exist are supplied as arguments to this function
-    GGML_API ggml_opt_params ggml_opt_default_params(
-            ggml_backend_sched_t      backend_sched,
-            struct ggml_context     * ctx_compute,
-            struct ggml_tensor      * inputs,
-            struct ggml_tensor      * outputs,
-            enum ggml_opt_loss_type   loss_type);
+    GGML_API struct ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t    backend_sched,
+            enum ggml_opt_loss_type loss_type);
 
     GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
     GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
@@ -120,7 +128,10 @@ extern "C" {
     // set gradients to zero, initilize loss, and optionally reset the optimizer
     GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
 
+    GGML_API bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx); // whether the graphs are allocated_statically
+
     // get underlying tensors that store data
+    // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc
     GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
     GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
     GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
@@ -128,11 +139,12 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
     GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
 
+    // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
     // ====== Optimization Result ======
 
-    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API ggml_opt_result_t ggml_opt_result_init(void);
     GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
     GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
 
@@ -144,11 +156,20 @@ extern "C" {
 
     // ====== Computation ======
 
-    // do forward pass, increment result if not NULL
-    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // if not using static graphs, this function must be called prior to ggml_opt_alloc
+    GGML_API void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs);
 
-    // do forward pass, increment result if not NULL, do backward pass
-    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // allocate the next graph for evaluation, either forward or forward + backward
+    // must be called exactly once prior to calling ggml_opt_eval
+    GGML_API void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward);
+
+    // do forward pass, increment result if not NULL, do backward pass if allocated
+    GGML_API void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
 
     // ############################################################################
     // ## The high-level functions start here. They do not depend on any private ##
@@ -200,9 +221,9 @@ extern "C" {
     // fit model defined by inputs and outputs to dataset
     GGML_API void ggml_opt_fit(
             ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
-            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
-            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
-            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            struct ggml_context           * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            struct ggml_tensor            * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)

ggml/include/ggml.h

@@ -536,6 +536,7 @@ extern "C" {
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
+        GGML_UNARY_OP_GELU_ERF,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -768,7 +769,7 @@ extern "C" {
     // Tensor flags
     GGML_API void ggml_set_input(struct ggml_tensor * tensor);
     GGML_API void ggml_set_output(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_tensor * tensor);
     GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
 
     //
@@ -938,7 +939,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
 
     // concat a and b along dim
     // used in stable-diffusion
@@ -1024,6 +1025,16 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // GELU using erf (error function) when possible
+    // some backends may fallback to approximation based on Abramowitz and Stegun formula
+    GGML_API struct ggml_tensor * ggml_gelu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_erf_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_gelu_quick(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -2049,15 +2060,14 @@ extern "C" {
 
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
-        struct ggml_context * ctx_compute, // context for gradient computation
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
+        struct ggml_context *  ctx,        // context for gradient computation
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs);
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
-    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads);
     GGML_API void                 ggml_graph_cpy       (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset     (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
     GGML_API void                 ggml_graph_clear     (struct ggml_cgraph * cgraph);

ggml/src/ggml-backend.cpp

@@ -674,6 +674,8 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
+    bool op_offload;
+
     int debug;
 };
 
@@ -766,7 +768,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
@@ -1109,7 +1111,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            assert(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
@@ -1452,7 +1454,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
         ggml_backend_buffer_type_t * bufts,
         int n_backends,
         size_t graph_size,
-        bool parallel) {
+        bool parallel,
+        bool op_offload) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1497,6 +1500,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
     }
 
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+    sched->op_offload = op_offload;
 
     ggml_backend_sched_reset(sched);
 

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -65,6 +65,7 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
+#include <aclnnop/aclnn_grouped_matmul_v2.h>
 #include <float.h>
 
 #include <cmath>
@@ -2587,3 +2588,276 @@ void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst){
 
     ggml_cann_release_resources(ctx, acl_src, acl_dst, alpha);
 }
+
+/**
+ * @brief Performs expert-specific matrix multiplication (MoE) with
+ * floating-point precision using the CANN backend.
+ *
+ * This function executes a matrix multiplication operation tailored for
+ * Mixture of Experts (MoE) models, where the input tensor is multiplied
+ * with expert-specific weight matrices. It uses the CANN backend for
+ * efficient computation and stores the result in the destination tensor `dst`.
+ * The operation may leverage identity-based optimizations or routing masks
+ * as part of sparse expert selection.
+ *
+ * @param ctx The context for executing CANN backend operations.
+ * @param dst The destination tensor where the MoE multiplication result
+ * will be stored.
+ *
+ * @note This function assumes floating-point data types and is designed for
+ * MoE architectures, possibly involving sparse expert routing.
+ */
+static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    //dst   [M, K, N, 1]
+    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
+    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
+    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // copy index from npu to cpu
+    int64_t n_as = ne02; // A
+    int64_t n_ids = ids->ne[0]; // K
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
+        ACL_MEMCPY_DEVICE_TO_HOST);
+    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+
+    char * src0_original = (char *) src0->data;
+    char * src1_original = (char *) src1->data;
+    char * dst_original  = (char *)  dst->data;
+    size_t ori_src0_nb[4] = {nb00, nb01, nb02, nb03};
+
+    // src0 is F16, src1 is F32, dst is F32
+    ggml_cann_pool_alloc src0_cast_allocator;
+    if (src0->type == GGML_TYPE_F16) {
+        src0_cast_allocator.alloc(ctx.pool(), sizeof(float) * ggml_nelements(src0));
+        void* src0_cast_buf = src0_cast_allocator.get();
+
+        size_t cast_nb[GGML_MAX_DIMS];
+        cast_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            cast_nb[i] = cast_nb[i - 1] * src0->ne[i - 1];
+        }
+
+        aclTensor* acl_src0_f16 = ggml_cann_create_tensor(src0);
+        aclTensor* acl_cast = ggml_cann_create_tensor(src0_cast_buf,
+            ACL_FLOAT, sizeof(float), src0->ne, cast_nb, 4);
+        GGML_CANN_CALL_ACLNN_OP(ctx, Cast, acl_src0_f16, ACL_FLOAT, acl_cast);
+        ggml_cann_release_resources(ctx, acl_cast, acl_src0_f16);
+
+        src0_original = (char *) src0_cast_buf;
+        memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
+    }
+
+    std::vector<aclTensor*> src0_tensor_vec;
+    std::vector<aclTensor*> src1_tensor_vec;
+    std::vector<aclTensor*> dst_tensor_vec;
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // src0_row [M, D] -> weight && permute
+            int64_t src0_ne[2] = {ne01, ne00};
+            size_t src0_nb[2] = {ori_src0_nb[1], ori_src0_nb[0]};
+            // src1_row [D, 1] -> input
+            int64_t src1_ne[2] = {ne10, 1};
+            size_t src1_nb[2] = {nb10, nb11};
+            // dst_row [M, 1] -> out
+            int64_t dst_ne[2] = {ne0, 1};
+            size_t dst_nb[2] = {nb0, nb1};
+
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            aclTensor* acl_src0 = ggml_cann_create_tensor(src0_tmp_ptr,
+                ACL_FLOAT, sizeof(float),
+                src0_ne, src0_nb, 2);
+            aclTensor* acl_src1 = ggml_cann_create_tensor(src1_tmp_ptr,
+                ACL_FLOAT, sizeof(float),
+                src1_ne, src1_nb, 2);
+            aclTensor* acl_dst = ggml_cann_create_tensor(dst_tmp_ptr,
+                ACL_FLOAT, sizeof(float),
+                dst_ne, dst_nb, 2);
+
+            src0_tensor_vec.push_back(acl_src0);
+            src1_tensor_vec.push_back(acl_src1);
+            dst_tensor_vec.push_back(acl_dst);
+        }
+    }
+
+    size_t GROUP_SIZE = 128;
+    // GroupedMatmulV2 required tensor_list.size < 128
+    for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
+        // split and call GroupedMatmulV2
+        size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
+        std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
+        std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
+        std::vector<aclTensor*> dst_tensor_vec_split(dst_tensor_vec.begin() + i, dst_tensor_vec.begin() + end);
+
+        aclTensorList* src0_tensor_list = aclCreateTensorList(src0_tensor_vec_split.data(), src0_tensor_vec_split.size());
+        aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
+        aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+            nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
+
+        ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
+    }
+    return;
+}
+
+/**
+ * @brief Performs expert-specific matrix multiplication (MoE) with
+ * quantized precision using the CANN backend.
+ *
+ * This function executes a matrix multiplication operation tailored for
+ * Mixture of Experts (MoE) models, where the input tensor is multiplied
+ * with expert-specific quantized weight matrices. It leverages the CANN
+ * backend to perform efficient low-precision computations and stores the
+ * quantized result in the destination tensor `dst`.
+ *
+ * Quantization techniques reduce memory footprint and improve performance
+ * by using lower-bit representations (e.g., int8) instead of floating-point.
+ * This function is designed to work with such formats and may incorporate
+ * optimizations like identity-based fast paths or routing masks for sparse
+ * expert selection.
+ *
+ * @param ctx The context for executing CANN backend operations.
+ * @param dst The destination tensor where the quantized MoE multiplication result
+ * will be stored.
+ *
+ * @note This function assumes quantized data types and is designed for
+ * MoE architectures with potential sparse expert routing.
+ */
+static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    // TODO: Use aclnnGroupedMatMul
+    //dst   [M, K, N, 1]
+    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
+    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
+    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // copy index from npu to cpu
+    int64_t n_as = ne02; // A
+    int64_t n_ids = ids->ne[0]; // K
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
+        ACL_MEMCPY_DEVICE_TO_HOST);
+    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+
+    char * src0_original = (char *) src0->data;
+    char * src1_original = (char *) src1->data;
+    char * dst_original  = (char *)  dst->data;
+
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    const enum ggml_type type = dst->src[0]->type;
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    } else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    } else {
+        GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = weight_elem_size;
+    src0_row.nb[1] = weight_elem_size * ne00;
+    src0_row.nb[2] = weight_elem_size * ne00;
+    src0_row.nb[3] = weight_elem_size * ne00;
+    size_t weight_stride = ne00 * ne01 * weight_elem_size;
+    size_t weight_size = weight_stride * ne02 * ne03;
+
+    // scale [D, M, 1, 1] -> scale && permute
+    size_t scale_elem_size = sizeof(uint16_t);
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
+
+    // src1_row [D, 1, 1, 1] -> input
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    // dst_row [M, 1, 1, 1] -> out
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    //create weight for one row
+    ggml_cann_pool_alloc weight_allocator(ctx.pool());
+    void* weight_buffer = weight_allocator.alloc(nb02);
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*weight_stride;
+            void* scale_tmp_ptr = src0_original + weight_size + i02*scale_stride;
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            // mem cpy
+            ggml_cann_async_memcpy(ctx, weight_buffer, src0_tmp_ptr, weight_stride,
+                ACL_MEMCPY_DEVICE_TO_DEVICE);
+            void* scale_buffer = (char*)weight_buffer + weight_stride;
+            ggml_cann_async_memcpy(ctx, scale_buffer, scale_tmp_ptr, scale_stride,
+                ACL_MEMCPY_DEVICE_TO_DEVICE);
+
+            src0_row.data = weight_buffer;
+            src1_row.data = src1_tmp_ptr;
+            dst_row.data = dst_tmp_ptr;
+            dst_row.src[0] = &src0_row;
+            dst_row.src[1] = &src1_row;
+
+            ggml_cann_mul_mat(ctx, &dst_row);
+        }
+    }
+    return;
+}
+
+void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    const enum ggml_type type = dst->src[0]->type;
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+            ggml_cann_mul_mat_id_fp(ctx, dst);
+            break;
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            ggml_cann_mul_mat_id_quant(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupported type for mul_mat_id");
+            break;
+    }
+}

