update hw info

enhance FA stable in UT
[SYCL] ehance UPSCALE to support all UT cases (#20637 )
2026-03-31 09:24:40 +08:00 · 2026-03-17 15:57:02 +08:00 · 2026-03-17 10:01:52 +08:00 · 2026-03-17 01:42:04 +01:00 · 2026-03-17 00:31:14 +01:00 · 2026-03-16 22:38:59 +02:00
95 changed files with 3904 additions and 2321 deletions
@@ -53,10 +53,11 @@ RUN apt-get update \
    && apt-get install -y \
    build-essential \
    git \
-    python3 \
-    python3-dev \
+    python3.13 \
+    python3.13-dev \
    python3-pip \
    python3-wheel \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.13 100 \
    && pip install --break-system-packages --upgrade setuptools \
    && pip install --break-system-packages -r requirements.txt \
    && apt autoremove -y \
@@ -104,3 +104,20 @@ OpenCL:
        - any-glob-to-any-file:
            - ggml/include/ggml-opencl.h
            - ggml/src/ggml-opencl/**
+            - docs/backend/OPENCL.md
+Hexagon:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-hexagon.h
+            - ggml/src/ggml-hexagon/**
+WebGPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-webgpu.h
+            - ggml/src/ggml-webgpu/**
+OpenVINO:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-openvino.h
+            - ggml/src/ggml-openvino/**
+            - docs/backend/OPENVINO.md
@@ -0,0 +1,57 @@
+name: CI (3rd-party)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-3rd-party.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-24-llguidance:
+    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libssl-dev
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_LLGUIDANCE=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
@@ -0,0 +1,140 @@
+name: CI (android)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-android.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-android.yml',
+      'examples/llama.android/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  android:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+
+      # Disabled due to size (400MB) and always 0 cache hits
+      # - name: ccache
+      #   uses: ggml-org/ccache-action@v1.2.16
+      #   with:
+      #     key: android-build
+      #     evict-old-files: 1d
+
+      - name: Set up JDK
+        uses: actions/setup-java@v5
+        with:
+          java-version: 17
+          distribution: zulu
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v3
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Build
+        run: |
+          cd examples/llama.android
+          ./gradlew build --no-daemon
+
+  android-ndk:
+    runs-on: ubuntu-latest
+
+    env:
+      OPENCL_VERSION: 2025.07.22
+
+    strategy:
+      matrix:
+        include:
+          - build: 'arm64-cpu'
+            defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_OPENSSL=OFF -D GGML_OPENMP=OFF'
+          - build: 'arm64-snapdragon'
+            defines: '--preset arm64-android-snapdragon-release'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        run: |
+          mkdir opencl
+          curl -L -o opencl/clhpp.tar.gz      https://github.com/KhronosGroup/OpenCL-CLHPP/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          curl -L -o opencl/headers.tar.gz    https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          curl -L -o opencl/icd-loader.tar.gz https://github.com/KhronosGroup/OpenCL-ICD-Loader/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          tar -xaf opencl/headers.tar.gz    -C opencl
+          tar -xaf opencl/clhpp.tar.gz      -C opencl
+          tar -xaf opencl/icd-loader.tar.gz -C opencl
+          sudo cp -r opencl/OpenCL-Headers-${OPENCL_VERSION}/CL         ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+          sudo cp -r opencl/OpenCL-CLHPP-${OPENCL_VERSION}/include/CL/* ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/CL
+          cd opencl/OpenCL-ICD-Loader-${OPENCL_VERSION}
+          cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -DOPENCL_ICD_LOADER_HEADERS_DIR=${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=31 -DANDROID_STL=c++_shared
+          cmake --build build
+          sudo cp build/libOpenCL.so ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+          rm -rf opencl
+
+      - name: Install Hexagon SDK
+        id: install_hexsdk
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        env:
+          HEXSDK_VER: 6.4.0.2
+          HEXTLS_VER: 19.0.04
+        run: |
+          curl -L -o hex-sdk.tar.gz https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v$HEXSDK_VER/hexagon-sdk-v$HEXSDK_VER-amd64-lnx.tar.xz
+          mkdir hex-sdk
+          tar -xaf hex-sdk.tar.gz -C hex-sdk
+          ls -l hex-sdk
+          sudo mv hex-sdk /opt/hexagon
+          echo "HEXAGON_SDK_ROOT=/opt/hexagon/$HEXSDK_VER"                                     >> "$GITHUB_ENV"
+          echo "HEXAGON_TOOLS_ROOT=/opt/hexagon/$HEXSDK_VER/tools/HEXAGON_Tools/$HEXTLS_VER"   >> "$GITHUB_ENV"
+          echo "DEFAULT_HLOS_ARCH=64"                                                          >> "$GITHUB_ENV"
+          echo "DEFAULT_TOOLS_VARIANT=toolv19"                                                 >> "$GITHUB_ENV"
+          echo "DEFAULT_NO_QURT_INC=0"                                                         >> "$GITHUB_ENV"
+          echo "DEFAULT_DSP_ARCH=v73"                                                          >> "$GITHUB_ENV"
+
+      - name: Update CMake presets
+        id: update_presets
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        run: |
+          cp docs/backend/snapdragon/CMakeUserPresets.json .
+
+      - name: Build
+        id: ndk_build
+        run: |
+          cmake ${{ matrix.defines }} -B build
+          cmake --build build
+          cmake --install build --prefix pkg-adb/llama.cpp
+
+      - name: Test
+        id: cmake_test
+        run: |
+          echo "FIXME: test on devices"
@@ -0,0 +1,214 @@
+name: CI (apple)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-apple.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-apple.yml',
+      'ggml/src/ggml-metal/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  macOS-latest-ios:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-ios
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=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
+
+  macos-latest-ios-xcode:
+    runs-on: macos-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Setup Xcode
+        uses: ggml-org/setup-xcode@v1
+        with:
+          xcode-version: latest-stable
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_OPENSSL=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: Upload xcframework artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+          retention-days: 1
+
+      - name: Build Xcode project
+        run: |
+          xcodebuild -downloadPlatform iOS
+          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
+
+  macOS-latest-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-tvos
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -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
+
+  macOS-latest-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macOS-latest-swift:
+    runs-on: macos-latest
+    needs: macos-latest-ios-xcode
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-swift
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Download xcframework artifact
+        uses: actions/download-artifact@v7
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+
+      - name: Build llama.cpp with CMake
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
@@ -37,37 +37,37 @@ jobs:
          path: ./vulkan_sdk
          version: ${{ env.VULKAN_SDK_VERSION }}

