server: fix regression on streamed non-chat completion w/ stops (#13785)

* more forgiving message diffs: partial stop words aren't erased, full stops are

* Add (slow) server test for completion + stream + stop

.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/get-tag-name/action.yml

@@ -0,0 +1,22 @@
+name: "Determine tag name"
+description: "Determine the tag name to use for a release"
+outputs:
+  name:
+    description: "The name of the tag"
+    value: ${{ steps.tag.outputs.name }}
+
+runs:
+  using: "composite"
+  steps:
+    - name: Determine tag name
+      id: tag
+      shell: bash
+      run: |
+        BUILD_NUMBER="$(git rev-list --count HEAD)"
+        SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+        if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+          echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+        else
+          SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+          echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+        fi

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

@@ -0,0 +1,67 @@
+name: "Windows - Setup CUDA Toolkit"
+description: "Setup CUDA Toolkit for Windows"
+inputs:
+  cuda_version:
+    description: "CUDA toolkit version"
+    required: true
+
+runs:
+  using: "composite"
+  steps:
+    - name: Install Cuda Toolkit 11.7
+      if: ${{ inputs.cuda_version == '11.7' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
+          choco install unzip -y
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+    - name: Install Cuda Toolkit 12.4
+      if: ${{ inputs.cuda_version == '12.4' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
+          choco install unzip -y
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

.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

@@ -2,30 +2,19 @@ name: CI
 
 on:
   workflow_dispatch: # allows manual triggering
-    inputs:
-      create_release:
-        description: 'Create new release'
-        required: true
-        type: boolean
   push:
     branches:
       - master
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
-# Fine-grant permission
-# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
-permissions:
-  contents: write # for creating release
-
 env:
-  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
   LLAMA_LOG_COLORS: 1
@@ -40,8 +29,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -74,33 +61,6 @@ jobs:
           cd build
           ctest -L 'main|curl' --verbose --timeout 900
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
-          name: llama-bin-macos-arm64.zip
-
   macOS-latest-cmake-x64:
     runs-on: macos-13
 
@@ -108,8 +68,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -143,33 +101,6 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
-          name: llama-bin-macos-x64.zip
-
   ubuntu-cpu-cmake:
     strategy:
       matrix:
@@ -185,8 +116,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -225,33 +154,6 @@ jobs:
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
           ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
-          name: llama-bin-ubuntu-${{ matrix.build }}.zip
-
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
 
@@ -378,8 +280,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -407,34 +307,7 @@ jobs:
         run: |
           cd build
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 2700
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
-          name: llama-bin-ubuntu-vulkan-x64.zip
+          ctest -L main --verbose --timeout 3600
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
@@ -478,7 +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
@@ -831,8 +704,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -935,35 +806,6 @@ jobs:
       #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
       #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        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\*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
-          name: llama-bin-win-${{ matrix.build }}.zip
-
   ubuntu-latest-cmake-cuda:
     runs-on: ubuntu-latest
     container: nvidia/cuda:12.6.2-devel-ubuntu24.04
@@ -972,8 +814,6 @@ jobs:
         - name: Clone
           id: checkout
           uses: actions/checkout@v4
-          with:
-            fetch-depth: 0
 
         - name: Install dependencies
           env:
@@ -1005,77 +845,23 @@ jobs:
     strategy:
       matrix:
         cuda: ['12.4', '11.7']
-        build: ['cuda']
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
 
       - name: Install ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
         with:
-          key: ${{ github.job }}-${{ matrix.cuda }}-${{ matrix.build }}
+          key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
           evict-old-files: 1d
 
-      - name: Install Cuda Toolkit 11.7
-        if: ${{ matrix.cuda == '11.7' }}
-        run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-      - name: Install Cuda Toolkit 12.4
-        if: ${{ matrix.cuda == '12.4' }}
-        run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+      - name: Install Cuda Toolkit
+        uses: ./.github/actions/windows-setup-cuda
+        with:
+          cuda_version: ${{ matrix.cuda }}
 
       - name: Install Ninja
         id: install_ninja
@@ -1105,51 +891,6 @@ jobs:
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        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-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
-
-      - name: Copy and pack Cuda runtime
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        run: |
-          echo "Cuda install location: ${{ env.CUDA_PATH }}"
-          $dst='.\build\bin\cudart\'
-          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
-
-      - name: Upload Cuda runtime
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
-
   windows-latest-cmake-sycl:
     runs-on: windows-latest
 
@@ -1158,15 +899,13 @@ 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
@@ -1185,52 +924,6 @@ jobs:
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Build the release package
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
-          cp "${{ 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/*
-
-      - name: Upload the release package
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
-          name: llama-bin-win-sycl-x64.zip
-
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-latest
@@ -1288,110 +981,12 @@ jobs:
             -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
-  # TODO: reuse windows-latest-cmake-hip instead of duplicating this job
-  windows-latest-cmake-hip-release:
-    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-    runs-on: windows-latest
-
-    strategy:
-      matrix:
-        gpu_target: [gfx1100, gfx1101, gfx1030]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
-
-      - name: Clone rocWMMA repository
-        id: clone_rocwmma
-        run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: windows-latest-cmake-hip-release
-          evict-old-files: 1d
-
-      - name: Install
-        id: depends
-        run: |
-          $ErrorActionPreference = "Stop"
-          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP SDK"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        id: verify
-        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_BUILD_TYPE=Release `
-            -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_HIP=ON `
-            -DGGML_RPC=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
-          cmake --build build -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
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - 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-${{ matrix.gpu_target }}.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-
   ios-xcode-build:
     runs-on: macos-latest
 
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: Build
         id: cmake_build
@@ -1418,32 +1013,6 @@ jobs:
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
-          name: llama-${{ steps.tag.outputs.name }}-xcframework
-
   android-build:
     runs-on: ubuntu-latest
 
@@ -1471,283 +1040,8 @@ jobs:
       - name: Build
         run: |
           cd examples/llama.android
-
           ./gradlew build --no-daemon
 
-  release:
-    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-
-    runs-on: ubuntu-latest
-
-    needs:
-      - ubuntu-cpu-cmake
-      - ubuntu-22-cmake-vulkan
-      - windows-latest-cmake
-      - windows-2019-cmake-cuda
-      - windows-latest-cmake-sycl
-      - windows-latest-cmake-hip-release
-      - macOS-latest-cmake-arm64
-      - macOS-latest-cmake-x64
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: release
-          evict-old-files: 1d
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Download artifacts
-        id: download-artifact
-        uses: actions/download-artifact@v4
-        with:
-          path: ./artifact
-
-      - name: Move artifacts
-        id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
-
-      - name: Create release
-        id: create_release
-        uses: ggml-org/action-create-release@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          tag_name: ${{ steps.tag.outputs.name }}
-
-      - name: Upload release
-        id: upload_release
-        uses: actions/github-script@v3
-        with:
-          github-token: ${{secrets.GITHUB_TOKEN}}
-          script: |
-            const path = require('path');
-            const fs = require('fs');
-            const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact/release')) {
-              if (path.extname(file) === '.zip') {
-                console.log('uploadReleaseAsset', file);
-                await github.repos.uploadReleaseAsset({
-                  owner: context.repo.owner,
-                  repo: context.repo.repo,
-                  release_id: release_id,
-                  name: file,
-                  data: await fs.readFileSync(`./artifact/release/${file}`)
-                });
-              }
-            }
-
-#  ubuntu-latest-gcc:
-#    runs-on: ubuntu-latest
-#
-#    strategy:
-#      matrix:
-#        build: [Debug, Release]
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#          sudo apt-get install cmake
-#
-#      - name: Configure
-#        run: cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-#
-#      - name: Build
-#        run: |
-#          make
-#
-#  ubuntu-latest-clang:
-#    runs-on: ubuntu-latest
-#
-#    strategy:
-#      matrix:
-#        build: [Debug, Release]
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#          sudo apt-get install cmake
-#
-#      - name: Configure
-#        run: cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
-#
-#      - name: Build
-#        run: |
-#          make
-#
-#  ubuntu-latest-gcc-sanitized:
-#    runs-on: ubuntu-latest
-#
-#    strategy:
-#      matrix:
-#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#          sudo apt-get install cmake
-#
-#      - name: Configure
-#        run: cmake . -DCMAKE_BUILD_TYPE=Debug -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON
-#
-#      - name: Build
-#        run: |
-#          make
-#
-#  windows:
-#    runs-on: windows-latest
-#
-#    strategy:
-#      matrix:
-#        build: [Release]
-#        arch: [Win32, x64]
-#        include:
-#          - arch: Win32
-#            s2arc: x86
-#          - arch: x64
-#            s2arc: x64
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Add msbuild to PATH
-#        uses: microsoft/setup-msbuild@v1
-#
-#      - name: Configure
-#        run: >
-#          cmake -S . -B ./build -A ${{ matrix.arch }}
-#          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-#
-#      - name: Build
-#        run: |
-#          cd ./build
-#          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
-#
-#      - name: Upload binaries
-#        uses: actions/upload-artifact@v4
-#        with:
-#          name: llama-bin-${{ matrix.arch }}
-#          path: build/bin/${{ matrix.build }}
-#
-#  windows-blas:
-#    runs-on: windows-latest
-#
-#    strategy:
-#      matrix:
-#        build: [Release]
-#        arch: [Win32, x64]
-#        blas: [ON]
-#        include:
-#          - arch: Win32
-#            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
-#            s2arc: x86
-#          - arch: x64
-#            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
-#            s2arc: x64
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Add msbuild to PATH
-#        uses: microsoft/setup-msbuild@v1
-#
-#      - name: Fetch OpenBLAS
-#        if: matrix.blas == 'ON'
-#        run: |
-#          C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
-#          7z x blas.zip -oblas -y
-#          copy blas/include/cblas.h .
-#          copy blas/include/openblas_config.h .
-#          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
-#
-#      - name: Configure
-#        run: >
-#          cmake -S . -B ./build -A ${{ matrix.arch }}
-#          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-#          -DLLAMA_SUPPORT_OPENBLAS=${{ matrix.blas }}
-#          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
-#
-#      - name: Build
-#        run: |
-#          cd ./build
-#          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
-#
-#      - name: Copy libopenblas.dll
-#        if: matrix.blas == 'ON'
-#        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
-#
-#      - name: Upload binaries
-#        if: matrix.blas == 'ON'
-#        uses: actions/upload-artifact@v4
-#        with:
-#          name: llama-blas-bin-${{ matrix.arch }}
-#          path: build/bin/${{ matrix.build }}
-#
-#  emscripten:
-#    runs-on: ubuntu-latest
-#
-#    strategy:
-#      matrix:
-#        build: [Release]
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        run: |
-#          wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
-#          tar -xvf master.tar.gz
-#          emsdk-master/emsdk update
-#          emsdk-master/emsdk install latest
-#          emsdk-master/emsdk activate latest
-#
-#      - name: Configure
-#        run: echo "tmp"
-#
-#      - name: Build
-#        run: |
-#          pushd emsdk-master
-#          source ./emsdk_env.sh
-#          popd
-#          emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-#          make
-
   openEuler-latest-cmake-cann:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:

.github/workflows/docker.yml

@@ -36,10 +36,13 @@ jobs:
       matrix:
         config:
           # Multi-stage build
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
+          # Note: the arm64 images are failing, which prevents the amd64 images from being built
+          # https://github.com/ggml-org/llama.cpp/issues/11888
+          #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { 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 }
-          - { 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