ggml/src/ggml-cann/aclnn_ops.h

@@ -978,6 +978,33 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
     }
 }
 
+/**
+ * @brief   Performs sparse expert-based matrix multiplication using the CANN backend.
+ *
+ * @details This function implements a MoE-style batched matrix multiplication, where each input token
+ *          is routed to one or more experts, and each expert corresponds to a specific [D, M] weight matrix
+ *          in the source tensor `src0`. The routing indices are provided via the `ids` tensor.
+ *
+ *          For each token (from `src1`), the function selects the corresponding expert(s) as specified by `ids`,
+ *          performs the matrix multiplication with the selected expert's weight submatrix (from `src0`),
+ *          and stores the results in `dst`. This operation is optimized and executed on the CANN backend.
+ *
+ *          Dimensions:
+ *              - src0: [D, M, A, 1], where A is the number of experts
+ *              - src1: [D, B, N, 1], where N is batch size and B is the slot count per sample
+ *              - ids : [K, N],       where K is the number of experts each token is routed to
+ *              - dst : [M, K, N, 1], output tensor storing the result of expert × token multiplication
+ *
+ *          The function handles two main modes:
+ *              - If `ne12 == 1`, a simpler per-token loop is used.
+ *              - TODO: If `ne12 > 1`, grouped multiplication and memory copying is used for efficiency.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the expert-weighted token outputs are stored.
+ *            Expected to be of shape [M, K, N, 1].
+ */
+void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.

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

@@ -1672,7 +1672,8 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_mul_mat(ctx, dst);
             break;
         case GGML_OP_MUL_MAT_ID:
-            return false;
+            ggml_cann_mul_mat_id(ctx, dst);
+            break;
         case GGML_OP_SCALE:
             ggml_cann_scale(ctx, dst);
             break;
@@ -2030,7 +2031,22 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
         }
         case GGML_OP_MUL_MAT_ID:
-            return false;
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
+                    return true;
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
+                    // only support contiguous for quantized types.
+                    return ggml_is_contiguous(op->src[0]) &&
+                            ggml_is_contiguous(op->src[1]);
+                default:
+                    return false;
+            }
         // embedding
         case GGML_OP_GET_ROWS: {
             switch (op->src[0]->type) {

ggml/src/ggml-cpu/CMakeLists.txt

@@ -385,9 +385,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.5.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "ea22e1aefb800e9bc8c74d91633cc58e")
+        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -428,6 +428,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
 
         set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -438,17 +439,19 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
 
-        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS})
+        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
 
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
+        list(APPEND GGML_KLEIDIAI_SOURCES
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
 
         if (NOT DOTPROD_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
         endif()
 
         if (NOT I8MM_ENABLED MATCHES -1)
@@ -456,9 +459,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (NOT SME_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c)
-            set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2")
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c)
+            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
         endif()
 
         set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")

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

@@ -8519,7 +8519,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+#ifdef __ARM_FEATURE_MATMUL_INT8
+    assert((nrc == 2) || (nrc == 1));
+#else
     assert(nrc == 1);
+#endif
     UNUSED(nrc);
     UNUSED(bx);
     UNUSED(by);
@@ -8530,6 +8534,197 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     const int nb = n / QK_K;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q6_K * GGML_RESTRICT x0 = x;
+        const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
+        const block_q8_K * GGML_RESTRICT y0 = y;
+        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
+
+        float32x4_t vfsum = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
+            const uint8_t * GGML_RESTRICT ql0 = x0->ql;
+            const uint8_t * GGML_RESTRICT ql1 = x1->ql;
+            const uint8_t * GGML_RESTRICT qh0 = x0->qh;
+            const uint8_t * GGML_RESTRICT qh1 = x1->qh;
+            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
+            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
+
+            const uint8x16_t mone = vdupq_n_u8(0x30);
+            const uint8x16_t  m4b = vdupq_n_u8(0x0f);
+
+            int32x4_t visum = vdupq_n_s32(0);
+
+            // process 8 blocks per iteration, totally 16 blocks
+            for (int j = 0; j < 2; ++j, qh0 += 32, ql0 += 64, qh1 += 32, ql1 += 64) {
+                int8x16_t vx0[8], vx1[8];
+
+                // de-quantize vx0[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh0);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql0);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx0[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx0[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx0[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx0[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx0[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx0[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // de-quantize vx1[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh1);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql1);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx1[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx1[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx1[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx1[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx1[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx1[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // process 16 elements (one block with same scale) per iteration
+                // - vx = concat(ql, qh) - 32
+                // - r1,r2,r3,r4 = smmla(vx, vy)
+                for (int k = 0; k < 8; ++k) {
+                    const int blk = j * 8 + k;
+
+                    const int8x16_t vy0 = vld1q_s8(qy0);
+                    const int8x16_t vy1 = vld1q_s8(qy1);
+                    qy0 += 16;
+                    qy1 += 16;
+
+                    const int32x4_t block_scale = {
+                        x0->scales[blk],
+                        x0->scales[blk],
+                        x1->scales[blk],
+                        x1->scales[blk],
+                    };
+
+                    // calculate four results at once with outer product
+                    const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    int32x4_t vr = vdupq_n_s32(0);
+                    vr = vmmlaq_s32(vr, vx_l, vy_l);
+                    vr = vmmlaq_s32(vr, vx_h, vy_h);
+
+                    // apply block scale, will NOT overflow
+                    // block_scale * sum_256(int6*int8) <= 2^(8+8+6+8) = 30 bits
+                    visum = vmlaq_s32(visum, vr, block_scale);
+                }
+            }
+
+            // adjust bias, apply superblock scale
+            {
+                int32_t bias[4];
+#ifdef __ARM_FEATURE_SVE
+                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
+                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
+                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
+                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
+                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
+                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
+                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
+                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
+                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
+                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
+                const svint64_t zero = svdup_n_s64(0);
+                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
+                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
+                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
+                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
+#else
+                // NEON doesn't support int16 dot product, fallback to separated mul and add
+                const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
+                const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
+
+                int8x16_t scales_s8 = vld1q_s8(x0->scales);
+                const int16x8x2_t q6scales0 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+                scales_s8 = vld1q_s8(x1->scales);
+                const int16x8x2_t q6scales1 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+
+                int32x4_t prod;
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[0] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[1] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[2] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[3] = vaddvq_s32(prod);
+
+#endif
+                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
+
+                const float32x4_t superblock_scale = {
+                    GGML_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                };
+
+                visum = vsubq_s32(visum, vibias);
+                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
+            }
+        }
+
+        // vfsum = ABCD -> ACBD
+        // AC -> s, BD -> (s+bs)
+        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
+        vst1_f32(s,      vget_low_f32 (vfsum));
+        vst1_f32(s + bs, vget_high_f32(vfsum));
+
+        return;
+    }
+#endif
+
 #ifdef __ARM_FEATURE_SVE
     const int vector_length = ggml_cpu_get_sve_cnt()*8;
     float sum = 0;