-  ubuntu-24-spacemit-cache:
-    runs-on: ubuntu-24.04
+  #ubuntu-24-spacemit-cache:
+  #  runs-on: ubuntu-24.04

-    env:
-      # Make sure this is in sync with build-linux-cross.yml
-      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
+  #  env:
+  #    # Make sure this is in sync with build-linux-cross.yml
+  #    SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"

-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
+  #  steps:
+  #    - name: Clone
+  #      id: checkout
+  #      uses: actions/checkout@v6

-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-toolchain
-        with:
-          path: ./spacemit_toolchain
-          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
+  #    - name: Setup Cache
+  #      uses: actions/cache@v5
+  #      id: cache-toolchain
+  #      with:
+  #        path: ./spacemit_toolchain
+  #        key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}

-      - name: Setup SpacemiT Toolchain
-        if: steps.cache-toolchain.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-spacemit
-        with:
-          path: ./spacemit_toolchain
-          version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
+  #    - name: Setup SpacemiT Toolchain
+  #      if: steps.cache-toolchain.outputs.cache-hit != 'true'
+  #      uses: ./.github/actions/linux-setup-spacemit
+  #      with:
+  #        path: ./spacemit_toolchain
+  #        version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}

  ubuntu-24-openvino-cache:
    runs-on: ubuntu-24.04

    env:
-      # Sync versions in build.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
      OPENVINO_VERSION_MAJOR: "2026.0"
      OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"