@@ -0,0 +1,742 @@
+name: Release
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/release.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+  CMAKE_ARGS: "-DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON"
+
+jobs:
+  macOS-arm64:
+    runs-on: macos-14
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-cmake-arm64
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+          brew install curl
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_RPC=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          name: llama-bin-macos-arm64.zip
+
+  macOS-x64:
+    runs-on: macos-13
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-cmake-x64
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+          brew install curl
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
+          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
+          cmake -B build \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_METAL=OFF \
+            -DGGML_RPC=ON
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          name: llama-bin-macos-x64.zip
+
+  ubuntu-22-cpu:
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-22.04
+          - build: 'arm64'
+            os: ubuntu-22.04-arm
+
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: ubuntu-cpu-cmake
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
+          name: llama-bin-ubuntu-${{ matrix.build }}.zip
+
+  ubuntu-22-vulkan:
+    runs-on: ubuntu-22.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: ubuntu-22-cmake-vulkan
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DGGML_VULKAN=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          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
+        shell: cmd
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
+          cmake -S . -B build -G "Ninja Multi-Config" ^
+            -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
+            -DGGML_OPENMP=ON ^
+            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%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\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          7z a llama-bin-win-cpu-${{ 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
+
+    env:
+      OPENBLAS_VERSION: 0.3.23
+      VULKAN_VERSION: 1.4.309.0
+
+    strategy:
+      matrix:
+        include:
+          - 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="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
+            target: 'ggml-opencl'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
+          variant: ccache
+          evict-old-files: 1d
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.backend == 'vulkan' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        if: ${{ matrix.backend == 'opencl-adreno' && matrix.arch == 'arm64' }}
+        run: |
+          git clone https://github.com/KhronosGroup/OpenCL-Headers
+          cd OpenCL-Headers
+          cmake -B build `
+            -DBUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build build --target install
+          git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+          cd OpenCL-ICD-Loader
+          cmake -B build-arm64-release `
+            -A arm64 `
+            -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build build-arm64-release --target install --config release
+
+      - name: Build
+        id: cmake_build
+        run: |
+          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 ${{ matrix.target }}
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
+          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
+
+  windows-cuda:
+    runs-on: windows-2019
+
+    strategy:
+      matrix:
+        cuda: ['12.4', '11.7']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-cuda-${{ matrix.cuda }}
+          variant: ccache
+          evict-old-files: 1d
+
+      - name: Install Cuda Toolkit
+        uses: ./.github/actions/windows-setup-cuda
+        with:
+          cuda_version: ${{ matrix.cuda }}
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        shell: cmd
+        run: |
+          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          cmake -S . -B build -G "Ninja Multi-Config" ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
+            -DGGML_CUDA=ON ^
+            -DLLAMA_CURL=OFF
+          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+
+      - name: Copy and pack Cuda runtime
+        run: |
+          echo "Cuda install location: ${{ env.CUDA_PATH }}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{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 $dst\*
+
+      - name: Upload Cuda runtime
+        uses: actions/upload-artifact@v4
+        with:
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+
+  windows-sycl:
+    runs-on: windows-latest
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
+      WINDOWS_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
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-sycl
+          variant: ccache
+          evict-old-files: 1d
+
+      - name: Install
+        run:  |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Build
+        id: cmake_build
+        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
+        run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
+
+          echo "cp oneAPI running time dll files to ./build/bin done"
+          7z a llama-bin-win-sycl-x64.zip ./build/bin/*
+
+      - name: Upload the release package
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-sycl-x64.zip
+          name: llama-bin-win-sycl-x64.zip
+
+  windows-hip:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        include:
+          - name: "radeon"
+            gpu_targets: "gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Clone rocWMMA repository
+        id: clone_rocwmma
+        run: |
+          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-hip-${{ matrix.name }}-x64
+          evict-old-files: 1d
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        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/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=OFF `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DGGML_HIP=ON `
+            -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: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+
+  ios-xcode-build:
+    runs-on: macos-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          ./build-xcframework.sh
+
+      - name: Build Xcode project
+        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
+          name: llama-${{ steps.tag.outputs.name }}-xcframework
+
+  release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+
+    # Fine-grant permission
+    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+    permissions:
+        contents: write # for creating release
+
+    runs-on: ubuntu-latest
+
+    needs:
+      - windows
+      - windows-cpu
+      - windows-cuda
+      - windows-sycl
+      - windows-hip
+      - ubuntu-22-cpu
+      - ubuntu-22-vulkan
+      - macOS-arm64
+      - macOS-x64
+      - ios-xcode-build
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+          merge-multiple: true
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: |
+          mkdir -p release
+
+          echo "Adding CPU backend files to existing zips..."
+          for arch in x64 arm64; do
+            cpu_zip="artifact/llama-bin-win-cpu-${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
+        uses: ggml-org/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ steps.tag.outputs.name }}
+
+      - name: Upload release
+        id: upload_release
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./release')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: 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

CMakeLists.txt

@@ -252,20 +252,3 @@ configure_file(cmake/llama.pc.in
 
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
-
-#
-# copy the license files
-#
-
-# Check if running in GitHub Actions
-if(DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
-    message(STATUS "Running inside GitHub Actions - copying license files")
-
-    # Copy all files from licenses/ to build/bin/
-    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
-    foreach(LICENSE_FILE ${LICENSE_FILES})
-        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
-        configure_file(${LICENSE_FILE} "${CMAKE_BINARY_DIR}/bin/${FILENAME}" COPYONLY)
-    endforeach()
-endif()
-

CMakePresets.json

@@ -38,15 +38,6 @@
         }
     },
 
-    {
-        "name": "arm64-windows-msvc", "hidden": true,
-        "architecture": { "value": "arm64",    "strategy": "external" },
-        "toolset":      { "value": "host=x64", "strategy": "external" },
-        "cacheVariables": {
-            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
-        }
-    },
-
     {
         "name": "arm64-windows-llvm", "hidden": true,
         "architecture": { "value": "arm64",    "strategy": "external" },
@@ -73,10 +64,6 @@
     { "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
     { "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
 
-    { "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
-    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
-    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
-
     { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] },
     { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] },
     { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] },

Makefile

@@ -1187,11 +1187,6 @@ llama-cli: tools/main/main.cpp \
 	@echo '====  Run ./llama-cli -h for help.  ===='
 	@echo
 
-llama-infill: examples/infill/infill.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 llama-run: tools/run/run.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

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

@@ -60,12 +60,16 @@ add_library(${TARGET} STATIC
     base64.hpp
     chat.cpp
     chat.h
+    chat-parser.cpp
+    chat-parser.h
     common.cpp
     common.h
     console.cpp
     console.h
     json-schema-to-grammar.cpp
     json.hpp
+    json-partial.h
+    json-partial.cpp
     llguidance.cpp
     log.cpp
     log.h
@@ -73,6 +77,8 @@ add_library(${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 +125,8 @@ 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 ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE
@@ -144,3 +150,27 @@ endif ()
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
+
+
+#
+# copy the license files
+#
+
+# Check if running in GitHub Actions
+if (DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
+    message(STATUS "Running inside GitHub Actions - copying license files")
+
+    # Copy all files from licenses/ to build/bin/
+    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
+    foreach(LICENSE_FILE ${LICENSE_FILES})
+        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
+        add_custom_command(
+            POST_BUILD
+            TARGET ${TARGET}
+            COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                "${LICENSE_FILE}"
+                "$<TARGET_FILE_DIR:llama>/${FILENAME}"
+            COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
+        message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}")
+    endforeach()
+endif()

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) {
@@ -1283,7 +1283,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.use_color = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
     add_opt(common_arg(
         {"-t", "--threads"}, "N",
         string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads),
@@ -1416,7 +1416,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"-n", "--predict", "--n-predict"}, "N",
         string_format(
-            ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL
+            ex == LLAMA_EXAMPLE_MAIN
                 ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)"
                 : "number of tokens to predict (default: %d, -1 = infinity)",
             params.n_predict),
@@ -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"),
@@ -1655,7 +1663,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.input_prefix = value;
             params.enable_chat_template = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--in-suffix"}, "STRING",
         "string to suffix after user inputs with (default: empty)",
@@ -1663,14 +1671,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.input_suffix = value;
             params.enable_chat_template = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--no-warmup"},
         "skip warming up the model with an empty run",
         [](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(
@@ -1680,7 +1688,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.spm_infill = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_INFILL}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--samplers"}, "SAMPLERS",
         string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
@@ -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",
@@ -2097,13 +2098,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.cache_type_v = kv_cache_type_from_str(value);
         }
     ).set_env("LLAMA_ARG_CACHE_TYPE_V"));
-    add_opt(common_arg(
-        {"--perplexity", "--all-logits"},
-        string_format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"),
-        [](common_params & params) {
-            params.logits_all = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--hellaswag"},
         "compute HellaSwag score over random tasks from datafile supplied with -f",
@@ -2211,39 +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",
@@ -2443,6 +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)",
@@ -2584,7 +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),
@@ -2634,13 +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",
@@ -2839,15 +2848,24 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
-        "reasoning format (default: deepseek; allowed values: deepseek, none)\n"
-        "controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).\n"
-        "only supported for non-streamed responses",
+        "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
+        "- none: leaves thoughts unparsed in `message.content`\n"
+        "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
+        "(default: deepseek)",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
-            else { std::invalid_argument("invalid value"); }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
+    add_opt(common_arg(
+        {"--reasoning-budget"}, "N",
+        "controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)",
+        [](common_params & params, int value) {
+            if (value != 0 && value != -1) { throw std::invalid_argument("invalid value"); }
+            params.reasoning_budget = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK_BUDGET"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(
@@ -2859,7 +2877,7 @@ 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(
@@ -2872,6 +2890,16 @@ 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),
@@ -2892,7 +2920,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.simple_io = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--positive-file"}, "FNAME",
         string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()),
@@ -2936,7 +2964,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             /**/ if (value == "jsonl") { params.batched_bench_output_jsonl = true; }
             else if (value == "md") { params.batched_bench_output_jsonl = false; }
-            else { std::invalid_argument("invalid value"); }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_BENCH}));
     add_opt(common_arg(

common/chat-parser.cpp

@@ -0,0 +1,376 @@
+#include "chat-parser.h"
+#include "common.h"
+#include "log.h"
+#include "regex-partial.h"
+
+#include <optional>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
+    : input_(input), is_partial_(is_partial), syntax_(syntax)
+{
+    result_.role = "assistant";
+
+    while (true) {
+        std::string id = std::to_string(std::rand());
+        if (input.find(id) == std::string::npos) {
+            healing_marker_ = id;
+            break;
+        }
+    }
+}
+
+std::string common_chat_msg_parser::str(const common_string_range & rng) const {
+    GGML_ASSERT(rng.begin <= rng.end);
+    return input_.substr(rng.begin, rng.end - rng.begin);
+}
+
+void common_chat_msg_parser::add_content(const std::string &content) {
+    result_.content += content;
+}
+
+void common_chat_msg_parser::add_reasoning_content(const std::string &reasoning_content) {
+    result_.reasoning_content += reasoning_content;
+}
+
+bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::string & id, const std::string & arguments) {
+    if (name.empty()) {
+        return false;
+    }
+
+    common_chat_tool_call tool_call;
+    tool_call.name = name;
+    tool_call.arguments = arguments;
+    tool_call.id = id;
+
+    // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
+    result_.tool_calls.emplace_back(tool_call);
+    return true;
+}
+bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
+    std::string name = tool_call.contains("name") ? tool_call.at("name") : "";
+    std::string id = tool_call.contains("id") ? tool_call.at("id") : "";
+    std::string arguments = tool_call.contains("arguments") ? tool_call.at("arguments") : "";
+    return add_tool_call(name, id, arguments);
+}
+
+bool common_chat_msg_parser::add_tool_calls(const json & arr) {
+    for (const auto & item : arr) {
+        if (!add_tool_call(item)) {
+            return false;
+        }
+    }
+    return true;
+}
+void common_chat_msg_parser::finish() {
+    if (!is_partial_ && pos_ != input_.size()) {
+        throw std::runtime_error("Unexpected content at end of input");// + input_.substr(pos_));
+    }
+}
+
+bool common_chat_msg_parser::consume_spaces() {
+    const auto length = input_.size();
+    auto consumed = false;
+    while (pos_ < length && std::isspace(input_[pos_])) {
+        ++pos_;
+        consumed = true;
+    }
+    return consumed;
+}
+
+bool common_chat_msg_parser::try_consume_literal(const std::string & literal) {
+    auto pos = pos_;
+    for (auto i = 0u; i < literal.size(); ++i) {
+        if (pos >= input_.size()) {
+            return false;
+        }
+        if (input_[pos] != literal[i]) {
+            return false;
+        }
+        ++pos;
+    }
+    pos_ = pos;
+    return true;
+}
+
+std::optional<common_chat_msg_parser::find_regex_result>  common_chat_msg_parser::try_find_literal(const std::string & literal) {
+    auto idx = input_.find(literal, pos_);
+    if (idx != std::string::npos) {
+        find_regex_result res;
+        res.prelude = input_.substr(pos_, idx - pos_);
+        auto end = idx + literal.size();
+        res.groups.emplace_back(common_string_range{idx, end});
+        move_to(end);
+        return res;
+    }
+    if (is_partial_) {
+        idx = string_find_partial_stop(input_, literal);
+        if (idx != std::string::npos && idx >= pos_) {
+            find_regex_result res;
+            res.prelude = input_.substr(pos_, idx - pos_);
+            auto end = input_.size();
+            res.groups.emplace_back(common_string_range{idx, end});
+            move_to(end);
+            return res;
+        }
+    }
+    return std::nullopt;
+}
+
+void common_chat_msg_parser::consume_literal(const std::string & literal) {
+    if (!try_consume_literal(literal)) {
+        throw common_chat_msg_partial_exception(literal);
+    }
+}
+
+bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think, const std::string & end_think) {
+    auto handle_reasoning = [&](const std::string & reasoning, bool closed) {
+        auto stripped_reasoning = string_strip(reasoning);
+        if (stripped_reasoning.empty()) {
+            return;
+        }
+        if (syntax_.reasoning_in_content) {
+            add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "<think>" : start_think);
+            add_content(stripped_reasoning);
+            if (closed) {
+                add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "</think>" : end_think);
+            }
+        } else {
+            add_reasoning_content(stripped_reasoning);
+        }
+    };
+    if (syntax_.reasoning_format != COMMON_REASONING_FORMAT_NONE) {
+        if (syntax_.thinking_forced_open || try_consume_literal(start_think)) {
+            if (auto res = try_find_literal(end_think)) {
+                handle_reasoning(res->prelude, /* closed */ true);
+                consume_spaces();
+                return true;
+            }
+            auto rest = consume_rest();
+            if (!rest.empty()) {
+                handle_reasoning(rest, /* closed */ !is_partial());
+            }
+            if (!syntax_.thinking_forced_open) {
+                throw common_chat_msg_partial_exception(end_think);
+            }
+            return true;
+        }
+    }
+    return false;
+}
+
+std::string common_chat_msg_parser::consume_rest() {
+    auto rest = input_.substr(pos_);
+    pos_ = input_.size();
+    return rest;
+}
+
+// Tries to find the regex, consumes it (pos right after it) and gives the prelude (right before it) and the groups to the callback.
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_find_regex(const common_regex & regex, size_t from) {
+    auto m = regex.search(input_, from == std::string::npos ? pos_ : from);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    auto prelude = input_.substr(pos_, m.groups[0].begin - pos_);
+    pos_ = m.groups[0].end;
+
+    return find_regex_result{prelude, m.groups};
+}
+
+common_chat_msg_parser::find_regex_result common_chat_msg_parser::consume_regex(const common_regex & regex) {
+    if (auto result = try_consume_regex(regex)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception(regex.str());
+}
+
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_consume_regex(const common_regex & regex) {
+    auto m = regex.search(input_, pos_);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    if (m.groups[0].begin != pos_) {
+        // Didn't match at the current position.
+        return std::nullopt;
+    }
+    pos_ = m.groups[0].end;
+
+    return find_regex_result {
+        /* .prelude = */ "",
+        m.groups,
+    };
+}
+
+std::optional<common_json> common_chat_msg_parser::try_consume_json() {
+    auto it = input_.cbegin() + pos_;
+    const auto end = input_.cend();
+    common_json result;
+    if (!common_json_parse(it, end, healing_marker_, result)) {
+        return std::nullopt;
+    }
+    pos_ = std::distance(input_.cbegin(), it);
+    if (result.healing_marker.marker.empty()) {
+        // No healing marker, just return the parsed json
+        return result;
+    }
+    if (!is_partial()) {
+        throw common_chat_msg_partial_exception("JSON");
+    }
+    return result;
+}
+
+common_json common_chat_msg_parser::consume_json() {
+    if (auto result = try_consume_json()) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+common_chat_msg_parser::consume_json_result common_chat_msg_parser::consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    if (auto result = try_consume_json_with_dumped_args(args_paths, content_paths)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parser::try_consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    auto partial = try_consume_json();
+    if (!partial) {
+        return std::nullopt;
+    }
+    auto is_arguments_path = [&](const std::vector<std::string> & path) {
+        return std::find(args_paths.begin(), args_paths.end(), path) != args_paths.end();
+    };
+    auto is_content_path = [&](const std::vector<std::string> & path) {
+        return std::find(content_paths.begin(), content_paths.end(), path) != content_paths.end();
+    };
+
+    if (partial->healing_marker.marker.empty()) {
+        if (args_paths.empty()) {
+            // No arguments to dump, and JSON was parsed fully.
+            return consume_json_result {
+                partial->json,
+                /* .is_partial = */ false,
+            };
+        }
+        if (is_arguments_path({})) {
+            // Entire JSON is the arguments and was parsed fully.
+            return consume_json_result {
+                partial->json.dump(),
+                /* .is_partial = */ false,
+            };
+        }
+    }
+
+    LOG_DBG("Parsed partial JSON: %s (json_healing_marker: %s)\n", partial->json.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+
+    auto found_healing_marker = false;
+    std::vector<std::string> path;
+    std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
+        if (is_arguments_path(path)) {
+            auto arguments = j.dump();
+            if (is_partial() && !partial->healing_marker.marker.empty()) {
+                auto idx = arguments.find(partial->healing_marker.json_dump_marker);
+                if (idx != std::string::npos) {
+                    arguments.resize(idx);
+                    found_healing_marker = true;
+                }
+                if (arguments == "\"") {
+                    // This happens because of completing `:"$magic` after `"arguments"`
+                    arguments = "";
+                }
+            }
+            return arguments;
+        }
+        if (is_content_path(path)) {
+            if (!j.is_string()) {
+                throw std::runtime_error("Content path must be a string");
+            }
+            std::string str = j;
+            auto idx = str.find(partial->healing_marker.marker); // not using json_dump_marker as we're inside a string
+            if (idx != std::string::npos) {
+                str.resize(idx);
+                found_healing_marker = true;
+            }
+            return str;
+        }
+        if (j.is_object()) {
+            auto obj = json::object();
+            for (const auto & p : j.items()) {
+                const auto & key = p.key();
+                const auto & value = p.value();
+                const std::string key_str = key; // NOLINT
+                auto idx = key_str.find(healing_marker_);
+                if (idx != std::string::npos) {
+                    found_healing_marker = true;
+                    break;
+                }
+                path.push_back(key_str);
+                if (value.is_string()) {
+                    const std::string value_str = value;
+                    if (value_str.find(healing_marker_) != std::string::npos) {
+                        found_healing_marker = true;
+                        if (is_content_path(path)) {
+                            if (partial->healing_marker.marker == partial->healing_marker.json_dump_marker) {
+                                // The healing occurred inside the string: good. Otherwise we just ditch the entire key/value pair.
+                                obj[key] = remove_unsupported_healings_and_dump_args(value);
+                            }
+                        }
+                        break;
+                    }
+                    obj[key] = value;
+                } else {
+                    obj[key] = remove_unsupported_healings_and_dump_args(value);
+                }
+                path.pop_back();
+            }
+            return obj;
+        }
+        if (j.is_array()) {
+            auto arr = json::array();
+            for (const auto & value : j) {
+                if (value.is_string()) {
+                    std::string str = value;
+                    auto idx = str.find(healing_marker_);
+                    if (idx != std::string::npos) {
+                        // Don't heal array values that aren't in the arguments.
+                        found_healing_marker = true;
+                        break;
+                    }
+                }
+                arr.push_back(remove_unsupported_healings_and_dump_args(value));
+            }
+            return arr;
+        }
+        return j;
+    };
+
+    auto cleaned = remove_unsupported_healings_and_dump_args(partial->json);
+    LOG_DBG("Cleaned up JSON %s to %s (json_healing_marker : '%s')\n", partial->json.dump().c_str(), cleaned.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+    return consume_json_result {
+        cleaned,
+        /* .is_partial = */ found_healing_marker,
+    };
+}