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

@@ -282,7 +282,11 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q6_K,
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
         .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_IQ2_XXS] = {
         .from_float               = NULL,
@@ -2198,6 +2202,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     } break;
 
                 case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
                     {

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

@@ -4,16 +4,22 @@
 
 // KleidiAI micro-kernels
 #include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
-#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
-#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
-#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
-#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
+#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
+
+#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
+
+#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
+#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
+#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
+
 #include "kai_common.h"
 
 #include "kernels.h"
@@ -61,6 +67,53 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+    {
+        /* SME GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .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,
+        },
+        /* SME GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .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,
+        },
+        /* .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,
+        },
+        /* .rhs_info = */ {
+            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .pack_func   = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_F16,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__APPLE__)
@@ -105,6 +158,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__ARM_FEATURE_MATMUL_INT8)
@@ -148,6 +204,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #else
@@ -192,6 +251,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__ARM_FEATURE_DOTPROD)
@@ -235,12 +297,33 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #endif
 };
 
-ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) {
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
+    ggml_kleidiai_kernels * kernel = nullptr;
+
+    if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
+        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 &&
+                gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
+                gemm_gemv_kernels[i].op_type  == tensor->type) {
+                kernel = &gemm_gemv_kernels[i];
+                break;
+            }
+        }
+    }
+
+    return kernel;
+}
+
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {

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

@@ -4,6 +4,10 @@
 
 #pragma once
 
+#include <functional>
+#include <variant>
+#include "ggml.h"
+
 enum cpu_feature {
     CPU_FEATURE_NONE    = 0,
     CPU_FEATURE_DOTPROD = 1,
@@ -26,26 +30,53 @@ struct kernel_info {
     size_t (*get_nr)(void);
     size_t (*get_kr)(void);
     size_t (*get_sr)(void);
-    size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl);
-    size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl);
+    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);
-    void (*run_kernel)(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::variant<
+        std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
+            float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
+        std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
+            size_t dst_stride_col, float clamp_min, float clamp_max)>
+    > run_kernel;
 };
 
 struct lhs_packing_info {
     size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
-    size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-    size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-    void (*pack_func)(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::variant<
+        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
+        std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
+    > get_packed_offset;
+    std::variant<
+        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
+        std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
+    > packed_size;
+    std::variant<
+        std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
+            size_t lhs_stride, void* lhs_packed)>,
+        std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
+        void* lhs_packed)>
+    > pack_func;
 };
 
 struct rhs_packing_info {
-    size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
-    void (*pack_func)(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::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;
+    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;
 };
 
 struct ggml_kleidiai_kernels {
@@ -55,6 +86,10 @@ struct ggml_kleidiai_kernels {
     rhs_packing_info rhs_info;
 
     cpu_feature required_cpu;
+    ggml_type lhs_type;
+    ggml_type rhs_type;
+    ggml_type op_type;
 };
 
-ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);

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

@@ -3,7 +3,9 @@
 //
 #include <arm_neon.h>
 #include <assert.h>
+#include <atomic>
 #include <cfloat>
+#include <stdexcept>
 #include <stdint.h>
 #include <string.h>
 #if defined(__linux__)
@@ -34,8 +36,9 @@
 #include "ggml-common.h"
 
 struct ggml_kleidiai_context {
+    cpu_feature features;
     ggml_kleidiai_kernels * kernels;
-} static ctx = { NULL };
+} static ctx = { CPU_FEATURE_NONE, NULL };
 
 static void init_kleidiai_context(void) {
 
@@ -47,18 +50,18 @@ static void init_kleidiai_context(void) {
         const char *env_var = getenv("GGML_KLEIDIAI_SME");
         int sme_enabled = 0;
 
-        cpu_feature features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
-                                (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
-                                (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
+        ctx.features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
+                        (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
+                        (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
 
         if (env_var) {
             sme_enabled = atoi(env_var);
         }
 
         if (sme_enabled != 0) {
-            features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
+            ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
         }
-        ctx.kernels = ggml_kleidiai_select_kernels(features);
+        ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
     }
     ggml_critical_section_end();
 }
@@ -68,95 +71,275 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
     return tensor->ne[dim];
 }
 
+template<typename Ret, typename Variant, typename... Args>
+static Ret variant_call(const Variant & var, Args&&... args) {
+    return std::visit([&](auto&& func) -> Ret {
+        if constexpr (std::is_invocable_r_v<Ret, decltype(func), Args...>) {
+            return func(std::forward<Args>(args)...);
+        } else {
+            throw std::runtime_error("Invalid function type in variant_call");
+        }
+    }, var);
+}
+
 namespace ggml::cpu::kleidiai {
+
+static size_t round_down(size_t x, size_t y) {
+    return y == 0 ? x : x - (x % y);
+}
+
+static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint16_t * src, size_t rhs_stride) {
+    size_t src_stride = rhs_stride / sizeof(uint16_t);
+    size_t dst_stride = n;
+
+    for (size_t k_idx = 0; k_idx < k; ++k_idx) {
+        for (size_t n_idx = 0; n_idx < n; ++n_idx) {
+            uint16_t v = *(src + k_idx + n_idx * src_stride);
+            *(dst + n_idx + k_idx * dst_stride) = kai_cast_f32_f16(v);
+        }
+    }
+}
+
 class tensor_traits : public ggml::cpu::tensor_traits {
     bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
-        GGML_ASSERT(ctx.kernels);
-        kernel_info * kernel = op->src[1]->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
+        GGML_ASSERT(kernels);
+        kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
 
         size_t k = op->src[0]->ne[0];
+        size_t n = op->src[0]->ne[1];
         size_t m = op->src[1]->ne[1];
 
         size_t mr = kernel->get_mr();
         size_t kr = kernel->get_kr();
         size_t sr = kernel->get_sr();
 
-        size = ctx.kernels->lhs_info.packed_size(m, k, QK4_0, mr, kr, sr);
+        if (kernels->rhs_type == GGML_TYPE_Q4_0) {
+            size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, QK4_0, mr, kr, sr);
+        } else if (kernels->rhs_type == GGML_TYPE_F16) {
+            size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr) +
+                   variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
+                   k * n * sizeof(float) + n * sizeof(float);
+        } else {
+            GGML_ASSERT(false);
+        }
 
         return true;
     }
 
+
     bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
         if (dst->op == GGML_OP_MUL_MAT) {
-            const ggml_tensor * src0 = dst->src[0];
-            const ggml_tensor * src1 = dst->src[1];
+            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+                return compute_forward_q4_0(params, dst);
+            } else if (dst->src[0]->type == GGML_TYPE_F16) {
+                return compute_forward_kv_cache(params, dst);
+            }
+        }
+        return false;
+    }
 