@@ -0,0 +1,102 @@
+name: CI (cann)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-cann.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-cann.yml',
+      'ggml/src/ggml-cann/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  openEuler-latest-cann:
+    defaults:
+      run:
+        shell: bash -el {0}
+    strategy:
+      matrix:
+        arch: [x86, aarch64]
+        chip_type: ['910b', '310p']
+        build: ['Release']
+        use_acl_graph: ['on', 'off']
+        exclude:
+          # 310P does not support USE_ACL_GRAPH=on
+          - chip_type: '310p'
+            use_acl_graph: 'on'
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
+
+      - name: Set container image
+        id: cann-image
+        run: |
+          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-openeuler24.03-py3.11' }}"
+          echo "image=${image}" >> "${GITHUB_OUTPUT}"
+
+      - name: Pull container image
+        run: docker pull "${{ steps.cann-image.outputs.image }}"
+
+      - name: Build
+        env:
+          BUILD_TYPE: ${{ matrix.build }}
+          SOC_TYPE: ascend${{ matrix.chip_type }}
+          USE_ACL_GRAPH: ${{ matrix.use_acl_graph }}
+        run: |
+          HOST_UID=$(id -u)
+          HOST_GID=$(id -g)
+
+          docker run --rm \
+            -v "${PWD}:/workspace" \
+            -w /workspace \
+            -e SOC_TYPE=${SOC_TYPE} \
+            -e BUILD_TYPE=${BUILD_TYPE} \
+            -e USE_ACL_GRAPH=${USE_ACL_GRAPH} \
+            "${{ steps.cann-image.outputs.image }}" \
+            bash -lc '
+              set -e
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
+              yum clean all && rm -rf /var/cache/yum
+              git config --global --add safe.directory "/workspace"
+              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+              cmake -S . -B build \
+                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DGGML_CANN=on \
+                  -DSOC_TYPE=${SOC_TYPE} \
+                  -DUSE_ACL_GRAPH=${USE_ACL_GRAPH}
+              cmake --build build -j $(nproc)
+
+              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
+            '
@@ -1,7 +1,24 @@
-name: Build on Linux using cross-compiler
+name: CI (cross)
 on:
+  # only manual triggers due to low-importance of the workflows
+  # TODO: for regular runs, provision dedicated self-hosted runners
  workflow_dispatch:
-  workflow_call:
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-cross.yml',
+      'ggml/src/spacemit/*',
+      'ggml/src/arch/loongarch/*'
+    ]
+  # run once every week
+  schedule:
+    - cron: '0 0 * * 0'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+

 jobs:
  # ubuntu-24-riscv64-cpu-cross:
@@ -264,15 +281,15 @@ jobs:
    steps:
      - uses: actions/checkout@v6

-      - name: Use SpacemiT Toolchain Cache
-        uses: actions/cache@v5
-        id: cache-toolchain
-        with:
-          path: ./spacemit_toolchain
-          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
+      #- name: Use SpacemiT Toolchain Cache
+      #  uses: actions/cache@v5
+      #  id: cache-toolchain
+      #  with:
+      #    path: ./spacemit_toolchain
+      #    key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}