common/chat-parser.h

@@ -0,0 +1,116 @@
+#pragma once
+
+#include "chat.h"
+#include "json-partial.h"
+#include "json.hpp"
+#include "regex-partial.h"
+
+#include <optional>
+#include <string>
+#include <vector>
+
+class common_chat_msg_partial_exception : public std::runtime_error {
+  public:
+    common_chat_msg_partial_exception(const std::string & message) : std::runtime_error(message) {}
+};
+
+class common_chat_msg_parser {
+    std::string input_;
+    bool is_partial_;
+    common_chat_syntax syntax_;
+    std::string healing_marker_;
+
+    size_t pos_ = 0;
+    common_chat_msg result_;
+
+  public:
+    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+    const std::string & input() const { return input_; }
+    size_t pos() const { return pos_; }
+    const std::string & healing_marker() const { return healing_marker_; }
+    const bool & is_partial() const { return is_partial_; }
+    const common_chat_msg & result() const { return result_; }
+
+    void move_to(size_t pos) {
+        if (pos > input_.size()) {
+            throw std::runtime_error("Invalid position!");
+        }
+        pos_ = pos;
+    }
+    void move_back(size_t n) {
+        if (pos_ < n) {
+            throw std::runtime_error("Can't move back that far!");
+        }
+        pos_ -= n;
+    }
+
+    // Get the substring of the input at the given range
+    std::string str(const common_string_range & rng) const;
+
+    // Appends to the result.content field
+    void add_content(const std::string & content);
+
+    // Appends to the result.reasoning_content field
+    void add_reasoning_content(const std::string & reasoning_content);
+
+    // Adds a tool call to the result. If the tool call is too incomplete (e.g. name empty), it won't add anything.
+    bool add_tool_call(const std::string & name, const std::string & id, const std::string & arguments);
+
+    // Adds a tool call using the "name", "id" and "arguments" fields of the json object
+    bool add_tool_call(const nlohmann::ordered_json & tool_call);
+
+    // Adds an array of tool calls using their "name", "id" and "arguments" fields.
+    bool add_tool_calls(const nlohmann::ordered_json & arr);
+
+    void finish();
+
+    bool consume_spaces();
+
+    void consume_literal(const std::string & literal);
+
+    bool try_parse_reasoning(const std::string & start_think, const std::string & end_think);
+
+    std::string consume_rest();
+
+    struct find_regex_result {
+        std::string prelude;
+        std::vector<common_string_range> groups;
+    };
+
+    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos);
+
+    bool try_consume_literal(const std::string & literal);
+
+    std::optional<find_regex_result> try_find_literal(const std::string & literal);
+
+    find_regex_result consume_regex(const common_regex & regex);
+
+    std::optional<find_regex_result> try_consume_regex(const common_regex & regex);
+
+    std::optional<common_json> try_consume_json();
+    common_json consume_json();
+
+    struct consume_json_result {
+        nlohmann::ordered_json value;
+        bool is_partial;
+    };
+
+    /*
+        Consume (possibly partial) json and converts specific subtrees to (possibly truncated) JSON strings.
+
+        By default, object keys can't be truncated, nor can string values (their corresponding key is removed,
+        e.g. `{"foo": "bar", "baz": "b` -> `{"foo": "bar"}`
+
+        But one can allow subpaths to be kept truncated, and possibly json-dumped to truncated json strings
+        - with `content_paths={{"foo"}}` -> `{"foo": "b` -> {"foo": "b"}`
+        - with `args_paths={{"foo"}}` -> `{"foo": {"b` -> `{"foo": "{b"}`
+    */
+    consume_json_result consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+    std::optional<consume_json_result> try_consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+};

common/chat.h

@@ -3,6 +3,8 @@
 #pragma once
 
 #include "common.h"
+#include <functional>
+#include <chrono>
 #include <string>
 #include <vector>
 
@@ -12,11 +14,19 @@ struct common_chat_tool_call {
     std::string name;
     std::string arguments;
     std::string id;
+
+    bool operator==(const common_chat_tool_call & other) const {
+        return name == other.name && arguments == other.arguments && id == other.id;
+    }
 };
 
 struct common_chat_msg_content_part {
     std::string type;
     std::string text;
+
+    bool operator==(const common_chat_msg_content_part & other) const {
+        return type == other.type && text == other.text;
+    }
 };
 
 struct common_chat_msg {
@@ -27,6 +37,51 @@ struct common_chat_msg {
     std::string reasoning_content;
     std::string tool_name;
     std::string tool_call_id;
+
+    template <class T> T to_json_oaicompat() const;
+
+    bool empty() const {
+        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
+    }
+    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+        for (auto i = 0u; i < tool_calls.size(); i++) {
+            if (ids_cache.size() <= i) {
+                auto id = tool_calls[i].id;
+                if (id.empty()) {
+                    id = gen_tool_call_id();
+                }
+                ids_cache.push_back(id);
+            }
+            tool_calls[i].id = ids_cache[i];
+        }
+    }
+    bool operator==(const common_chat_msg & other) const {
+        return role == other.role
+            && content == other.content
+            && content_parts == other.content_parts
+            && tool_calls == other.tool_calls
+            && reasoning_content == other.reasoning_content
+            && tool_name == other.tool_name
+            && tool_call_id == other.tool_call_id;
+    }
+    bool operator!=(const common_chat_msg & other) const {
+        return !(*this == other);
+    }
+};
+
+struct common_chat_msg_diff {
+    // std::string reasoning_content_delta;
+    std::string content_delta;
+    size_t tool_call_index = std::string::npos;
+    common_chat_tool_call tool_call_delta;
+
+    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg);
+
+    bool operator==(const common_chat_msg_diff & other) const {
+        return content_delta == other.content_delta
+        && tool_call_index == other.tool_call_index
+        && tool_call_delta == other.tool_call_delta;
+    }
 };
 
 struct common_chat_tool {
@@ -48,14 +103,11 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_LLAMA_3_X,
     COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
     COMMON_CHAT_FORMAT_DEEPSEEK_R1,
-    COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING,
     COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
-    COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
-    COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };
@@ -70,7 +122,9 @@ struct common_chat_templates_inputs {
     std::vector<common_chat_tool> tools;
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
-    bool extract_reasoning     = true;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    bool enable_thinking = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
 struct common_chat_params {
@@ -78,11 +132,20 @@ struct common_chat_params {
     std::string                         prompt;
     std::string                         grammar;
     bool                                grammar_lazy = false;
+    bool                                thinking_forced_open = false;
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
 };
 
+struct common_chat_syntax {
+    common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
+    // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
+    bool                     reasoning_in_content  = false;
+    bool                     thinking_forced_open  = false;
+};
+
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
 bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
 
@@ -119,8 +182,9 @@ std::string common_chat_format_example(
     const struct common_chat_templates * tmpls,
     bool use_jinja);
 
-std::string               common_chat_format_name(common_chat_format format);
-common_chat_msg           common_chat_parse(      const std::string & input, common_chat_format format);
+const char*               common_chat_format_name(common_chat_format format);
+const char*               common_reasoning_format_name(common_reasoning_format format);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
@@ -133,3 +197,5 @@ template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common
 // T can be std::string containing JSON or nlohmann::ordered_json
 template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
 template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+
+template <class T> T common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);

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");
@@ -830,7 +849,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -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;
 }
 
@@ -1096,7 +1118,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_threads         = params.cpuparams.n_threads;
     cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
                                 params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
-    cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
     cparams.rope_freq_base    = params.rope_freq_base;
@@ -1114,6 +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;
@@ -1306,81 +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
 //
@@ -1565,3 +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>
 
@@ -66,7 +67,6 @@ enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_MAIN,
-    LLAMA_EXAMPLE_INFILL,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -76,7 +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,
@@ -96,6 +96,7 @@ enum common_sampler_type {
     COMMON_SAMPLER_TYPE_XTC         = 8,
     COMMON_SAMPLER_TYPE_INFILL      = 9,
     COMMON_SAMPLER_TYPE_PENALTIES   = 10,
+    COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -114,7 +115,7 @@ enum common_grammar_trigger_type {
     COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
     COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
     COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
 };
 
 struct common_grammar_trigger {
@@ -161,6 +162,7 @@ struct common_params_sampling {
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
+        COMMON_SAMPLER_TYPE_TOP_N_SIGMA,
         COMMON_SAMPLER_TYPE_TOP_K,
         COMMON_SAMPLER_TYPE_TYPICAL_P,
         COMMON_SAMPLER_TYPE_TOP_P,
@@ -321,17 +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 logits_all        = false; // return logits for all tokens in the batch
     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 +368,8 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int reasoning_budget = -1;
+    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;
 