-            GGML_TENSOR_BINARY_OP_LOCALS
+    bool compute_forward_kv_cache(ggml_compute_params * params, struct ggml_tensor * dst) {
+        static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
 
-            GGML_ASSERT(ctx.kernels);
-            kernel_info * kernel = src1->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
-            lhs_packing_info * lhs_info = &ctx.kernels->lhs_info;
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
 
-            GGML_ASSERT(kernel);
+        GGML_TENSOR_BINARY_OP_LOCALS
 
-            const int ith = params->ith;
-            const int nth = params->nth;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        GGML_ASSERT(kernels);
 
-            const size_t k = ne00;
-            const size_t m = ne11;
-            const size_t n = ne01;
+        kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
+        GGML_ASSERT(kernel);
 
-            const size_t n_step = kernel->get_n_step();
-            const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
-            const size_t n_start = ith * num_n_per_thread;
+        const int nth = params->nth;
+        const int ith = params->ith;
 
-            size_t n_to_process = num_n_per_thread;
-            if ((n_start + n_to_process) > n) {
-                n_to_process = n - n_start;
+        const int64_t lhs_batch_size0 = ne12;
+        const int64_t rhs_batch_size0 = ne02;
+        const int64_t batch_size      = rhs_batch_size0;
+
+        const int64_t r = lhs_batch_size0 / rhs_batch_size0;
+
+        const int64_t m = ne11 * r;
+        const int64_t n = ne01;
+        const int64_t k = ne00;
+
+        const size_t lhs_stride = src1->nb[1];
+        const size_t rhs_stride = src0->nb[1];
+        const size_t dst_stride = dst->nb[1];
+
+        const int64_t mr = static_cast<int64_t>(kernel->get_mr());
+        const int64_t nr = static_cast<int64_t>(kernel->get_nr());
+        const int64_t kr = static_cast<int64_t>(kernel->get_kr());
+        const int64_t sr = static_cast<int64_t>(kernel->get_sr());
+
+        const size_t lhs_packed_size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr);
+        const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, n, k);
+        const size_t kxn_size        = k * n * sizeof(float);
+        const size_t bias_size       = n * sizeof(float);
+
+        const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
+        GGML_ASSERT(wsize_required <= params->wsize);
+
+        uint8_t * lhs_packed = static_cast<uint8_t *>(params->wdata);
+        uint8_t * rhs_packed = lhs_packed + lhs_packed_size;
+        uint8_t * rhs_kxn    = rhs_packed + rhs_packed_size;
+        uint8_t * bias       = rhs_kxn + kxn_size;
+
+        for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
+            const uint8_t * lhs_batch = static_cast<const uint8_t *>(src1->data) + batch_idx * m * lhs_stride;
+            const uint8_t * rhs_batch = static_cast<const uint8_t *>(src0->data) + batch_idx * n * rhs_stride;
+            uint8_t * dst_batch       = static_cast<uint8_t *>(dst->data) + batch_idx * m * dst_stride;
+
+            // LHS packing
+            {
+                const int64_t m_roundup_mr = kai_roundup(m, mr);
+                const int64_t num_threads  = KAI_MIN(m_roundup_mr / mr, nth);
+
+                if (ith < num_threads) {
+                    const int64_t num_m_per_thread0   = round_down(m_roundup_mr / num_threads, mr);
+                    const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
+
+                    const int64_t m_start          = ith * num_m_per_thread0;
+                    const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
+
+                    const size_t lhs_offset        = variant_call<size_t>(kernels->gemm.get_lhs_offset, m_start, lhs_stride);
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernels->lhs_info.get_packed_offset, m_start, k, mr, kr, sr);
+
+                    const void * src_ptr = static_cast<const uint8_t *>(lhs_batch) + lhs_offset;
+                    void * dst_ptr       = static_cast<uint8_t *>(lhs_packed) + lhs_packed_offset;
+
+                    variant_call<void>(kernels->lhs_info.pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
+                }
             }
 
-            const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
-            uint8_t * lhs_packed       = (uint8_t*)params->wdata;
-            const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+            // RHS packing
+            if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) {
+                // First thread to reach this point handles RHS packing
+                memset(bias, 0, n * sizeof(float));
+                transpose_f32kxn_f16nxk(n, k, reinterpret_cast<float *>(rhs_kxn),
+                                        reinterpret_cast<const uint16_t *>(rhs_batch), rhs_stride);
 
-            size_t mr = kernel->get_mr();
-            size_t kr = kernel->get_kr();
-            size_t sr = kernel->get_sr();
-
-            // Calculate number of columns to be processed per thread
-            const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
-            const size_t m_start = ith * num_m_per_thread;
-            size_t m_to_process = num_m_per_thread;
-            if ((m_start + m_to_process) > m) {
-                m_to_process = m - m_start;
-            }
-
-            if(m_start < m) {
-                // Transform LHS
-                const size_t src_stride        = src1->nb[1];
-                const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
-                const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr);
-                void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
-
-                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+                variant_call<void>(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float),
+                             rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
             }
 
             ggml_barrier(params->threadpool);
 
-            // Perform the operation
-            const size_t dst_stride        = dst->nb[1];
-            const size_t lhs_packed_offset = lhs_info->get_packed_offset(0, k, QK4_0, mr, kr, sr);
-            const size_t rhs_packed_offset = kernel->get_rhs_packed_offset(n_start, k, QK4_0);
-            const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
-            const void * rhs_ptr           = static_cast<const void *>(rhs_packed + rhs_packed_offset);
-            const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
-            float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
+            first_to_arrive.clear(std::memory_order_release);
 
-            kernel->run_kernel(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr,
-                               dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
-            return true;
+            // Perform the matmul
+            {
+                const int64_t m_to_process = m;
+                const int64_t m_start      = 0;
+
+                const int64_t n_step      = static_cast<int64_t>(kernel->get_n_step());
+                const int64_t num_threads = KAI_MIN(n / n_step, nth);
+
+                if (ith < num_threads) {
+                    const int64_t num_n_per_thread0   = round_down(n / num_threads, n_step);
+                    const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0;
+
+                    const int64_t n_start      = ith * num_n_per_thread0;
+                    const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
+
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernel->get_lhs_offset, m_start, k);
+                    const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k);
+                    const size_t dst_offset        = kernel->get_dst_offset(m_start, n_start, dst_stride);
+
+                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * rhs_ptr = rhs_packed + rhs_packed_offset;
+                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch + dst_offset);
+
+                    variant_call<void>(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
+                }
+            }
+
+            if (batch_idx != batch_size - 1) {
+                // This barrier is necessary when the batch size is larger than 1. While processing a batch,
+                // the work data buffer (params->wdata) is used as temporary storage which means that only
+                // a single batch can be processed at any given time. No barrier is needed for the last
+                // batch since GGML inserts a barrier between the execution of every operator.
+                ggml_barrier(params->threadpool);
+            }
         }
-        return false;
+
+        return true;
+    }
+
+    bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        GGML_ASSERT(kernels);
+
+        kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
+        lhs_packing_info * lhs_info = &kernels->lhs_info;
+
+        GGML_ASSERT(kernel);
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+
+        size_t mr = kernel->get_mr();
+        size_t kr = kernel->get_kr();
+        size_t sr = kernel->get_sr();
+
+        const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
+        uint8_t * lhs_packed       = (uint8_t*)params->wdata;
+        const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+
+        const size_t n_step = kernel->get_n_step();
+        const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
+        const size_t n_start = ith * num_n_per_thread;
+
+        size_t n_to_process = num_n_per_thread;
+        if ((n_start + n_to_process) > n) {
+            n_to_process = n - n_start;
+        }
+
+        // Calculate number of columns to be processed per thread
+        const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
+        const size_t m_start = ith * num_m_per_thread;
+        size_t m_to_process = num_m_per_thread;
+        if ((m_start + m_to_process) > m) {
+            m_to_process = m - m_start;
+        }
+
+        if (m_start < m) {
+            // 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);
+            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);
+        }
+
+        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 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);
+
+        variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+                           sizeof(float), -FLT_MAX, FLT_MAX);
+
+        return true;
     }
 
 public:
@@ -169,13 +352,13 @@ public:
         size_t sr      = ctx.kernels->gemm.get_sr();
 
 #ifndef NDEBUG
-        const size_t repacked_size = ctx.kernels->rhs_info.packed_size(n, k, nr, kr, QK4_0);
+        const size_t repacked_size = variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
         GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
 #endif
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
-        ctx.kernels->rhs_info.pack_func(1, n, k, nr, kr, sr, QK4_0, (const uint8_t *)data, NULL, tensor->data, 0, &params);
+        variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
 
         return 0;
 
@@ -189,7 +372,7 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc
 }
 }  // namespace ggml::cpu::kleidiai
 
-GGML_API enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
 
     GGML_UNUSED(buffer);
@@ -238,12 +421,11 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
 namespace ggml::cpu::kleidiai {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        if (    op->op == GGML_OP_MUL_MAT &&
-                op->src[0]->type == GGML_TYPE_Q4_0 &&
-                op->src[0]->buffer &&
-                (ggml_n_dims(op->src[0]) == 2) &&
-                op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels
-                ) {
+        if (op->op == GGML_OP_MUL_MAT &&
+            op->src[0]->type == GGML_TYPE_Q4_0 &&
+            op->src[0]->buffer &&
+            (ggml_n_dims(op->src[0]) == 2) &&
+            op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
             if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                 return false;
             }
@@ -260,6 +442,19 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }
+            else if (ggml_kleidiai_select_kernels(ctx.features, op) &&
+                     op->src[0]->op == GGML_OP_VIEW &&
+                     (op->src[1]->op == GGML_OP_PERMUTE || op->src[1]->op ==  GGML_OP_SOFT_MAX) &&
+                     op->src[1]->ne[1] > 1) {
+                if ((op->src[0]->nb[0] != 2) ||
+                    (op->src[1]->nb[0] != 4) ||
+                    (op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
+                    (op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
+                    return nullptr;
+                }
+
+                return ggml::cpu::kleidiai::get_tensor_traits(NULL, NULL);
+            }
         }
         return nullptr;
     }

ggml/src/ggml-cpu/ops.cpp

@@ -2691,6 +2691,109 @@ static void ggml_compute_forward_gelu(
     }
 }
 
+// ggml_compute_forward_gelu_erf
+
+static void ggml_compute_forward_gelu_erf_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_erf_f32(nc,
+                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu_erf_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_erf_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            GGML_UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu_erf(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_gelu_erf_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_gelu_erf_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_gelu_quick
 
 static void ggml_compute_forward_gelu_quick_f32(
@@ -7749,6 +7852,10 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_gelu(params, dst);
             } break;
+        case GGML_UNARY_OP_GELU_ERF:
+            {
+                ggml_compute_forward_gelu_erf(params, dst);
+            } break;
         case GGML_UNARY_OP_GELU_QUICK:
             {
                 ggml_compute_forward_gelu_quick(params, dst);

ggml/src/ggml-cpu/vec.h

@@ -428,6 +428,7 @@ inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp
 static const float GELU_COEF_A     = 0.044715f;
 static const float GELU_QUICK_COEF = -1.702f;
 static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+static const float SQRT_2_INV      = 0.70710678118654752440084436210484f;
 
 inline static float ggml_gelu_f32(float x) {
     return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
@@ -440,6 +441,14 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp
     }
 }
 
+inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float xi = GGML_FP16_TO_FP32(x[i]);
+        float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
+        y[i] = GGML_FP32_TO_FP16(res);
+    }
+}
+
 #ifdef GGML_GELU_FP16
 inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
     uint16_t t;
@@ -463,6 +472,13 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
 }
 #endif
 
+inline static void ggml_vec_gelu_erf_f32(const int n, float * y, const float * x) {
+    for (int i = 0; i < n; ++i) {
+        float xi = x[i];
+        y[i] = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
+    }
+}
+
 inline static float ggml_gelu_quick_f32(float x) {
     return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x)));
 }

ggml/src/ggml-cuda/CMakeLists.txt

@@ -118,7 +118,7 @@ if (CUDAToolkit_FOUND)
 
     set(CUDA_CXX_FLAGS "")
 
-    set(CUDA_FLAGS -use_fast_math)
+    set(CUDA_FLAGS -use_fast_math -extended-lambda)
 
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
         # Options are:

ggml/src/ggml-cuda/acc.cu

@@ -1,47 +1,61 @@
 #include "acc.cuh"
 
-static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset) {
-    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int64_t ne,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
+    const int64_t i = blockDim.x * blockIdx.x + threadIdx.x;
+
     if (i >= ne) {
         return;
     }
-    int src1_idx = i - offset;
-    int oz = src1_idx / nb2;
-    int oy = (src1_idx - (oz * nb2)) / nb1;
-    int ox = src1_idx % nb1;
-    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
-        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
-    } else {
-        dst[i] = x[i];
+
+    int64_t src1_idx = i - offset;
+
+    int64_t tmp = src1_idx;
+    const int64_t i13 = tmp / s13;
+    tmp -= i13 * s13;
+    const int64_t i12 = tmp / s12;
+    tmp -= i12 * s12;
+    const int64_t i11 = tmp / s11;
+    tmp -= i11 * s11;
+    const int64_t i10 = tmp;
+
+    float val = x[i];
+    if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
+        val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10];
     }
+    dst[i] = val;
 }
 
-static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
-    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
-    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int64_t n_elements,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s1, const int64_t s2, const int64_t s3, const int64_t offset, cudaStream_t stream) {
+    const int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
 }
 
 void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *)  dst->data;
+
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
 
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
+    GGML_ASSERT(ggml_is_contiguously_allocated(dst));
 
-    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
+    const int64_t s1     = dst->op_params[0] / sizeof(float);
+    const int64_t s2     = dst->op_params[1] / sizeof(float);
+    const int64_t s3     = dst->op_params[2] / sizeof(float);
+    const int64_t offset = dst->op_params[3] / sizeof(float);
+
+    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], s1, s2, s3, offset, stream);
 }

ggml/src/ggml-cuda/common.cuh

@@ -296,6 +296,25 @@ static __device__ void no_device_code(
 #define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
 #endif // __CUDA_ARCH__
 
+// The compiler is always able to unroll loops if they contain continue expressions.
+// In such cases loop unrolling can still be achieved via recursion:
+template <int n>
+struct ggml_cuda_unroll {
+    template <typename Func, typename... Args>
+    __device__ void operator()(const Func & f, Args... args) const {
+        f(n - 1, args...);
+        ggml_cuda_unroll<n - 1>{}(f, args...);
+    }
+};
+
+template <>
+struct ggml_cuda_unroll<1> {
+    template <typename Func, typename... Args>
+    __device__ void operator()(const Func & f, Args... args) const {
+        f(0, args...);
+    }
+};
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE

ggml/src/ggml-cuda/cp-async.cuh

@@ -2,6 +2,17 @@
 
 #include "common.cuh"
 
+
+static __device__ __forceinline__ unsigned int ggml_cuda_cvta_generic_to_shared(void * generic_ptr) {
+#ifdef CP_ASYNC_AVAILABLE
+    return __cvta_generic_to_shared(generic_ptr);
+#else
+    GGML_UNUSED(generic_ptr);
+    NO_DEVICE_CODE;
+    return 0;
+#endif // CP_ASYNC_AVAILABLE
+}
+
 // Copies data from global to shared memory, cg == cache global.
 // Both the src and dst pointers must be aligned to 16 bit.
 // Shared memory uses 32 bit addressing, the pointer is passed as unsigned int.

ggml/src/ggml-cuda/cpy.cu

@@ -1,5 +1,8 @@
 #include "cpy.cuh"
 #include "dequantize.cuh"
+#ifdef GGML_USE_MUSA
+#include "ggml-musa/mudnn.cuh"
+#endif // GGML_USE_MUSA
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
@@ -597,7 +600,14 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 #endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
-        CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
+#ifdef GGML_USE_MUSA
+        if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
+            CUDA_CHECK(mudnnMemcpyAsync(ctx, src1, src0));
+        } else
+#endif // GGML_USE_MUSA
+        {
+            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_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {

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

@@ -516,7 +516,7 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
         nullptr;
 }
 
-template<int D, int ncols1, int ncols2, int KQ_stride> // D == head size
+template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
         float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne11) {
@@ -665,23 +665,27 @@ static void on_no_fattn_vec_case(const int D) {
         fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
         GGML_ABORT("fatal error");
     } else {
-        fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
+        fprintf(stderr, "Unsupported KV type combination for head_size %d.\n", D);
         fprintf(stderr, "Only f16 is supported.\n");
         GGML_ABORT("fatal error");
     }
 }
 
-template <int D, int ncols1, int ncols2, int KQ_stride>
+template <int DV, int ncols1, int ncols2>
 void launch_fattn(
     ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
     const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
+    const bool is_mla = DV == 512; // TODO better parameterization
+
     const ggml_tensor * Q = dst->src[0];
     const ggml_tensor * K = dst->src[1];
     const ggml_tensor * V = dst->src[2];
 