@@ -409,6 +413,7 @@ 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 +429,11 @@ 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 +511,9 @@ 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 +622,6 @@ 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 +663,9 @@ 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/json-partial.cpp

@@ -0,0 +1,255 @@
+#include <json-partial.h>
+#include "ggml.h"
+#include "log.h"
+#include <string>
+
+#include <json.hpp>
+
+using json = nlohmann::ordered_json;
+
+enum common_json_stack_element_type {
+    COMMON_JSON_STACK_ELEMENT_OBJECT,
+    COMMON_JSON_STACK_ELEMENT_KEY,
+    COMMON_JSON_STACK_ELEMENT_ARRAY,
+};
+
+struct common_json_stack_element {
+    common_json_stack_element_type type;
+    std::string key;
+};
+
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    std::string::const_iterator it = input.begin();
+    const auto end = input.end();
+    return common_json_parse(it, end, healing_marker, out);
+}
+
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    // // https://json.nlohmann.me/features/parsing/sax_interface/
+    struct json_error_locator : public nlohmann::json_sax<json> {
+        std::size_t position;
+        bool found_error;
+        std::string last_token;
+        std::string exception_message;
+        std::vector<common_json_stack_element> stack;
+
+        json_error_locator() : position(0), found_error(false) {}
+
+        bool parse_error(std::size_t position, const std::string & last_token, const json::exception & ex) override { // NOLINT
+            this->position = position - 1;
+            this->found_error = true;
+            this->last_token = last_token;
+            this->exception_message = ex.what();
+            return false;
+        }
+        void close_value() {
+            if (!stack.empty() && (stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY)) {
+                stack.pop_back();
+            }
+        }
+        bool null() override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool boolean(bool) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_integer(number_integer_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_unsigned(number_unsigned_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_float(number_float_t, const string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool string(string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool binary(binary_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool start_object(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_OBJECT, ""});
+            return true;
+        }
+        bool end_object() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+        bool key(string_t & key) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_KEY, key});
+            return true;
+        }
+        bool start_array(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_ARRAY, ""});
+            return true;
+        }
+        bool end_array() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+    };
+    json_error_locator err_loc;
+    auto start = it;
+    json::sax_parse(it, end, &err_loc);
+
+    if (err_loc.found_error) {
+        it = start;
+        auto temptative_end = it + err_loc.position;
+        // LOG_DBG("Error at position %zu (is_end = %s): %s\n", err_loc.position, temptative_end == end ? "true" : "false", err_loc.exception_message.c_str());
+
+        auto input = std::string(it, temptative_end);
+        try {
+            out.json = json::parse(input);
+            // out.json = json::parse(it, temptative_end);
+            it = temptative_end;
+            return true;
+        } catch (const std::exception & ex) {
+            // No, needs healing.
+            LOG_DBG("Failed to parse up to error: %s: <<<%s>>>\n", ex.what(), std::string(it, temptative_end).c_str());
+        }
+        auto can_parse = [](const std::string & str) {
+            try {
+                auto _ = json::parse(str); // NOLINT
+                return true;
+            } catch (const std::exception &) {
+                return false;
+            }
+        };
+        if (!healing_marker.empty() && !err_loc.stack.empty()) {
+            std::string str(it, temptative_end);
+            auto last_non_sp_pos = str.find_last_not_of(" \n\r\t");
+            if (last_non_sp_pos == std::string::npos) {
+                throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+            }
+            auto last_non_sp_char = str[last_non_sp_pos];
+            // Used to detect stops on a number, which may not be complete.
+            auto was_maybe_number = [&]() {
+                if (!str.empty() && std::isspace(str.back())) {
+                    return false;
+                }
+                return std::isdigit(last_non_sp_char) ||
+                    last_non_sp_char == '.' ||
+                    last_non_sp_char == 'e' ||
+                    last_non_sp_char == 'E' ||
+                    last_non_sp_char == '-';
+            };
+
+            std::string closing;
+            for (size_t i = err_loc.stack.size(); i > 0; i--) {
+                auto & el = err_loc.stack[i - 1];
+                if (el.type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                    closing += "}";
+                } else if (el.type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                    closing += "]";
+                } else if (el.type != COMMON_JSON_STACK_ELEMENT_KEY) {
+                    throw std::runtime_error("Unexpected stack element type");
+                }
+            }
+
+            const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
+
+            if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
+                // We're inside an object value
+                if (last_non_sp_char == ':' && can_parse(str + "1" + closing)) {
+                    // Was about to create an object value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + ": 1" + closing)) {
+                    str += (out.healing_marker.json_dump_marker = ":\"" + magic_seed) + "\"" + closing;
+                } else if (last_non_sp_char == '{' && can_parse(str + closing)) {
+                    // Was about to create an object
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an object value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an object value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else {
+                    // find last :
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+                    }
+                    // Cutting back to opening : for object value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                if ((last_non_sp_char == ',' || last_non_sp_char == '[') && can_parse(str + "1" + closing)) {
+                    // Was about to create an array value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an array value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an array value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
+                    // Had just finished a value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
+                } else {
+                    auto last_pos = str.find_last_of("[,");
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON array stopped in an unknown location");
+                    }
+                    // Cutting back to last [ or , for array value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                if ((last_non_sp_char == '{' && can_parse(str + closing)) ||
+                        (last_non_sp_char == ',' && can_parse(str + "\"\": 1" + closing))) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ",\"\": 1" + closing)) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\": 1" + closing)) {
+                    // Was inside an object key string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\": 1" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
+                    // Was inside an object key string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
+                } else {
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+                    }
+                    // fprintf(stderr, "Cutting back to last : for object key+value\n");
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else {
+                throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+            }
+            // fprintf(stderr, "HEALED:\nSTRING <<<\n%s\n>>>\n\nmagic_cut: <<<\n%s\n>>>\n\n", str.c_str(), out.healing_marker.json_dump_marker.c_str());
+            out.json = json::parse(str);
+            it = temptative_end;
+            return true;
+        }
+        // TODO: handle unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
+        // fprintf(stderr, "Closing: TODO\n");
+        return false;
+    }
+    out.json = json::parse(it, end);
+    it = end;
+    return true;
+}

common/json-partial.h

@@ -0,0 +1,37 @@
+#pragma once
+#include <json.hpp>
+
+// Healing marker (empty if the JSON was fully parsed / wasn't healed).
+struct common_healing_marker {
+    // Raw marker.
+    std::string marker;
+
+    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
+    std::string json_dump_marker;
+};
+
+// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
+struct common_json {
+    nlohmann::ordered_json json;
+
+    common_healing_marker healing_marker;
+};
+
+// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
+//
+// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
+// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
+// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
+//
+// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out);
+
+// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out);

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);

common/sampling.cpp

@@ -1,6 +1,7 @@
 #include "sampling.h"
 
 #include "common.h"
+#include "log.h"
 
 #include <cmath>
 #include <unordered_map>
@@ -160,7 +161,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
-        std::vector<std::string> patterns_at_start;
+        std::vector<std::string> trigger_patterns;
         std::vector<std::string> patterns_anywhere;
         std::vector<llama_token> trigger_tokens;
         for (const auto & trigger : params.grammar_triggers) {
@@ -172,10 +173,13 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
-                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START:
                 {
-                    const auto & pattern = trigger.value;
-                    (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern);
+                    patterns_anywhere.push_back(trigger.value);
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
+                {
+                    trigger_patterns.push_back(trigger.value);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -189,10 +193,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
         }
 
-        std::vector<std::string> trigger_patterns;
-        if (!patterns_at_start.empty()) {
-            trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*");
-        }
         if (!patterns_anywhere.empty()) {
             trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
         }
@@ -229,51 +229,48 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 params.logit_bias.data()));
 
     if (params.mirostat == 0) {
-        if (params.top_n_sigma >= 0) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
-        } else {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case COMMON_SAMPLER_TYPE_DRY:
-                        {
-                            std::vector<const char *> c_breakers;
-                            c_breakers.reserve(params.dry_sequence_breakers.size());
-                            for (const auto & str : params.dry_sequence_breakers) {
-                                c_breakers.push_back(str.c_str());
-                            }
-
-                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+        for (const auto & cnstr : params.samplers) {
+            switch (cnstr) {
+                case COMMON_SAMPLER_TYPE_DRY:
+                    {
+                        std::vector<const char *> c_breakers;
+                        c_breakers.reserve(params.dry_sequence_breakers.size());
+                        for (const auto & str : params.dry_sequence_breakers) {
+                            c_breakers.push_back(str.c_str());
                         }
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_XTC:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    case COMMON_SAMPLER_TYPE_INFILL:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
-                        break;
-                    case COMMON_SAMPLER_TYPE_PENALTIES:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                    }
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_K:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
+                    break;
+                case COMMON_SAMPLER_TYPE_MIN_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_XTC:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    break;
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    break;
+                case COMMON_SAMPLER_TYPE_INFILL:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
+                    break;
+                case COMMON_SAMPLER_TYPE_PENALTIES:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    break;
+                default:
+                    GGML_ASSERT(false && "unknown sampler type");
             }
         }
         llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
@@ -475,6 +472,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
         case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
+        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return 's';
         case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
         case COMMON_SAMPLER_TYPE_XTC:         return 'x';
@@ -490,6 +488,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
         case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return "top_n_sigma";
         case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
         case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
@@ -504,6 +503,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "dry",         COMMON_SAMPLER_TYPE_DRY },
         { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
         { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
         { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
         { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
@@ -517,6 +517,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
     std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
         { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
         { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
         { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
         { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
@@ -533,14 +534,16 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         auto sampler = sampler_canonical_name_map.find(name);
         if (sampler != sampler_canonical_name_map.end()) {
             samplers.push_back(sampler->second);
-        } else {
-            if (allow_alt_names) {
-                sampler = sampler_alt_name_map.find(name);
-                if (sampler != sampler_alt_name_map.end()) {
-                    samplers.push_back(sampler->second);
-                }
+            continue;
+        }
+        if (allow_alt_names) {
+            sampler = sampler_alt_name_map.find(name);
+            if (sampler != sampler_alt_name_map.end()) {
+                samplers.push_back(sampler->second);
+                continue;
             }
         }
+        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str());
     }
 
     return samplers;
@@ -552,6 +555,7 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_N_SIGMA), COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
@@ -566,6 +570,8 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
         const auto sampler = sampler_name_map.find(c);
         if (sampler != sampler_name_map.end()) {
             samplers.push_back(sampler->second);
+        } else {
+            LOG_WRN("%s: unable to match sampler by char '%c'\n", __func__, c);
         }
     }
 

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")
@@ -1388,10 +1424,10 @@ class BaichuanModel(TextModel):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_file_type(self.ftype)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         head_count = self.hparams["num_attention_heads"]
@@ -1512,10 +1548,10 @@ class XverseModel(TextModel):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_file_type(self.ftype)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -1828,10 +1864,10 @@ class LlamaModel(TextModel):
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
@@ -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
@@ -2206,10 +2265,10 @@ class DeciModel(TextModel):
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
@@ -2449,10 +2508,10 @@ class MiniCPMModel(TextModel):
         logit_scale = self.hparams["hidden_size"] / self.hparams["dim_model_base"]
         self.gguf_writer.add_logit_scale(logit_scale)
         logger.info(f"gguf: (minicpm) logit_scale = {logit_scale}")
-        if self.hparams.get("rope_scaling") is not None:
-            if self.hparams["rope_scaling"].get("type") == "longrope":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
-                logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "longrope":
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
+            logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
@@ -2584,7 +2643,7 @@ class QwenModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
 
-@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM")
+@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM", "Qwen2AudioForConditionalGeneration")
 class Qwen2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2
 
@@ -2597,19 +2656,25 @@ class Qwen2Model(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self._try_set_pooling_type()
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
     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("multi_modal_projector") \
+                or name.startswith("vision_model") or name.startswith("audio_tower"):
+            # skip vision and audio 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 +2698,8 @@ 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 +2714,9 @@ 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 +2774,62 @@ 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
@@ -2761,6 +2882,13 @@ class Qwen2MoeModel(TextModel):
         if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
             self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
             logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
+        # YaRN is not enabled by default
+        # To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -3028,7 +3156,7 @@ class Phi3MiniModel(TextModel):
 
         scale = max_pos_embds / orig_max_pos_embds
 
-        rope_scaling_type = rope_scaling.get('type', '').lower()
+        rope_scaling_type = rope_scaling.get('rope_type', rope_scaling.get('type', '')).lower()
         if len(rope_scaling_type) == 0:
             raise KeyError('Missing the required key rope_scaling.type')
 
@@ -3340,10 +3468,10 @@ class InternLM2Model(TextModel):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
         self.gguf_writer.add_file_type(self.ftype)
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
@@ -3353,6 +3481,11 @@ 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
 
@@ -3418,14 +3551,18 @@ class InternLM3Model(TextModel):
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear" or self.hparams["rope_scaling"].get("rope_type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     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")):
@@ -3900,14 +4037,24 @@ 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)
+        # calculate proj_scale_factor (used by tinygemma3 test model)
+        image_seq_length = self.preprocessor_config.get("image_seq_length", 256)
+        n_per_side = int(image_seq_length ** 0.5)
+        image_size = self.hparams["image_size"]
+        patch_size = self.hparams["patch_size"]
+        proj_scale_factor = (image_size // patch_size) // n_per_side
+        if proj_scale_factor > 0 and proj_scale_factor != 4:
+            # we only need to write this if it's not the default value
+            # in this case, we are converting a test model
+            self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         del bid, new_name, n_dims  # unused
@@ -3921,6 +4068,9 @@ class Gemma3VisionModel(VisionModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
+        if "vision_model.head." in name:
+            return [] # skip redundant tensors for tinygemma3
+
         if name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
                 or name.startswith("multimodal_projector.") or name.startswith("vision_model."):
             # process vision tensors
@@ -4846,12 +4996,12 @@ class DeepseekV2Model(TextModel):
 