+    GGML_ASSERT(V || is_mla);
+
     const ggml_tensor * mask = dst->src[3];
 
     ggml_tensor * KQV = dst;
@@ -689,9 +693,13 @@ void launch_fattn(
     GGML_ASSERT(Q->type == GGML_TYPE_F32);
     GGML_ASSERT(KQV->type == GGML_TYPE_F32);
 
+    GGML_ASSERT(      Q->nb[0] == ggml_element_size(Q));
+    GGML_ASSERT(      K->nb[0] == ggml_element_size(K));
+    GGML_ASSERT(!V || V->nb[0] == ggml_element_size(V));
+
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
     GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
-                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+        "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.");
 
@@ -713,10 +721,10 @@ void launch_fattn(
     size_t nb12 = K->nb[2];
     size_t nb13 = K->nb[3];
 
-    const char * V_data = (const char *) V->data;
-    size_t nb21 = V->nb[1];
-    size_t nb22 = V->nb[2];
-    size_t nb23 = V->nb[3];
+    const char * V_data = V ? (const char *) V->data : nullptr;
+    size_t nb21 = V ? V->nb[1] : nb11;
+    size_t nb22 = V ? V->nb[2] : nb12;
+    size_t nb23 = V ? V->nb[3] : nb13;
 
     if (need_f16_K && K->type != GGML_TYPE_F16) {
         GGML_ASSERT(ggml_is_contiguously_allocated(K));
@@ -733,7 +741,7 @@ void launch_fattn(
         nb13 = nb13*bs*sizeof(half)/ts;
     }
 
-    if (need_f16_V && V->type != GGML_TYPE_F16) {
+    if (V && need_f16_V && V->type != GGML_TYPE_F16) {
         GGML_ASSERT(ggml_is_contiguously_allocated(V));
         V_f16.alloc(ggml_nelements(V));
         to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
@@ -754,10 +762,13 @@ void launch_fattn(
     const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
 
     const dim3 block_dim(warp_size, nwarps, 1);
+    int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
+    CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+
     dim3 blocks_num;
     if (stream_k) {
         // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
-        const int max_blocks = 2*nsm;
+        const int max_blocks = max_blocks_per_sm*nsm;
         const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
         const int tiles_efficiency_percent = 100 * ntiles_total / (max_blocks*tiles_nwaves);
 
@@ -769,14 +780,11 @@ void launch_fattn(
         blocks_num.y = 1;
         blocks_num.z = 1;
 
-        dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float));
+        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.
 
-        int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
-        CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
-
         // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave:
         parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1);
 
@@ -853,19 +861,19 @@ void launch_fattn(
 
     if (stream_k) {
         if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
-            const dim3 block_dim_combine(D, 1, 1);
+            const dim3 block_dim_combine(DV, 1, 1);
             const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};
 
-            flash_attn_stream_k_fixup<D, ncols1, ncols2, KQ_stride>
+            flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
                 ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]);
         }
     } else if (parallel_blocks > 1) {
-        const dim3 block_dim_combine(D, 1, 1);
+        const dim3 block_dim_combine(DV, 1, 1);
         const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
         const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
-        flash_attn_combine_results<D>
+        flash_attn_combine_results<DV>
             <<<blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream>>>
             (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
     }

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

@@ -307,7 +307,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         case 128: {
@@ -315,7 +315,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         default: {

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -318,7 +318,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         case 128: {
@@ -326,7 +326,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         default: {

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

@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // GGML_USE_HIP
 static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -48,6 +48,12 @@ static __global__ void flash_attn_vec_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    if (ncols > 1) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
@@ -91,6 +97,13 @@ static __global__ void flash_attn_vec_ext_f16(
             kqsum_shared[j][threadIdx.x] = 0.0f;
         }
     }
+
+    __shared__ half maskh_shared[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        maskh_shared[j*D + tid] = 0.0f;
+    }
+
     __syncthreads();
 
     // Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
@@ -168,12 +181,43 @@ static __global__ void flash_attn_vec_ext_f16(
     for (int j = 0; j < ncols; ++j) {
         KQ[j*D + tid] = -HALF_MAX_HALF;
     }
+    __syncthreads();
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 
     for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
+        if (mask) {
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                maskh_shared[j*D + tid] = slopeh*maskh[j*ne11 + k_VKQ_0 + tid];
+            }
+
+            __syncthreads();
+
+            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
+            // In such cases, skip the KV slice.
+            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
+#ifndef GGML_USE_HIP
+            bool skip = true;
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    const float2 tmp = __half22float2(((const half2 *) maskh_shared)[j*(D/2) + i]);
+                    skip = skip && isinf(tmp.x) && isinf(tmp.y);
+                }
+            }
+            if (__all_sync(0xFFFFFFFF, skip)) {
+                __syncthreads();
+                continue;
+            }
+#endif // GGML_USE_HIP
+        }
+
         // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
         // see https://github.com/ggerganov/llama.cpp/pull/7061 .
         // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
@@ -201,7 +245,7 @@ static __global__ void flash_attn_vec_ext_f16(
                     sum = logit_softcap*tanhf(sum);
                 }
 
-                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+                sum += maskh_shared[j*D + i_KQ];
 
                 if (ncols == 1) {
                     kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
@@ -315,7 +359,7 @@ void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx,
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>
@@ -334,7 +378,9 @@ void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
-    if (Q->ne[1] == 1) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (Q->ne[1] == 1 || GGML_CUDA_CC_IS_NVIDIA(cc)) {
         constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // GGML_USE_HIP
 static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -60,6 +60,12 @@ static __global__ void flash_attn_vec_ext_f32(
         NO_DEVICE_CODE;
         return;
     }
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    if (ncols > 1) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
@@ -104,6 +110,13 @@ static __global__ void flash_attn_vec_ext_f32(
             kqsum_shared[j][threadIdx.x] = 0.0f;
         }
     }
+
+    __shared__ float maskf_shared[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        maskf_shared[j*D + tid] = 0.0f;
+    }
+
     __syncthreads();
 
     // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
@@ -181,6 +194,35 @@ static __global__ void flash_attn_vec_ext_f32(
     for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
+        if (mask) {
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + k_VKQ_0 + tid]);
+            }
+
+            __syncthreads();
+
+            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
+            // In such cases, skip the KV slice.
+            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
+#ifndef GGML_USE_HIP
+            bool skip = true;
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+                for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    skip = skip && isinf(maskf_shared[j*D + i]);
+                }
+            }
+            if (__all_sync(0xFFFFFFFF, skip)) {
+                __syncthreads();
+                continue;
+            }
+#endif // GGML_USE_HIP
+        }
+
         float kqmax_new_arr[ncols];
 #pragma unroll
         for (int j = 0; j < ncols; ++j) {
@@ -204,7 +246,7 @@ static __global__ void flash_attn_vec_ext_f32(
                     sum = logit_softcap*tanhf(sum);
                 }
 
-                sum += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+                sum += maskf_shared[j*D + i_KQ];
 
                 kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum);
 
@@ -310,7 +352,7 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx,
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>
@@ -326,7 +368,9 @@ void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
-    if (Q->ne[1] == 1) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (Q->ne[1] == 1 || GGML_CUDA_CC_IS_NVIDIA(cc)) {
         constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;

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

@@ -490,7 +490,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
         fattn_kernel = flash_attn_ext_f16<
             D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
     }
-    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
 }
 
 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/fattn.cu

@@ -8,58 +8,33 @@
 #include "fattn-wmma-f16.cuh"
 #include "fattn.cuh"
 
-template <int D, int ncols2>
+template <int DKQ, int DV, int ncols2>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const ggml_tensor * Q = dst->src[0];
 
-    if (Q->ne[1] <= 8/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<D, 8/ncols2, ncols2>(ctx, dst);
-        return;
+    if constexpr (ncols2 <= 8) {
+        if (Q->ne[1] <= 8/ncols2) {
+            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
+            return;
+        }
     }
 
     if (Q->ne[1] <= 16/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<D, 16/ncols2, ncols2>(ctx, dst);
+        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
         return;
     }
 
-    if (Q->ne[1] <= 32/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<D, 32/ncols2, ncols2>(ctx, dst);
+    if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || Q->ne[1] <= 32/ncols2) {
+        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
         return;
     }
 
-    ggml_cuda_flash_attn_ext_mma_f16_case<D, 64/ncols2, ncols2>(ctx, dst);
+    ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 64/ncols2, ncols2>(ctx, dst);
 }
 
-template <int ncols2>
-static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-
-    switch (Q->ne[0]) {
-        case 64:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 64, ncols2>(ctx, dst);
-            break;
-        case 80:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 80, ncols2>(ctx, dst);
-            break;
-        case 96:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 96, ncols2>(ctx, dst);
-            break;
-        case 112:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<112, ncols2>(ctx, dst);
-            break;
-        case 128:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<128, ncols2>(ctx, dst);
-            break;
-        case 256:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<256, ncols2>(ctx, dst);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
-}
-
-static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+template <int DKQ, int DV>
+static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
@@ -68,27 +43,79 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
     float max_bias = 0.0f;
     memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
 
-    const float use_gqa_opt = mask && max_bias == 0.0f;
+    const bool use_gqa_opt = mask && max_bias == 0.0f;
 
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
 
     if (use_gqa_opt && gqa_ratio % 8 == 0) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<8>(ctx, dst);
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio == 4) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<4>(ctx, dst);
+    if (use_gqa_opt && gqa_ratio % 4 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio == 2) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<2>(ctx, dst);
+    if (use_gqa_opt && gqa_ratio % 2 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
         return;
     }
 
-    ggml_cuda_flash_attn_ext_mma_f16_switch_hs<1>(ctx, dst);
+    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+}
+
+static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    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];
+    const ggml_tensor * mask = dst->src[3];
+
+    switch (Q->ne[0]) {
+        case 64:
+            GGML_ASSERT(V->ne[0] == 64);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 64,  64>(ctx, dst);
+            break;
+        case 80:
+            GGML_ASSERT(V->ne[0] == 80);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 80,  80>(ctx, dst);
+            break;
+        case 96:
+            GGML_ASSERT(V->ne[0] == 96);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 96,  96>(ctx, dst);
+            break;
+        case 112:
+            GGML_ASSERT(V->ne[0] == 112);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<112, 112>(ctx, dst);
+            break;
+        case 128:
+            GGML_ASSERT(V->ne[0] == 128);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<128, 128>(ctx, dst);
+            break;
+        case 256:
+            GGML_ASSERT(V->ne[0] == 256);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
+            break;
+        case 576: {
+            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+            GGML_ASSERT(V->ne[0] == 512);
+            float max_bias = 0.0f;
+            memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+            const bool use_gqa_opt = mask && max_bias == 0.0f;
+            GGML_ASSERT(use_gqa_opt);
+
+            GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+            const int gqa_ratio = Q->ne[2] / K->ne[2];
+            GGML_ASSERT(gqa_ratio % 16 == 0);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+        } break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
 }
 
 #define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
@@ -299,7 +326,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
     const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
     const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
-    const bool can_use_vector_kernel = Q->ne[0] % (2*warp_size) == 0;
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
     if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);

ggml/src/ggml-cuda/getrows.cu

@@ -10,10 +10,11 @@ static __global__ void k_get_rows(
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
-    const int i10 =  blockDim.y*blockIdx.y + threadIdx.y;
-    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
-    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+    const int i00 = (blockIdx.y * blockDim.x + threadIdx.x)*2;
+    const int i10 =  blockIdx.x;
+    const int i11 =  blockIdx.z / ne12;
+    const int i12 =  blockIdx.z % ne12;
 
     if (i00 >= ne00) {
         return;
@@ -46,10 +47,11 @@ static __global__ void k_get_rows_float(
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    const int i00 =  blockIdx.x*blockDim.x + threadIdx.x;
-    const int i10 =  blockDim.y*blockIdx.y + threadIdx.y;
-    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
-    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+    const int i00 = blockIdx.y * blockDim.x + threadIdx.x;
+    const int i10 = blockIdx.x;
+    const int i11 = blockIdx.z / ne12;
+    const int i12 = blockIdx.z % ne12;
 
     if (i00 >= ne00) {
         return;
@@ -94,8 +96,8 @@ static void get_rows_cuda_q(
         const size_t nb1, const size_t nb2, const size_t nb3,
         cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
-    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+    const int block_num_y = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(ne10, block_num_y, ne11*ne12);
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);
@@ -127,8 +129,8 @@ static void get_rows_cuda_float(
         const size_t nb1, const size_t nb2, const size_t nb3,
         cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
-    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+    const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(ne10, block_num_y, ne11*ne12);
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);

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

@@ -1909,13 +1909,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
+    // If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q.
+    // But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data.
+    // Therefore, in such cases use cuBLAS.
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
+        && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
+
     bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
-    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
     bool any_gpus_with_slow_fp16   = false;
@@ -3215,16 +3221,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return false;
 #endif // FLASH_ATTN_AVAILABLE
             if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
-                // different head sizes of K and V are not supported yet
-                return false;
+                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
+                if (!new_mma_available(cc)) {
+                    return false;
+                }
+                const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
+                return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
             }
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek MLA
-                return false;
-            }
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cuda/mmq.cu

@@ -91,11 +91,11 @@ void ggml_cuda_mul_mat_q(
 
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-        GGML_ASSERT(!src0->view_src);
         const size_t size_data  = ggml_nbytes(src0);
         const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
         if (size_alloc > size_data) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
         }
     }
@@ -122,6 +122,7 @@ void ggml_cuda_mul_mat_q(
             const int64_t s13 = src1->nb[3] / ts_src1;
             quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
                 ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+            CUDA_CHECK(cudaGetLastError());
         }
 
         const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
@@ -205,6 +206,7 @@ void ggml_cuda_mul_mat_q(
         const int64_t s13 = src1->nb[2] / ts_src1;
         quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
             ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+        CUDA_CHECK(cudaGetLastError());
     }
 
     const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));

ggml/src/ggml-cuda/mmvq.cu

@@ -515,11 +515,11 @@ void ggml_cuda_mul_mat_vec_q(
 
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-        GGML_ASSERT(!src0->view_src);
         const size_t size_data  = ggml_nbytes(src0);
         const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
         if (size_alloc > size_data) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
         }
     }

ggml/src/ggml-cuda/quantize.cu

@@ -56,13 +56,13 @@ static __global__ void quantize_mmq_q8_1(
     constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
     constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
 
-    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.y + threadIdx.x)*4;
 
     if (i0 >= ne0) {
         return;
     }
 
-    const int64_t i1 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
     const int64_t i2 = blockIdx.z % ne2;
     const int64_t i3 = blockIdx.z / ne2;
 
@@ -75,8 +75,8 @@ static __global__ void quantize_mmq_q8_1(
 
     block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
 
-    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.x*gridDim.y*blockDim.x/QK8_1); // first block of channel
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.x;                    // block index in channel
     const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block
 
     // Load 4 floats per thread and calculate max. abs. value between them:
@@ -166,8 +166,9 @@ void quantize_mmq_q8_1_cuda(
     GGML_ASSERT(ne00 % 4 == 0);
     GGML_ASSERT(ne0 % (4*QK8_1) == 0);
 
-    const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
-    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
+    // ne1 tends to assume the highest values, therefore use it as the "x" dimension of the CUDA grid:
+    const int64_t block_num_y = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(ne1, block_num_y, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
     switch (mmq_get_q8_1_ds_layout(type_src0)) {
         case MMQ_Q8_1_DS_LAYOUT_D4:

ggml/src/ggml-cuda/sum.cu

@@ -31,7 +31,7 @@ void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguously_allocated(src0));
 
     const float * src0_d = (const float *) src0->data;
     float * dst_d = (float *) dst->data;