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
-                self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_scaling["mscale_all_dim"])
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -5343,11 +5493,11 @@ class Glm4Model(TextModel):
         super().set_gguf_parameters()
         rope_dim = self.hparams["head_dim"]
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
 
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
@@ -5580,10 +5730,10 @@ class ExaoneModel(TextModel):
         rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
         rotary_factor = rotary_factor if rotary_factor is not None else 1.0
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
-        if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
-            if hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
@@ -5649,11 +5799,20 @@ 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
@@ -5664,12 +5823,21 @@ 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)
 
 
@@ -5686,10 +5854,11 @@ class BailingMoeModel(TextModel):
         rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
-        if (self.hparams.get("rope_scaling") or {}).get("type") == "yarn" and "factor" in self.hparams["rope_scaling"]:
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
         else:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
         self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
@@ -5819,6 +5988,65 @@ 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. Instead, it uses Llama or other models for text. If you want to get the audio encoder, please use --mmproj argument")
+
+
+@ModelBase.register("Qwen2AudioForConditionalGeneration")
+class WhisperEncoderModel(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.QWEN2A)
+        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))
+
+    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
+
+        if name.startswith("language_model."):
+            # skip language model tensors
+            return []
+
+        # 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)]
+
+
+@ModelBase.register("UltravoxModel")
+class UltravoxWhisperEncoderModel(WhisperEncoderModel):
+    has_vision_encoder = False # no vision encoder
+    has_audio_encoder = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_audio_stack_factor(self.global_config["stack_factor"])
+
 ###### CONVERSION LOGIC ######
 
 
@@ -5994,13 +6222,15 @@ 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
 
@@ -6063,7 +6293,7 @@ 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                   |   √   |   √  |   √  |
 
 
 
@@ -258,6 +280,15 @@ cmake --build build --config release
 ### **GitHub contribution**:
 Please add the **[CANN]** prefix/tag in issues/PRs titles to help the CANN-team check/address them without delay.
 
+## Updates
+### Basic Flash Attention Support
+The basic FA kernel with aclnnops has been added in aclnn_ops.cpp.
+Currently, the FA only supports the cases with FP16 KV tensors and NO logit softcap.
+Since the aclnn interface for flash attention cannot support the logit softcap, we will only update the quantized version in the future.
+
+Authors from Peking University: Bizhao Shi (bshi@pku.edu.cn), Yuxin Yang (yxyang@pku.edu.cn), Ruiyang Ma (ruiyang@stu.pku.edu.cn), and Guojie Luo (gluo@pku.edu.cn).
+
+We would like to thank Tuo Dai, Shanni Li, and all of the project maintainers from Huawei Technologies Co., Ltd for their help during the code development and pull request.
 
 ## TODO
 - Support more models and data types.

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/function-calling.md

@@ -325,36 +325,65 @@ To get the official template from original HuggingFace repos, you can use [scrip
 > [!TIP]
 > If there is no official `tool_use` Jinja template, you may want to set `--chat-template chatml` to use a default that works with many models (YMMV!), or write your own (e.g. we provide a custom [llama-cpp-deepseek-r1.jinja](../models/templates/llama-cpp-deepseek-r1.jinja) for DeepSeek R1 distills)
 
+> [!CAUTION]
+> Beware of extreme KV quantizations (e.g. `-ctk q4_0`), they can substantially degrade the model's tool calling performance.
+
 Test in CLI (or with any library / software that can use OpenAI-compatible API backends):
 
 ```bash
 curl http://localhost:8080/v1/chat/completions -d '{
-"model": "gpt-3.5-turbo",
-"tools": [
-    {
-    "type":"function",
-    "function":{
-        "name":"python",
-        "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
-        "parameters":{
-        "type":"object",
-        "properties":{
-            "code":{
-            "type":"string",
-            "description":"The code to run in the ipython interpreter."
+    "model": "gpt-3.5-turbo",
+    "tools": [
+        {
+        "type":"function",
+        "function":{
+            "name":"python",
+            "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
+            "parameters":{
+            "type":"object",
+            "properties":{
+                "code":{
+                "type":"string",
+                "description":"The code to run in the ipython interpreter."
+                }
+            },
+            "required":["code"]
             }
-        },
-        "required":["code"]
         }
-    }
-    }
-],
-"messages": [
-    {
-    "role": "user",
-    "content": "Print a hello world message with python."
-    }
-]
+        }
+    ],
+    "messages": [
+        {
+        "role": "user",
+        "content": "Print a hello world message with python."
+        }
+    ]
+}'
+
+
+curl http://localhost:8080/v1/chat/completions -d '{
+    "model": "gpt-3.5-turbo",
+    "messages": [
+        {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things."},
+        {"role": "user", "content": "What is the weather in Istanbul?"}
+    ],
+    "tools": [{
+        "type":"function",
+        "function":{
+            "name":"get_current_weather",
+            "description":"Get the current weather in a given location",
+            "parameters":{
+                "type":"object",
+                "properties":{
+                    "location":{
+                        "type":"string",
+                        "description":"The city and country/state, e.g. `San Francisco, CA`, or `Paris, France`"
+                    }
+                },
+                "required":["location"]
+            }
+        }
+    }]
 }'
 ```
 

docs/multimodal.md

@@ -0,0 +1,100 @@
+# 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/collections/ggml-org/multimodal-ggufs-68244e01ff1f39e5bebeeedc
+
+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
+
+# Moondream2 20250414 version
+(tool_name) -hf ggml-org/moondream2-20250414-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
+
+# Qwen2-Audio and SeaLLM-Audio
+# note: no pre-quantized GGUF this model, as they have very poor result
+# ref: https://github.com/ggml-org/llama.cpp/pull/13760
+```

examples/CMakeLists.txt

@@ -21,7 +21,6 @@ else()
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf)
     add_subdirectory(gritlm)
-    add_subdirectory(infill)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(parallel)
@@ -33,6 +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/embedding/embedding.cpp

@@ -35,23 +35,14 @@ 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, int embd_norm) {
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    const struct llama_model * model = llama_get_model(ctx);
 
     // 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_model_has_encoder(model) && !llama_model_has_decoder(model)) {
-        // encoder-only model
-        if (llama_encode(ctx, batch) < 0) {
-            LOG_ERR("%s : failed to encode\n", __func__);
-        }
-    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
-        // decoder-only model
-        if (llama_decode(ctx, batch) < 0) {
-            LOG_ERR("%s : failed to decode\n", __func__);
-        }
+    if (llama_decode(ctx, batch) < 0) {
+        LOG_ERR("%s : failed to process\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {

examples/infill/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET llama-infill)
-add_executable(${TARGET} infill.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/infill/README.md

@@ -1,47 +0,0 @@
-# llama.cpp/example/infill
-
-This example shows how to use the infill mode with Code Llama models supporting infill mode.
-Currently the 7B and 13B models support infill mode.
-
-Infill supports most of the options available in the main example.
-
-For further information have a look at the main README.md in llama.cpp/example/main/README.md
-
-## Common Options
-
-In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models:
-
--   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
--   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
--   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
--   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 4096, but if a LLaMA model was built with a longer context, increasing this value will provide better results for longer input/inference.
--   `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.
-
-## Input Prompts
-
-The `infill` program provides several ways to interact with the LLaMA models using input prompts:
-
--   `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option.
--   `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option.
--   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
-
-## Interaction
-
-The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first`
-
-### Interaction Options
-
--   `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model.
--   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
--   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
-
-### Example
-
-Download a model that supports infill, for example CodeLlama:
-```console
-scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models
-```
-
-```bash
-./llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
-```

examples/infill/infill.cpp

@@ -1,590 +0,0 @@
-#include "arg.h"
-#include "common.h"
-#include "console.h"
-#include "sampling.h"
-#include "log.h"
-#include "llama.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <ctime>
-#include <fstream>
-#include <iostream>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-#include <signal.h>
-#include <unistd.h>
-#elif defined (_WIN32)
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-#define NOMINMAX
-#endif
-#include <windows.h>
-#include <signal.h>
-#endif
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-static llama_context           ** g_ctx;
-static llama_model             ** g_model;
-static common_sampler          ** g_smpl;
-static common_params            * g_params;
-static std::vector<llama_token> * g_input_tokens;
-static std::ostringstream       * g_output_ss;
-static std::vector<llama_token> * g_output_tokens;
-
-static bool is_interacting = false;
-
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-static void sigint_handler(int signo) {
-    if (signo == SIGINT) {
-        if (!is_interacting) {
-            is_interacting = true;
-        } else {
-            console::cleanup();
-            LOG("\n");
-            common_perf_print(*g_ctx, *g_smpl);
-
-            // make sure all logs are flushed
-            LOG("Interrupted by user\n");
-            common_log_pause(common_log_main());
-
-            _exit(130);
-        }
-    }
-}
-#endif
-
-int main(int argc, char ** argv) {
-    common_params params;
-    g_params = &params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
-        return 1;
-    }
-
-    common_init();
-
-    auto & sparams = params.sampling;
-
-    console::init(params.simple_io, params.use_color);
-    atexit([]() { console::cleanup(); });
-
-    if (params.logits_all) {
-        LOG_ERR("\n************\n");
-        LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
-        LOG_ERR("************\n\n");
-
-        return 0;
-    }
-
-    if (params.embedding) {
-        LOG_ERR("\n************\n");
-        LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
-        LOG_ERR("************\n\n");
-
-        return 0;
-    }
-
-    if (params.n_ctx != 0 && params.n_ctx < 8) {
-        LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__);
-        params.n_ctx = 8;
-    }
-
-    if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
-        LOG_ERR("\n************\n");
-        LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
-        LOG_ERR("************\n\n");
-
-        return 0;
-    }
-
-    if (params.rope_freq_base != 0.0) {
-        LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
-    }
-
-    if (params.rope_freq_scale != 0.0) {
-        LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
-    }
-
-    LOG_INF("%s: llama backend init\n", __func__);
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    llama_model * model = nullptr;
-    llama_context * ctx = nullptr;
-    common_sampler * smpl = nullptr;
-
-    g_model = &model;
-    g_ctx = &ctx;
-    g_smpl = &smpl;
-
-    // load the model and apply lora adapter, if any
-    LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
-    common_init_result llama_init = common_init_from_params(params);
-
-    model = llama_init.model.get();
-    ctx = llama_init.context.get();
-
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n", __func__);
-        return 1;
-    }
-
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    const int n_ctx_train = llama_model_n_ctx_train(model);
-    const int n_ctx = llama_n_ctx(ctx);
-    LOG_DBG("n_ctx: %d\n", n_ctx);
-
-    if (n_ctx > n_ctx_train) {
-        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
-    }
-
-    // print system information
-    {
-        LOG_INF("\n");
-        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
-    }
-    const bool add_bos = llama_vocab_get_add_bos(vocab);
-    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
-
-    std::vector<llama_token> embd_inp;
-    std::vector<llama_token> embd_end;
-    std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
-    std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
-
-    GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0);
-    GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0);
-
-    inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
-    inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
-
-    embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
-    embd_end = params.spm_infill ? inp_pfx : inp_sfx;
-    if (add_bos) {
-        embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
-    }
-    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
-
-    const llama_token middle_token = llama_vocab_fim_mid(vocab);
-    if (middle_token >= 0) {
-        embd_inp.push_back(middle_token);
-    }
-
-    LOG_DBG("add_bos: %d\n", add_bos);
-    LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
-    LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
-    LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
-
-    // Should not run without any tokens
-    if (embd_inp.empty()) {
-        embd_inp.push_back(llama_vocab_bos(vocab));
-        LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
-    }
-
-    if ((int) embd_inp.size() > n_ctx - 4) {
-        LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
-        return 1;
-    }
-
-    // number of tokens to keep when resetting context
-    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
-        params.n_keep = (int)embd_inp.size();
-    }
-
-    LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
-    LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str());
-
-    // enable interactive mode if interactive start is specified
-    if (params.interactive_first) {
-        params.interactive = true;
-    }
-
-    if (params.verbose_prompt) {
-        LOG_INF("\n");
-        LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-        for (int i = 0; i < (int) embd_inp.size(); i++) {
-            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
-        }
-
-        if (params.n_keep > 0) {
-        LOG_INF("%s: static prompt based on n_keep: '", __func__);
-            for (int i = 0; i < params.n_keep; i++) {
-                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
-            }
-            LOG_CNT("'\n");
-        }
-        LOG_INF("\n");
-    }
-
-    if (params.interactive) {
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-        struct sigaction sigint_action;
-        sigint_action.sa_handler = sigint_handler;
-        sigemptyset (&sigint_action.sa_mask);
-        sigint_action.sa_flags = 0;
-        sigaction(SIGINT, &sigint_action, NULL);
-#elif defined (_WIN32)
-        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
-        };
-        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
-#endif
-
-        LOG_INF("%s: interactive mode on.\n", __func__);
-
-        if (params.input_prefix_bos) {
-            LOG_INF("Input prefix with BOS\n");
-        }
-
-        if (!params.input_prefix.empty()) {
-            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
-        }
-
-        if (!params.input_suffix.empty()) {
-            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
-        }
-    }
-    smpl = common_sampler_init(model, sparams);
-
-    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
-    LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
-    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
-
-    LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-
-    LOG_INF("\n");
-    LOG_INF("\n#####  Infill mode  #####\n\n");
-    if (params.interactive) {
-        const char *control_message;
-        if (params.multiline_input) {
-            control_message = " - To return control to LLaMA, end your input with '\\'.\n"
-                              " - To return control without starting a new line, end your input with '/'.\n";
-        } else {
-            control_message = " - Press Return to return control to LLaMA.\n"
-                              " - To return control without starting a new line, end your input with '/'.\n"
-                              " - If you want to submit another line, end your input with '\\'.\n";
-        }
-        LOG_INF("== Running in interactive mode. ==\n");
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
-#endif
-        LOG_INF(       "%s\n", control_message);
-
-        is_interacting = params.interactive_first;
-    }
-
-    bool input_echo = true;
-
-    int n_past     = 0;
-    int n_remain   = params.n_predict;
-    int n_consumed = 0;
-
-    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
-    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
-    std::ostringstream output_ss;     g_output_ss     = &output_ss;
-
-    // the first thing we will do is to output the prompt, so set color accordingly
-    console::set_display(console::prompt);
-
-    std::vector<llama_token> embd;
-
-    while (n_remain != 0 || params.interactive) {
-        // predict
-        if (!embd.empty()) {
-            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
-            // --prompt or --file which uses the same value.
-            int max_embd_size = n_ctx - 4;
-
-            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
-            if ((int) embd.size() > max_embd_size) {
-                const int skipped_tokens = (int) embd.size() - max_embd_size;
-                embd.resize(max_embd_size);
-
-                console::set_display(console::error);
-                LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
-                console::set_display(console::reset);
-            }
-
-            // infinite text generation via context swapping
-            // if we run out of context:
-            // - take the n_keep first tokens from the original prompt (via n_past)
-            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() > n_ctx) {
-                if (params.n_predict == -2) {
-                    LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                    break;
-                }
-
-                const int n_left    = n_past - params.n_keep - 1;
-                const int n_discard = n_left/2;
-
-                LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                    n_past, n_left, n_ctx, params.n_keep, n_discard);
-
-                llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
-
-                n_past -= n_discard;
-
-                LOG_DBG("after swap: n_past = %d\n", n_past);
-
-                LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
-
-            }
-
-            // evaluate tokens in batches
-            // embd is typically prepared beforehand to fit within a batch, but not always
-            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
-                int n_eval = (int) embd.size() - i;
-                if (n_eval > params.n_batch) {
-                    n_eval = params.n_batch;
-                }
-
-                LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
-
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
-                    LOG_ERR("%s : failed to eval\n", __func__);
-                    return 1;
-                }
-
-                n_past += n_eval;
-
-                LOG_DBG("n_past = %d\n", n_past);
-            }
-
-        }
-
-        embd.clear();
-
-        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            const llama_token id = common_sampler_sample(smpl, ctx, -1);
-
-            common_sampler_accept(smpl, id, true);
-
-            // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
-
-            embd.push_back(id);
-
-            // echo this to console
-            input_echo = true;
-
-            // decrement remaining sampling budget
-            --n_remain;
-
-            LOG_DBG("n_remain: %d\n", n_remain);
-        } else {
-            // some user input remains from prompt or interaction, forward it to processing
-            LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
-            while ((int) embd_inp.size() > n_consumed) {
-                embd.push_back(embd_inp[n_consumed]);
-
-                // push the prompt in the sampling context in order to apply repetition penalties later
-                // for the prompt, we don't apply grammar rules
-                common_sampler_accept(smpl, embd_inp[n_consumed], false);
-
-                ++n_consumed;
-                if ((int) embd.size() >= params.n_batch) {
-                    break;
-                }
-            }
-        }
-
-        // display text
-        if (input_echo) {
-            for (auto id : embd) {
-                const std::string token_str = common_token_to_piece(ctx, id);
-                LOG("%s", token_str.c_str());
-
-                if (embd.size() > 1) {
-                    input_tokens.push_back(id);
-                } else {
-                    output_tokens.push_back(id);
-                    output_ss << token_str;
-                }
-            }
-        }
-        // reset color to default if we there is no pending user input
-        if (input_echo && (int) embd_inp.size() == n_consumed) {
-            console::set_display(console::reset);
-        }
-
-        // if not currently processing queued inputs;
-        if ((int) embd_inp.size() <= n_consumed) {
-            // deal with eot token in infill mode
-            if ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){
-                if (is_interacting && !params.interactive_first) {
-                    // print an eot token
-                    LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
-                }
-                LOG("\n");
-                console::set_display(console::user_input);
-                std::string buffer;
-                std::string line;
-                bool another_line=true;
-                // set a new prefix via stdin
-                do {
-                    another_line = console::readline(line, params.multiline_input);
-                    buffer += line;
-                } while (another_line);
-                // check if we got an empty line, if so we use the old input
-                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
-                    params.input_prefix = buffer;
-                }
-                buffer.clear();
-                // set a new suffix via stdin
-                do {
-                    another_line = console::readline(line, params.multiline_input);
-                    buffer += line;
-                } while (another_line);
-                // check if we got an empty line
-                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
-                    params.input_suffix = buffer;
-                }
-                buffer.clear();
-                // done taking input, reset color
-                console::set_display(console::reset);
-
-                if (params.escape) {
-                    //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
-                    string_process_escapes(params.input_prefix);
-                    string_process_escapes(params.input_suffix);
-                }
-
-                // tokenize new prefix and suffix
-                std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
-                std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
-
-                inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
-                inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
-
-                embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
-                embd_end = params.spm_infill ? inp_pfx : inp_sfx;
-                if (add_bos) {
-                    embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
-                }
-                embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
-
-                if (middle_token >= 0) {
-                    embd_inp.push_back(middle_token);
-                }
-
-                embd.clear();
-                n_remain = params.n_predict;
-                n_past = 0;
-                n_consumed = 0;
-                is_interacting = false;
-            }
-            // deal with end of generation tokens in interactive mode
-            else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
-                LOG_DBG("found EOS token\n");
-
-                if (params.interactive) {
-
-                    is_interacting = true;
-                    LOG("\n");
-                    console::set_display(console::user_input);
-               }
-            }
-
-            if (n_past > 0 && is_interacting && !params.interactive) {
-                LOG_DBG("waiting for user input\n");
-
-                if (params.input_prefix_bos) {
-                    LOG_DBG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_vocab_bos(vocab));
-                }
-
-                std::string buffer;
-                if (!params.input_prefix.empty()) {
-                    LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
-                    buffer += params.input_prefix;
-                    LOG("%s", buffer.c_str());
-                }
-
-                std::string line;
-                bool another_line = true;
-                do {
-                    another_line = console::readline(line, params.multiline_input);
-                    buffer += line;
-                } while (another_line);
-
-                // done taking input, reset color
-                console::set_display(console::reset);
-
-                // Add tokens to embd only if the input buffer is non-empty
-                // Entering a empty line lets the user pass control back
-                if (buffer.length() > 1) {
-                    // append input suffix if any
-                    if (!params.input_suffix.empty()) {
-                        LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
-                        buffer += params.input_suffix;
-                        LOG("%s", params.input_suffix.c_str());
-                    }
-
-                    LOG_DBG("buffer: '%s'\n", buffer.c_str());
-
-                    const size_t original_size = embd_inp.size();
-
-                    const auto line_inp = common_tokenize(ctx, buffer, false);
-                    LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
-
-                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
-
-                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
-                        const llama_token token = embd_inp[i];
-                        output_tokens.push_back(token);
-                        output_ss << common_token_to_piece(ctx, token);
-                    }
-
-                    n_remain -= line_inp.size();
-                    LOG_DBG("n_remain: %d\n", n_remain);
-                } else {
-                    LOG_DBG("empty line, passing control back\n");
-                }
-
-                input_echo = false; // do not echo this again
-            }
-
-            if (n_past > 0) {
-                if (is_interacting) {
-                    common_sampler_reset(smpl);
-                }
-                is_interacting = false;
-            }
-        }
-
-        // end of generation
-        if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !params.interactive) {
-            break;
-        }
-
-        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
-        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
-        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
-            n_remain = params.n_predict;
-            is_interacting = true;
-        }
-    }
-    if (!params.interactive && n_remain <= 0) {
-        LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
-    }
-
-    LOG("\n");
-    common_perf_print(ctx, smpl);
-
-    common_sampler_free(smpl);
-    llama_backend_free();
-
-    return 0;
-}

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_process(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__);
+        LOG_ERR("%s : failed to process\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_process(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_process(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_process(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
@@ -366,6 +367,8 @@ if (MSVC)
         /wd4005  # Macro redefinition
         /wd4244  # Conversion from one type to another type, possible loss of data
         /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+        /wd4996  # Disable POSIX deprecation warnings
+        /wd4702  # Unreachable code warnings
     )
     function(disable_msvc_warnings target_name)
         if(TARGET ${target_name})

ggml/include/ggml-backend.h

@@ -38,7 +38,7 @@ extern "C" {
     GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer  (ggml_backend_buffer_type_t buft, size_t size);
     GGML_API size_t                ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
     GGML_API size_t                ggml_backend_buft_get_max_size  (ggml_backend_buffer_type_t buft);
-    GGML_API size_t                ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API size_t                ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
     GGML_API bool                  ggml_backend_buft_is_host       (ggml_backend_buffer_type_t buft);
     GGML_API ggml_backend_dev_t    ggml_backend_buft_get_device    (ggml_backend_buffer_type_t buft);
 
@@ -59,7 +59,7 @@ extern "C" {
     GGML_API enum ggml_status               ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor);
     GGML_API void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
     GGML_API bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
     GGML_API void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
@@ -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,
     };
@@ -673,11 +674,15 @@ extern "C" {
     GGML_API bool ggml_is_3d        (const struct ggml_tensor * tensor);
     GGML_API int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
+    // returns whether the tensor elements can be iterated over with a flattened index (no gaps, no permutation)
     GGML_API bool ggml_is_contiguous  (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
     GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
     GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
 
+    // returns whether the tensor elements are allocated as one contiguous block of memory (no gaps, but permutation ok)
+    GGML_API bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor);
+
     // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
     GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
 
@@ -764,7 +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);
 
     //
@@ -934,7 +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
@@ -1020,6 +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);
@@ -2045,15 +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/CMakeLists.txt

@@ -214,7 +214,7 @@ add_library(ggml
 target_link_libraries(ggml PUBLIC ggml-base)
 
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    target_link_libraries(ggml PRIVATE dl stdc++fs)
+    target_link_libraries(ggml PRIVATE dl)
 endif()
 
 function(ggml_add_backend_library backend)

ggml/src/ggml-backend.cpp

@@ -56,7 +56,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
     return SIZE_MAX;
 }
 
-size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
     // get_alloc_size is optional, defaults to ggml_nbytes
     if (buft->iface.get_alloc_size) {
         size_t size = buft->iface.get_alloc_size(buft, tensor);
@@ -152,7 +152,7 @@ size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
 }
 
-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor) {
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
 }
 
@@ -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/acl_tensor.cpp

@@ -31,6 +31,8 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_FLOAT;
         case GGML_TYPE_F16:
             return ACL_FLOAT16;
+        case GGML_TYPE_BF16:
+            return ACL_BF16;
         case GGML_TYPE_I8:
             return ACL_INT8;
         case GGML_TYPE_I16:

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -65,6 +65,8 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
+#include <aclnnop/aclnn_grouped_matmul_v2.h>
+#include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
 #include <float.h>
 
 #include <cmath>
@@ -73,11 +75,13 @@
 #include <vector>
 
 #include "ggml-impl.h"
+#include "ggml.h"
 
 #define GGML_COMMON_DECL_C
 
 #include "../ggml-common.h"
 