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 1, 8);
-DECL_FATTN_MMA_F16_CASE(80, 1, 8);
-DECL_FATTN_MMA_F16_CASE(96, 1, 8);
-DECL_FATTN_MMA_F16_CASE(112, 1, 8);
-DECL_FATTN_MMA_F16_CASE(128, 1, 8);
-DECL_FATTN_MMA_F16_CASE(256, 1, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 16, 1);
-DECL_FATTN_MMA_F16_CASE(80, 16, 1);
-DECL_FATTN_MMA_F16_CASE(96, 16, 1);
-DECL_FATTN_MMA_F16_CASE(112, 16, 1);
-DECL_FATTN_MMA_F16_CASE(128, 16, 1);
-DECL_FATTN_MMA_F16_CASE(256, 16, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 16, 2);
-DECL_FATTN_MMA_F16_CASE(80, 16, 2);
-DECL_FATTN_MMA_F16_CASE(96, 16, 2);
-DECL_FATTN_MMA_F16_CASE(112, 16, 2);
-DECL_FATTN_MMA_F16_CASE(128, 16, 2);
-DECL_FATTN_MMA_F16_CASE(256, 16, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 16, 4);
-DECL_FATTN_MMA_F16_CASE(80, 16, 4);
-DECL_FATTN_MMA_F16_CASE(96, 16, 4);
-DECL_FATTN_MMA_F16_CASE(112, 16, 4);
-DECL_FATTN_MMA_F16_CASE(128, 16, 4);
-DECL_FATTN_MMA_F16_CASE(256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 2, 4);
-DECL_FATTN_MMA_F16_CASE(80, 2, 4);
-DECL_FATTN_MMA_F16_CASE(96, 2, 4);
-DECL_FATTN_MMA_F16_CASE(112, 2, 4);
-DECL_FATTN_MMA_F16_CASE(128, 2, 4);
-DECL_FATTN_MMA_F16_CASE(256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 2, 8);
-DECL_FATTN_MMA_F16_CASE(80, 2, 8);
-DECL_FATTN_MMA_F16_CASE(96, 2, 8);
-DECL_FATTN_MMA_F16_CASE(112, 2, 8);
-DECL_FATTN_MMA_F16_CASE(128, 2, 8);
-DECL_FATTN_MMA_F16_CASE(256, 2, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 32, 1);
-DECL_FATTN_MMA_F16_CASE(80, 32, 1);
-DECL_FATTN_MMA_F16_CASE(96, 32, 1);
-DECL_FATTN_MMA_F16_CASE(112, 32, 1);
-DECL_FATTN_MMA_F16_CASE(128, 32, 1);
-DECL_FATTN_MMA_F16_CASE(256, 32, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 32, 2);
-DECL_FATTN_MMA_F16_CASE(80, 32, 2);
-DECL_FATTN_MMA_F16_CASE(96, 32, 2);
-DECL_FATTN_MMA_F16_CASE(112, 32, 2);
-DECL_FATTN_MMA_F16_CASE(128, 32, 2);
-DECL_FATTN_MMA_F16_CASE(256, 32, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 4, 2);
-DECL_FATTN_MMA_F16_CASE(80, 4, 2);
-DECL_FATTN_MMA_F16_CASE(96, 4, 2);
-DECL_FATTN_MMA_F16_CASE(112, 4, 2);
-DECL_FATTN_MMA_F16_CASE(128, 4, 2);
-DECL_FATTN_MMA_F16_CASE(256, 4, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 4, 4);
-DECL_FATTN_MMA_F16_CASE(80, 4, 4);
-DECL_FATTN_MMA_F16_CASE(96, 4, 4);
-DECL_FATTN_MMA_F16_CASE(112, 4, 4);
-DECL_FATTN_MMA_F16_CASE(128, 4, 4);
-DECL_FATTN_MMA_F16_CASE(256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 4, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 4, 8);
-DECL_FATTN_MMA_F16_CASE(80, 4, 8);
-DECL_FATTN_MMA_F16_CASE(96, 4, 8);
-DECL_FATTN_MMA_F16_CASE(112, 4, 8);
-DECL_FATTN_MMA_F16_CASE(128, 4, 8);
-DECL_FATTN_MMA_F16_CASE(256, 4, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 64, 1);
-DECL_FATTN_MMA_F16_CASE(80, 64, 1);
-DECL_FATTN_MMA_F16_CASE(96, 64, 1);
-DECL_FATTN_MMA_F16_CASE(112, 64, 1);
-DECL_FATTN_MMA_F16_CASE(128, 64, 1);
-DECL_FATTN_MMA_F16_CASE(256, 64, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 64, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 64, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 64, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 64, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 64, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 64, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 1);
-DECL_FATTN_MMA_F16_CASE(80, 8, 1);
-DECL_FATTN_MMA_F16_CASE(96, 8, 1);
-DECL_FATTN_MMA_F16_CASE(112, 8, 1);
-DECL_FATTN_MMA_F16_CASE(128, 8, 1);
-DECL_FATTN_MMA_F16_CASE(256, 8, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 2);
-DECL_FATTN_MMA_F16_CASE(80, 8, 2);
-DECL_FATTN_MMA_F16_CASE(96, 8, 2);
-DECL_FATTN_MMA_F16_CASE(112, 8, 2);
-DECL_FATTN_MMA_F16_CASE(128, 8, 2);
-DECL_FATTN_MMA_F16_CASE(256, 8, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 4);
-DECL_FATTN_MMA_F16_CASE(80, 8, 4);
-DECL_FATTN_MMA_F16_CASE(96, 8, 4);
-DECL_FATTN_MMA_F16_CASE(112, 8, 4);
-DECL_FATTN_MMA_F16_CASE(128, 8, 4);
-DECL_FATTN_MMA_F16_CASE(256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 8);
-DECL_FATTN_MMA_F16_CASE(80, 8, 8);
-DECL_FATTN_MMA_F16_CASE(96, 8, 8);
-DECL_FATTN_MMA_F16_CASE(112, 8, 8);
-DECL_FATTN_MMA_F16_CASE(128, 8, 8);
-DECL_FATTN_MMA_F16_CASE(256, 8, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 8);

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

@@ -18,7 +18,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f
 
 """
 
-SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {ncols1}, {ncols2});\n"
+SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size_kq}, {head_size_v}, {ncols1}, {ncols2});\n"
 
 TYPES_MMQ = [
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
@@ -57,18 +57,21 @@ for vkq_size in [16, 32]:
                 with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
                     f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v))
 
-for ncols in [8, 16, 32, 64, 128]:
-    for ncols2 in [1, 2, 4, 8]:
+for ncols in [8, 16, 32, 64]:
+    for ncols2 in [1, 2, 4, 8, 16]:
+        if ncols2 > ncols:
+            continue
         ncols1 = ncols // ncols2
-        if ncols == 128:
-            continue  # Too much register pressure.
         with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f:
             f.write(SOURCE_FATTN_MMA_START)
 
-            for head_size in [64, 80, 96, 112, 128, 256]:
-                if ncols == 128 and head_size == 256:
-                    continue  # Needs too much shared memory.
-                f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size=head_size))
+            for head_size_kq in [64, 80, 96, 112, 128, 256, 576]:
+                if head_size_kq != 576 and ncols2 == 16:
+                    continue
+                if head_size_kq == 576 and ncols2 != 16:
+                    continue
+                head_size_v = head_size_kq if head_size_kq != 576 else 512
+                f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))
 
 for type in TYPES_MMQ:
     with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f:

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

@@ -207,6 +207,10 @@ typedef struct {
     float    attn_factor;
     float    beta_fast;
     float    beta_slow;
+    int32_t  sect_0;
+    int32_t  sect_1;
+    int32_t  sect_2;
+    int32_t  sect_3;
 } ggml_metal_kargs_rope;
 
 typedef struct {
@@ -299,21 +303,42 @@ typedef struct {
 } ggml_metal_kargs_mul_mv_ext;
 
 typedef struct {
-    int32_t  nei0;
-    int32_t  nei1;
-    uint64_t nbi1;
+    int32_t  ne10;
+    int32_t  ne11;  // n_expert_used (bcast)
+    uint64_t nb11;
+    uint64_t nb12;
+    int32_t  neh11; // n_tokens
+    uint64_t nbh11;
+    int32_t  ne20;  // n_expert_used
+    uint64_t nb21;
+} ggml_metal_kargs_mul_mm_id_map0;
+
+typedef struct {
+    int32_t  ne20; // n_expert_used
+    int32_t  neh0;
+    int32_t  neh1;
+    uint64_t nbh1;
+    uint64_t nbh2;
+    int32_t  ne0;
+    uint64_t nb1;
+    uint64_t nb2;
+} ggml_metal_kargs_mul_mm_id_map1;
+
+typedef struct {
     int32_t  ne00;
     int32_t  ne02;
     uint64_t nb01;
     uint64_t nb02;
-    int32_t  ne11;
-    int32_t  ne12;
-    int32_t  ne13;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    int32_t  ne0;
-    int32_t  ne1;
+    uint64_t nb03;
+    int32_t  neh12;
+    uint64_t nbh10;
+    uint64_t nbh11;
+    uint64_t nbh12;
+    uint64_t nbh13;
+    int32_t  neh0;
+    int32_t  neh1;
+    int16_t  r2;
+    int16_t  r3;
 } ggml_metal_kargs_mul_mm_id;
 
 typedef struct {