+
 void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclTensor ** acl_src0,
                  aclTensor ** acl_src1, aclTensor ** acl_dst) {
     GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_can_repeat(src1, src0));
@@ -2587,3 +2591,603 @@ 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;
+    }
+}
+
+void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
+
+    ggml_tensor* src0 = dst->src[0]; // q, fp32
+    ggml_tensor* src1 = dst->src[1]; // k, fp16
+    ggml_tensor* src2 = dst->src[2]; // v, fp16
+    ggml_tensor* src3 = dst->src[3]; // mask, fp16
+
+    float maxBias = 0.0f;
+    float scaleValue = 1.0f;
+    float logitSoftcap = 0.0f;
+    memcpy(&scaleValue,    (float*)dst->op_params + 0, sizeof(float));
+    memcpy(&maxBias,       (float*)dst->op_params + 1, sizeof(float));
+    memcpy(&logitSoftcap,  (float*)dst->op_params + 2, sizeof(float));
+
+    if(logitSoftcap == 0.0f){
+        size_t faElemSize = sizeof(uint16_t);
+        auto   faDataType = ACL_FLOAT16; //ACL_BF16;
+
+        aclTensor* acl_src0_f16_tensor = nullptr;
+        aclTensor* acl_src1_f16_tensor = nullptr;
+        aclTensor* acl_src2_f16_tensor = nullptr;
+        aclTensor* acl_dst_f16_tensor  = nullptr;
+
+        // Step 1: cast the src0 (Query) to fp16 if needed
+        ggml_cann_pool_alloc src0_f16_allocator(ctx.pool());
+        void* src0_f16_buffer = nullptr;
+
+        if(ggml_cann_type_mapping(src0->type) != faDataType){
+            aclTensor* acl_src0_f32_tensor = ggml_cann_create_tensor(src0);
+            src0_f16_buffer = src0_f16_allocator.alloc(
+                                    ggml_nelements(src0) * faElemSize);
+
+            int64_t* src0_f16_ne = src0->ne;
+            size_t   src0_f16_nb[GGML_MAX_DIMS];
+            src0_f16_nb[0] = sizeof(uint16_t);
+            for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                src0_f16_nb[i] = src0_f16_nb[i - 1] * src0_f16_ne[i - 1];
+            }
+
+            acl_src0_f16_tensor = ggml_cann_create_tensor(
+                src0_f16_buffer, faDataType, faElemSize,
+                src0_f16_ne, src0_f16_nb, GGML_MAX_DIMS
+            );
+            aclnn_cast(ctx, acl_src0_f32_tensor, acl_src0_f16_tensor, faDataType);
+            ggml_cann_release_resources(ctx, acl_src0_f32_tensor);
+        }else{
+            acl_src0_f16_tensor = ggml_cann_create_tensor(src0);
+        }
+
+        // Step 2: create the acl tensors for src1 (Key), src2 (Value),
+        //         and the direct output from FusedInferAttention
+
+        acl_src1_f16_tensor = ggml_cann_create_tensor(src1);
+        acl_src2_f16_tensor = ggml_cann_create_tensor(src2);
+
+        ggml_cann_pool_alloc out_f16_allocator(ctx.pool());
+        void* out_f16_buffer = out_f16_allocator.alloc(
+                                    ggml_nelements(dst) * faElemSize);
+
+        int64_t* out_f16_ne = src0->ne;
+        size_t out_f16_nb[GGML_MAX_DIMS];
+        out_f16_nb[0] = faElemSize;
+        for(int i = 1; i < GGML_MAX_DIMS; ++i){
+            out_f16_nb[i] = out_f16_nb[i - 1] * out_f16_ne[i - 1];
+        }
+
+        acl_dst_f16_tensor = ggml_cann_create_tensor(
+            out_f16_buffer, faDataType, faElemSize,
+            out_f16_ne, out_f16_nb, GGML_MAX_DIMS
+        );
+
+        // Step 3: create the PSEShift tensor if needed
+        //         this tensor is considered as mask (f16) in the llama.cpp
+
+        aclTensor* bcast_pse_tensor = nullptr;
+        int64_t bcast_pse_ne[GGML_MAX_DIMS];
+        size_t bcast_pse_nb[GGML_MAX_DIMS];
+        ggml_cann_pool_alloc bcast_pse_allocator(ctx.pool());
+        void* bcast_pse_buffer = nullptr;
+
+        if(src3 != nullptr){
+            bcast_pse_buffer = bcast_pse_allocator.alloc(
+                            ggml_nelements(src3) * src0->ne[2] * sizeof(uint16_t));
+
+            if(src0->ne[1] > 1){
+                // Case 1: broadcast pse for prefill stage with multiple head
+                aclTensor* acl_mask_f16_tensor = ggml_cann_create_tensor(src3);
+                bcast_pse_ne[0] = src3->ne[0];
+                bcast_pse_ne[1] = src3->ne[1];
+                bcast_pse_ne[2] = src0->ne[2];
+                bcast_pse_ne[3] = src3->ne[3];
+
+                bcast_pse_nb[0] = sizeof(uint16_t);
+                for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                    bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
+                }
+
+                bcast_pse_tensor = ggml_cann_create_tensor(
+                    bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
+
+                int64_t repeats[] = {1, src0->ne[2], 1, 1};
+                aclnn_repeat(ctx, acl_mask_f16_tensor, bcast_pse_tensor, repeats);
+
+                ggml_cann_release_resources(ctx, acl_mask_f16_tensor);
+            }else{
+                // Case 2: trunc the first row and broadcast pse for decode stage with multiple head
+                int64_t trunc_pse_ne[GGML_MAX_DIMS] = {src3->ne[0], src0->ne[1], src3->ne[2], src3->ne[3]};
+                size_t* trunc_pse_nb = src3->nb;
+
+                aclTensor* acl_mask_f16_trunc_tensor = ggml_cann_create_tensor(
+                    src3->data, ACL_FLOAT16, sizeof(uint16_t),
+                    trunc_pse_ne, trunc_pse_nb, GGML_MAX_DIMS);
+
+                bcast_pse_ne[0] = src3->ne[0];
+                bcast_pse_ne[1] = src0->ne[1];
+                bcast_pse_ne[2] = src0->ne[2];
+                bcast_pse_ne[3] = src3->ne[3];
+
+                bcast_pse_nb[0] = sizeof(uint16_t);
+                for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                    bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
+                }
+
+                bcast_pse_tensor = ggml_cann_create_tensor(
+                    bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
+
+                int64_t repeats[] = {1, src0->ne[2], 1, 1};
+                aclnn_repeat(ctx, acl_mask_f16_trunc_tensor, bcast_pse_tensor, repeats);
+
+                ggml_cann_release_resources(ctx, acl_mask_f16_trunc_tensor);
+            }
+
+            // Compute the slope if needed. Derived from ggml_cann_softmax().
+            if(maxBias != 0.0f){
+                // alibi
+                const int64_t ne2_ne3 = src0->ne[2] * src0->ne[3];
+                const int64_t n_head = src0->ne[2];
+                const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
+                float m0 = powf(2.0f, -(maxBias) / n_heads_log2_floor);
+                float m1 = powf(2.0f, -(maxBias / 2.0f) / n_heads_log2_floor);
+                // init arange
+                ggml_cann_pool_alloc arange_allocator(ctx.pool(),
+                                                    ne2_ne3 * faElemSize);
+                void* tmp_arange_buffer = arange_allocator.get();
+
+                // arange1: [1, ..., n_heads_log2_floor+1)
+                float start = 1;
+                float stop = n_heads_log2_floor + 1;
+                float step = 1;
+                int64_t n_elements_arange = n_heads_log2_floor;
+
+                int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
+                size_t tmp_arange1_nb[] = {faElemSize};
+                aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
+                    tmp_arange_buffer, faDataType, faElemSize,
+                    tmp_arange1_ne, tmp_arange1_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+                aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
+
+                aclTensor* tmp_arange2_tensor = nullptr;
+                if (n_heads_log2_floor < ne2_ne3) {
+                    // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
+                    start = 1;
+                    stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
+                    step = 2;
+                    n_elements_arange = ne2_ne3 - n_heads_log2_floor;
+                    int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
+                    size_t tmp_arange2_nb[] = {faElemSize};
+
+                    aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
+                        (char*)tmp_arange_buffer +
+                            n_heads_log2_floor * faElemSize,
+                        faDataType, faElemSize,
+                        tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                    aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
+                                n_elements_arange);
+                }
+
+                // init mk_base
+                ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
+                                                    ne2_ne3 * faElemSize);
+                void* tmp_mk_base_buffer = mk_base_allocator.get();
+                int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
+                size_t tmp_mk_base1_nb[] = {faElemSize};
+                aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
+                    tmp_mk_base_buffer, faDataType, faElemSize,
+                    tmp_mk_base1_ne, tmp_mk_base1_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+                aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
+
+                aclTensor* tmp_mk_base2_tensor = nullptr;
+                if (n_heads_log2_floor < ne2_ne3) {
+                    int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
+                    size_t tmp_mk_base2_nb[] = {faElemSize};
+                    aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
+                        (char*)tmp_mk_base_buffer +
+                            n_heads_log2_floor * faElemSize,
+                        faDataType, faElemSize,
+                        tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                    aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
+                }
+
+                // init mk
+                int64_t tmp_mk_base_ne[] = {ne2_ne3};
+                size_t tmp_mk_base_nb[] = {faElemSize};
+                aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
+                    tmp_mk_base_buffer, faDataType, faElemSize,
+                    tmp_mk_base_ne, tmp_mk_base_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
+                    tmp_arange_buffer, faDataType, faElemSize,
+                    tmp_mk_base_ne, tmp_mk_base_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
+
+                // reshape mk
+                int64_t tmp_mk_ne[] = {1, 1, src0->ne[2], src0->ne[3]};
+                size_t tmp_mk_nb[GGML_MAX_DIMS];
+                tmp_mk_nb[0] = faElemSize;
+                for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                    tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
+                }
+                aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
+                    tmp_mk_base_buffer, faDataType, faElemSize,
+                    tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
+                    ACL_FORMAT_ND);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, tmp_mk_tensor);
+
+                ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
+                    tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
+                    tmp_arange_tensor, tmp_mk_tensor);
+            }
+        }
+
+        // Step 4: set the inputs for FusedInferAttention.
+        int kvTensorNum = 1;
+        aclTensor* acl_q_tensor = acl_src0_f16_tensor;
+        aclTensor* acl_k_tensors[] = {acl_src1_f16_tensor};
+        aclTensor* acl_v_tensors[] = {acl_src2_f16_tensor};
+        auto acl_k_tensor_list = aclCreateTensorList(acl_k_tensors, kvTensorNum);
+        auto acl_v_tensor_list = aclCreateTensorList(acl_v_tensors, kvTensorNum);
+
+        int64_t numHeads = src0->ne[2]; // N
+        int64_t numKeyValueHeads = src1->ne[2];
+        // double  scaleValue = 1 / sqrt(src0->ne[0]); // 1/sqrt(d)
+        int64_t preTokens = 65535;
+        int64_t nextTokens = 65535;
+        char layout[5] = {'B', 'N', 'S', 'D', 0};
+        int64_t sparseMode = 0;
+        int64_t innerPrecise = (src0->ne[1] == 1) ? 0 : 2;
+        int64_t blockSize = 0;
+        int64_t antiquantMode = 0;
+        bool softmaxLseFlag = false;
+        int64_t keyAntiquantMode = 0;
+        int64_t valueAntiquantMode = 0;
+
+        // Step 5: launch the FusedInferAttentionScoreV2 kernel.
+        // Refer to https://gitee.com/ascend/cann-ops-adv/blob/master/docs/FusedInferAttentionScoreV2.md
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, FusedInferAttentionScoreV2,
+            acl_q_tensor, acl_k_tensor_list, acl_v_tensor_list, // q, k, v
+            bcast_pse_tensor, nullptr, // pse, mask
+            nullptr, nullptr, // actSeqLen, actSeqLenkv
+            nullptr, nullptr, // deqScale1, quantScale1
+            nullptr, nullptr, nullptr, // deqScale2, quantScale2, quantOffset2
+            nullptr, nullptr, // antiquantScale, antiquantOffset
+            nullptr, // blockTable
+            nullptr, nullptr, // qPadSize, kvPadSize
+            nullptr, nullptr, // kAntiquantScale, kAntiQuantOffset
+            nullptr, nullptr, // vAntiquantScale, vAntiQuantOffset
+            nullptr, nullptr, nullptr, // kSharedPrefix, vSharedPrefix, actSharedLen
+            numHeads, scaleValue, // heads, scaleValue
+            preTokens, nextTokens, // preTokens, nextTokens
+            layout, // inputLayout
+            numKeyValueHeads, // numKVHeads
+            sparseMode, innerPrecise, // sparseMode, innerPrecise
+            blockSize, antiquantMode, // blockSize, antiquantMode
+            softmaxLseFlag, // softmaxLseFlag
+            keyAntiquantMode, valueAntiquantMode, // keyAntiqMode, valueAntiqMode
+            acl_dst_f16_tensor, // attentionOut
+            nullptr // softmaxLse
+        );
+
+        // Step 6: post-processing, permute and cast to f32
+
+        int64_t new_dim[] = {0, 2, 1, 3};
+        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+
+        if(ggml_cann_type_mapping(dst->type) != faDataType){
+            ggml_cann_pool_alloc perm_out_f16_allocator(ctx.pool());
+            perm_out_f16_allocator.alloc(ggml_nelements(dst) * faElemSize);
+            void* perm_out_f16_buffer = perm_out_f16_allocator.get();
+
+            int64_t* perm_out_f16_ne = dst->ne;
+            size_t  perm_out_f16_nb[GGML_MAX_DIMS];
+            perm_out_f16_nb[0] = faElemSize;
+            for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                perm_out_f16_nb[i] = perm_out_f16_nb[i - 1] * perm_out_f16_ne[i - 1];
+            }
+            aclTensor* acl_perm_out_f16_tensor = ggml_cann_create_tensor(
+                perm_out_f16_buffer, faDataType, faElemSize,
+                perm_out_f16_ne, perm_out_f16_nb, GGML_MAX_DIMS);
+            aclnn_permute(ctx, acl_dst_f16_tensor, acl_perm_out_f16_tensor, new_dim, GGML_MAX_DIMS);
+            aclnn_cast(ctx,
+                acl_perm_out_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
+            ggml_cann_release_resources(ctx, acl_perm_out_f16_tensor);
+        }else{
+            // only need to permute
+            aclnn_permute(ctx, acl_dst_f16_tensor, acl_dst_tensor, new_dim, GGML_MAX_DIMS);
+        }
+        ggml_cann_release_resources(ctx, acl_src0_f16_tensor,
+                                         acl_src1_f16_tensor,
+                                         acl_src2_f16_tensor,
+                                         acl_dst_f16_tensor,
+                                         acl_dst_tensor);
+        if(src3 != nullptr){
+            ggml_cann_release_resources(ctx, bcast_pse_tensor);
+        }
+    }else{
+        GGML_ABORT("Function is not implemented.");
+    }
+}

ggml/src/ggml-cann/aclnn_ops.h

@@ -714,6 +714,21 @@ void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  */
 void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Performs the Flash Attention extended operator using the CANN backend.
+ *
+ * @details This function implements the memory-efficient Flash Attention algorithm
+ *          for computing scaled dot-product attention with hardware acceleration.
+ *          The result is stored in the destination tensor `dst`.
+ *
+ *          This operation is accelerated using the CANN backend to improve runtime performance.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored.
+ *            dst->op is expected to be `GGML_OP_FLASH_ATTN_EXT`.
+ */
+void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /*
  * @brief A generic wrapper for ACL resources with custom deleter support.
  */
@@ -978,6 +993,33 @@ 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

@@ -36,6 +36,7 @@
 #include "ggml-backend-impl.h"
 #include "ggml-cann/aclnn_ops.h"
 #include "ggml-cann/common.h"
+#include "ggml.h"
 
 #define GGML_COMMON_DECL_C
 
@@ -1672,7 +1673,8 @@ 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;
@@ -1747,6 +1749,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
         case GGML_OP_COUNT_EQUAL:
             ggml_cann_count_equal(ctx, dst);
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            ggml_cann_flash_attn_ext(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2030,7 +2035,22 @@ 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) {
@@ -2161,6 +2181,38 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:{
+            // derived from [ggml-cuda.cu]
+            if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
+                return false;
+            }
+            if(op->src[1]->type != GGML_TYPE_F16 && op->src[1]->type != GGML_TYPE_F32 && op->src[1]->type != GGML_TYPE_BF16){
+                return false;
+            }
+            if(op->type != GGML_TYPE_F16 && op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_BF16){
+                return false;
+            }
+            if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
+                // different head sizes of K and V are not supported yet
+                return false;
+            }
+            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;
+            }
+            float logitSoftcap = 0.0f;
+            memcpy(&logitSoftcap,  (float*)op->op_params + 2, sizeof(float));
+            if(logitSoftcap != 0.0f) {
+                return false;
+            }
+            return true;
+        }
         default:
             return false;
     }

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-aarch64.cpp

@@ -72,8 +72,6 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
-#elif defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #define UNUSED GGML_UNUSED

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

@@ -20,12 +20,6 @@
 #define GROUP_MAX_EPS_IQ1_M 1e-7f
 #define GROUP_MAX_EPS_IQ1_S 1e-12f
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid warnings for hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-#endif
-
 #define UNUSED GGML_UNUSED
 
 // some compilers don't provide _mm256_set_m128i, e.g. gcc 7
@@ -8525,7 +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);
@@ -8536,6 +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

@@ -50,19 +50,6 @@
 #include "llamafile/sgemm.h"
 #endif
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-
-// disable POSIX deprecation warnings
-// these functions are never going away, anyway
-#pragma warning(disable: 4996)
-
-// unreachable code because of multiple instances of code after GGML_ABORT
-#pragma warning(disable: 4702)
-#endif
-
 // Note: once we move threading into a separate C++ file
 // will use std::hardware_destructive_interference_size instead of hardcoding it here
 // and we'll use C++ attribute syntax.
@@ -295,7 +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,
@@ -2211,6 +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:
                     {
@@ -3492,6 +3484,19 @@ void ggml_cpu_init(void) {
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
 
             GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);
+
+#ifdef GGML_USE_OPENMP
+            //if (!getenv("OMP_WAIT_POLICY")) {
+            //    // set the wait policy to active, so that OpenMP threads don't sleep
+            //    putenv("OMP_WAIT_POLICY=active");
+            //}
+
+            if (!getenv("KMP_BLOCKTIME")) {
+                // set the time to wait before sleeping a thread
+                // this is less aggressive than setting the wait policy to active, but should achieve similar results in most cases
+                putenv("KMP_BLOCKTIME=200"); // 200ms
+            }
+#endif
         }
 
 #if defined(__ARM_ARCH)

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

@@ -8,19 +8,6 @@
 
 #include <float.h>
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-
-// disable POSIX deprecation warnings
-// these functions are never going away, anyway
-#pragma warning(disable: 4996)
-
-// unreachable code because of multiple instances of code after GGML_ABORT
-#pragma warning(disable: 4702)
-#endif
-
 // ggml_compute_forward_dup
 
 static void ggml_compute_forward_dup_same_cont(
@@ -2704,6 +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(
@@ -7762,6 +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.cpp

@@ -2,12 +2,6 @@
 
 #include <cassert>
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-#endif
-
 // precomputed gelu table for f16 (128 KB)
 ggml_fp16_t ggml_table_gelu_f16[1 << 16];
 

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

@@ -12,12 +12,30 @@ if (CUDAToolkit_FOUND)
         # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
         # 70     == V100, FP16 tensor cores
         # 75     == Turing, int8 tensor cores
+        # 80     == Ampere, asynchronous data loading, faster tensor core instructions
+        # 86     == RTX 3000, needs CUDA v11.1
+        # 89     == RTX 4000, needs CUDA v11.8
+        #
+        # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
+        # XX-real    == compile CUDA code as device code for this specific architecture
+        # no suffix  == compile as both PTX and device code
+        #
+        # The default behavior for a non-native is to build virtual architectures as needed to cover all features needed
+        #     for best performance and to also build real architectures for the most commonly used GPUs.
         if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
             set(CMAKE_CUDA_ARCHITECTURES "native")
         elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80")
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
+                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
+            else()
+                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
+            endif()
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80")
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
+                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
+            else()
+                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
+            endif()
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -100,7 +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

@@ -130,10 +130,6 @@ static int ggml_cuda_highest_compiled_arch(const int arch) {
 
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
 #define GGML_CUDA_MAX_STREAMS 8
 
 [[noreturn]]
@@ -300,6 +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.");
 
@@ -713,12 +721,13 @@ 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));
         K_f16.alloc(ggml_nelements(K));
         to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
         to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
@@ -732,7 +741,8 @@ 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);
         to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
@@ -752,10 +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);
 
@@ -767,14 +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);
 
@@ -851,19 +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

@@ -555,8 +555,8 @@ static enum ggml_status ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer
 
     if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
         // initialize padding to 0 to avoid possible NaN values
-        size_t original_size = ggml_nbytes(tensor);
-        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+        const size_t original_size = ggml_nbytes(tensor);
+        const size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
 
         if (padded_size > original_size) {
             ggml_cuda_set_device(ctx->device);
@@ -679,6 +679,7 @@ static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_t
 
     if (ggml_is_quantized(tensor->type)) {
         if (ne0 % MATRIX_ROW_PADDING != 0) {
+            GGML_ASSERT(tensor->nb[0] == ggml_element_size(tensor));
             size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
     }
@@ -800,6 +801,7 @@ static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buff
 
 static enum ggml_status ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");
 
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
@@ -851,6 +853,7 @@ static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buff
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
     GGML_ASSERT(size == ggml_nbytes(tensor));
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");
 
     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
 
@@ -889,6 +892,7 @@ static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buff
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
     GGML_ASSERT(size == ggml_nbytes(tensor));
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");
 
     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
 
@@ -970,6 +974,7 @@ static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buf
 
 static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");
 
     size_t total_size = 0;
 
@@ -1531,6 +1536,8 @@ static void ggml_cuda_op_mul_mat(
 
         // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
         if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             const size_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
@@ -1902,13 +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;
@@ -2062,9 +2075,11 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         }
 
         ggml_tensor src0_slice = *src0;
-        src0_slice.ne[2] = 1;
-        src0_slice.nb[3] = src0_slice.nb[2];
-        src0_slice.data  = (char *) src0->data + i02*nb02;
+        src0_slice.ne[2]    = 1;
+        src0_slice.nb[3]    = src0_slice.nb[2];
+        src0_slice.op       = GGML_OP_VIEW;
+        src0_slice.view_src = dst->src[0]; // non-const pointer to src0
+        src0_slice.data     = (char *) src0->data + i02*nb02;
 
         ggml_tensor src1_slice;
         memset(&src1_slice, 0, sizeof(src1_slice));
@@ -2177,6 +2192,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_SILU:
                     ggml_cuda_op_silu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_GELU_ERF:
+                    ggml_cuda_op_gelu_erf(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_GELU_QUICK:
                     ggml_cuda_op_gelu_quick(ctx, dst);
                     break;
@@ -2962,6 +2980,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
@@ -3206,16 +3225,16 @@ 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

@@ -89,6 +89,17 @@ void ggml_cuda_mul_mat_q(
     const float * src1_d = (const float *) src1->data;
     float       *  dst_d = (float       *)  dst->data;
 
+    // If src0 is a temporary compute buffer, clear any potential padding.
+    if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
+        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));
+        }
+    }
+
     const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
 
     const int64_t s01 = src0->nb[1] / ts_src0;
@@ -111,6 +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));
@@ -118,7 +130,7 @@ void ggml_cuda_mul_mat_q(
 
         const mmq_args args = {
             src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
-            ne00, ne01, ne1, s01, s1,
+            ne00, ne01, ne1, s01, ne11, s1,
             ne02, ne12, s02, s12, s2,
             ne03, ne13, s03, s13, s3,
             use_stream_k};
@@ -194,6 +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));
@@ -202,7 +215,7 @@ void ggml_cuda_mul_mat_q(
     // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
     const mmq_args args = {
         src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
-        ne00, ne01, ne_get_rows, s01, s1,
+        ne00, ne01, ne_get_rows, s01, ne_get_rows, s1,
         ne02, ne02, s02, s12, s2,
         ne03, ne13, s03, s13, s3,
         use_stream_k};
@@ -241,7 +254,7 @@ void ggml_cuda_op_mul_mat_q(
         ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
     const mmq_args args = {
         src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
-        ne00, row_diff, src1_ncols, stride01, nrows_dst,
+        ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst,
         1, 1, 0, 0, 0,
         1, 1, 0, 0, 0,
         use_stream_k};

ggml/src/ggml-cuda/mmq.cuh

@@ -2522,7 +2522,7 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
 static __device__ __forceinline__ void mul_mat_q_process_tile(
         const char * __restrict__ x, const int offset_x, const int * __restrict__ y,
         const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-        const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int stride_row_x, const int ncols_y, const int stride_col_dst,
         const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) {
 
     constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
@@ -2606,7 +2606,7 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 static __global__ void mul_mat_q(
         const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
         const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst,
         const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 
@@ -2619,8 +2619,8 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
-    const int ntx = (ncols_y + mmq_x - 1) / mmq_x; // Number of tiles x
-    const int nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y
+    const int ntx = (ncols_dst + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int nty = (nrows_x   + mmq_y - 1) / mmq_y; // Number of tiles y
 
     // Initialize the ids for writing back data with just the index.
     // For regular matrix multiplications this is never changed.
@@ -2648,8 +2648,8 @@ static __global__ void mul_mat_q(
 
         // Defaults for regular matrix multiplication:
         int col_low    = 0;
-        int col_high   = ncols_y;
-        int col_diff   = ncols_y;
+        int col_high   = ncols_dst;
+        int col_diff   = ncols_dst;
         int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
         int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
 
@@ -2689,7 +2689,7 @@ static __global__ void mul_mat_q(
 
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst,
              tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
         return;
     }
@@ -2720,8 +2720,8 @@ static __global__ void mul_mat_q(
 
         // Defaults for regular matrix multiplication:
         int col_low    = 0;
-        int col_high   = ncols_y;
-        int col_diff   = ncols_y;
+        int col_high   = ncols_dst;
+        int col_diff   = ncols_dst;
         int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
         int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
 
@@ -2767,7 +2767,7 @@ static __global__ void mul_mat_q(
 
         constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst,
              tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 
         kbc += blocks_per_ne00;
@@ -2792,8 +2792,8 @@ static __global__ void mul_mat_q(
 
     // Defaults for regular matrix multiplication:
     int col_low    = 0;
-    int col_high   = ncols_y;
-    int col_diff   = ncols_y;
+    int col_high   = ncols_dst;
+    int col_diff   = ncols_dst;
     int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
     int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
 
@@ -2834,7 +2834,7 @@ static __global__ void mul_mat_q(
 
     constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-        (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+        (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst,
          tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }
 
@@ -2842,7 +2842,7 @@ static __global__ void mul_mat_q(
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 static __global__ void mul_mat_q_stream_k_fixup(
         const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
-        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_col_dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst,
         const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
@@ -2851,8 +2851,8 @@ static __global__ void mul_mat_q_stream_k_fixup(
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int ntx  = (ncols_y + mmq_x - 1) / mmq_x;
-    const int nty  = (nrows_x + mmq_y - 1) / mmq_y;
+    const int ntx  = (ncols_dst + mmq_x - 1) / mmq_x;
+    const int nty  = (nrows_x   + mmq_y - 1) / mmq_y;
 
     const int bidx0 = blockIdx.x;
 
@@ -2925,8 +2925,8 @@ static __global__ void mul_mat_q_stream_k_fixup(
         const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y;
         dst += offset_dst;
 
-        const int i_max = nrows_x - it*mmq_y - 1;
-        const int j_max = ncols_y - jt*mmq_x - 1;
+        const int i_max = nrows_x   - it*mmq_y - 1;
+        const int j_max = ncols_dst - jt*mmq_x - 1;
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
@@ -2989,7 +2989,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
 
 struct mmq_args {
     const char * x; ggml_type type_x; const int * y; const int32_t * ids_dst; const int32_t * expert_bounds; float * dst;
-    int64_t ncols_x; int64_t nrows_x; int64_t ncols_y; int64_t stride_row_x; int64_t nrows_dst;
+    int64_t ncols_x; int64_t nrows_x; int64_t ncols_dst; int64_t stride_row_x; int64_t ncols_y; int64_t nrows_dst;
     int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
     int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
     bool use_stream_k;
@@ -3025,8 +3025,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     }
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
 
-    const int nty  = (args.nrows_x + mmq_y - 1) / mmq_y;
-    const int ntx  = (args.ncols_y + mmq_x - 1) / mmq_x;
+    const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
+    const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;
     const int ntzw = args.nchannels_y * args.nsamples_y;
     const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
 
@@ -3040,14 +3040,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
             constexpr bool need_check = false;
             mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
-                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
                  channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
                  sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         } else {
             constexpr bool need_check = true;
             mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
-                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
                  channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
                  sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         }
@@ -3068,7 +3068,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
         mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
-             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
              channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
              sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
@@ -3077,14 +3077,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
         }
 
         mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
-            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
              args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     } else {
         constexpr bool need_check = true;
 
         mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
-             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
              channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
              sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
@@ -3093,7 +3093,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
         }
 
         mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
-            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
              args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     }
 }