      - name: Setup SpacemiT Toolchain
-        if: steps.cache-toolchain.outputs.cache-hit != 'true'
+        #if: steps.cache-toolchain.outputs.cache-hit != 'true'
        uses: ./.github/actions/linux-setup-spacemit
        with:
          path: ./spacemit_toolchain
@@ -0,0 +1,72 @@
+name: CI (msys)
+
+on:
+  # only manual triggers due to low-importance of the workflows
+  # TODO: for regular runs, provision dedicated self-hosted runners
+  workflow_dispatch:
+  # run once every week
+  schedule:
+    - cron: '0 0 * * 0'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  windows-msys2:
+    runs-on: windows-2025
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
+          - { sys: CLANG64, env: clang-x86_64, build: Release }
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+
+      #- name: ccache
+      #  uses: ggml-org/ccache-action@v1.2.16
+      #  with:
+      #    key: windows-msys2
+      #    variant: ccache
+      #    evict-old-files: 1d
+      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Setup ${{ matrix.sys }}
+        uses: msys2/setup-msys2@v2
+        with:
+          update: true
+          msystem: ${{matrix.sys}}
+          install: >-
+            base-devel
+            git
+            mingw-w64-${{matrix.env}}-toolchain
+            mingw-w64-${{matrix.env}}-cmake
+            mingw-w64-${{matrix.env}}-openblas
+
+      - name: Build using CMake
+        shell: msys2 {0}
+        run: |
+            cmake -B build
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
+      - name: Clean after building using CMake
+        shell: msys2 {0}
+        run: |
+            rm -rf build
+
+      - name: Build using CMake w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
@@ -0,0 +1,136 @@
+name: CI (riscv)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-riscv.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-riscv.yml',
+      'ggml/src/ggml-cpu/arch/riscv/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-riscv64-native-sanitizer:
+    runs-on: RISCV64
+
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug]
+
+    steps:
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+
+          # Install necessary packages
+          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 rustup cmake build-essential wget ccache git-lfs
+
+          # Set gcc-14 and g++-14 as the default compilers
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
+          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
+          sudo ln -sf /usr/bin/gcc-14 /usr/bin/gcc
+          sudo ln -sf /usr/bin/g++-14 /usr/bin/g++
+
+          # Install Rust stable version
+          rustup install stable
+          rustup default stable
+
+          git lfs install
+
+      - name: GCC version check
+        run: |
+          gcc --version
+          g++ --version
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Setup ccache
+        run: |
+          # Unique cache directory per matrix combination
+          export CCACHE_DIR="$HOME/.ccache/sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}"
+          mkdir -p "$CCACHE_DIR"
+
+          # Configure ccache
+          ccache --set-config=max_size=5G
+          ccache --set-config=compression=true
+          ccache --set-config=compression_level=6
+          ccache --set-config=cache_dir="$CCACHE_DIR"
+          ccache --set-config=sloppiness=file_macro,time_macros,include_file_mtime,include_file_ctime
+          ccache --set-config=hash_dir=false
+
+          # Export for subsequent steps
+          echo "CCACHE_DIR=$CCACHE_DIR" >> $GITHUB_ENV
+          echo "PATH=/usr/lib/ccache:$PATH" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        if: ${{ matrix.sanitizer != 'THREAD' }}
+        run: |
+          cmake -B build \
+            -DLLAMA_OPENSSL=OFF \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+            -DGGML_OPENMP=ON \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          cmake -B build \
+            -DLLAMA_OPENSSL=OFF \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
@@ -0,0 +1,87 @@
+name: CI (sanitize)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-sanitize.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-latest-sanitizer:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-latest-sanitizer-${{ matrix.sanitizer }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libssl-dev
+
+      - name: Build
+        id: cmake_build
+        if: ${{ matrix.sanitizer != 'THREAD' }}
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+            -DGGML_OPENMP=OFF
+
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
@@ -0,0 +1,245 @@
+name: CI (self-hosted)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp',
+      '**/*.glsl',
+      '**/*.wgsl'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-self-hosted.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp',
+      '**/*.glsl',
+      '**/*.wgsl'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ggml-ci-nvidia-cuda:
+    runs-on: [self-hosted, Linux, NVIDIA]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          nvidia-smi
+          GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-nvidia-vulkan-cm:
+    runs-on: [self-hosted, Linux, NVIDIA]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-nvidia-vulkan-cm2:
+    runs-on: [self-hosted, Linux, NVIDIA, COOPMAT2]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  # TODO: provision AMX-compatible machine
+  #ggml-ci-cpu-amx:
+  #  runs-on: [self-hosted, Linux, CPU, AMX]
+
+  #  steps:
+  #    - name: Clone
+  #      id: checkout
+  #      uses: actions/checkout@v6
+
+  #    - name: Test
+  #      id: ggml-ci
+  #      run: |
+  #        bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  # TODO: provision AMD GPU machine
+  # ggml-ci-amd-vulkan:
+  #   runs-on: [self-hosted, Linux, AMD]
+
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v6
+
+  #     - name: Test
+  #       id: ggml-ci
+  #       run: |
+  #         vulkaninfo --summary
+  #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  # TODO: provision AMD GPU machine
+  # ggml-ci-amd-rocm:
+  #   runs-on: [self-hosted, Linux, AMD]
+
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v6
+
+  #     - name: Test
+  #       id: ggml-ci
+  #       run: |
+  #         amd-smi static
+  #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-mac-metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  ggml-ci-mac-webgpu:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          DAWN_VERSION="v2.0.0"
+          DAWN_OWNER="reeselevine"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+          curl -L -o artifact.zip \
+            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+          mkdir dawn
+          unzip artifact.zip
+          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  ggml-ci-mac-vulkan:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  ggml-ci-linux-intel-vulkan:
+    runs-on: [self-hosted, Linux, Intel]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  ggml-ci-intel-openvino-gpu-low-perf:
+    runs-on: [self-hosted, Linux, Intel, OpenVINO]
+
+    env:
+      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.0"
+      OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Setup OpenVINO Toolkit
+        uses: ./.github/actions/linux-setup-openvino
+        with:
+          path: ./openvino_toolkit
+          version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
+          version_full: ${{ env.OPENVINO_VERSION_FULL }}
+
+      - name: Install OpenVINO dependencies
+        run: |
+          cd ./openvino_toolkit
+          chmod +x ./install_dependencies/install_openvino_dependencies.sh
+          echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          source ./openvino_toolkit/setupvars.sh
+          GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
@@ -0,0 +1,96 @@
+name: CI (vulkan)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-vulkan.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.comp',
+      '**/*.glsl'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-vulkan.yml',
+      'ggml/src/ggml-vulkan/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-24-vulkan-llvmpipe:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-24-vulkan-llvmpipe
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo add-apt-repository -y ppa:kisak/kisak-mesa
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
+
+      - name: Get latest Vulkan SDK version
+        id: vulkan_sdk_version
+        run: |
+          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
+
+      - name: Use Vulkan SDK Cache
+        uses: actions/cache@v5
+        id: cache-sdk
+        with:
+          path: ./vulkan_sdk
+          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+
+      - name: Setup Vulkan SDK
+        if: steps.cache-sdk.outputs.cache-hit != 'true'
+        uses: ./.github/actions/linux-setup-vulkan-llvmpipe
+        with:
+          path: ./vulkan_sdk
+          version: ${{ env.VULKAN_SDK_VERSION }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source ./vulkan_sdk/setup-env.sh
+          cmake -B build \
+            -DGGML_VULKAN=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          export GGML_VK_VISIBLE_DEVICES=0
+          export GGML_VK_DISABLE_F16=1
+          export GGML_VK_DISABLE_COOPMAT=1
+          # This is using llvmpipe and runs slower than other backends
+          ctest -L main --verbose --timeout 4800
@@ -4,10 +4,16 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/python-lint.yml', '**/*.py']
+    paths: [
+      '.github/workflows/python-lint.yml',
+      '**/*.py'
+    ]
  pull_request:
    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/python-lint.yml', '**/*.py']
+    paths: [
+      '.github/workflows/python-lint.yml',
+      '**/*.py'
+    ]

 concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -10,7 +10,22 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/release.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/release.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp',
+      '**/*.glsl'
+    ]

 concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -34,7 +49,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: macOS-latest-cmake-arm64
+          key: macOS-latest-arm64
          evict-old-files: 1d

      - name: Build
@@ -81,7 +96,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: macOS-latest-cmake-x64
+          key: macOS-latest-x64
          evict-old-files: 1d

      - name: Build
@@ -140,7 +155,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: ubuntu-cpu-cmake-${{ matrix.build }}
+          key: ubuntu-cpu-${{ matrix.build }}
          evict-old-files: 1d

      - name: Dependencies
@@ -191,7 +206,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: ubuntu-22-cmake-vulkan
+          key: ubuntu-22-vulkan
          evict-old-files: 1d

      - name: Dependencies
@@ -238,7 +253,7 @@ jobs:
      openvino_version: ${{ steps.openvino_version.outputs.value }}

    env:
-      # Sync versions in build.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
      OPENVINO_VERSION_MAJOR: "2026.0"
      OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"

@@ -256,7 +271,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: ubuntu-24-cmake-openvino-release-no-preset-v1
+          key: ubuntu-24-openvino-release-no-preset-v1
          evict-old-files: 1d

      - name: Dependencies
@@ -329,7 +344,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: windows-latest-cmake-cpu-${{ matrix.arch }}
+          key: windows-latest-cpu-${{ matrix.arch }}
          variant: ccache
          evict-old-files: 1d

@@ -390,7 +405,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
+          key: windows-latest-${{ matrix.backend }}-${{ matrix.arch }}
          variant: ccache
          evict-old-files: 1d

@@ -536,7 +551,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: windows-latest-cmake-sycl
+          key: windows-latest-sycl
          variant: ccache
          evict-old-files: 1d

@@ -616,7 +631,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
          evict-old-files: 1d

      - name: Dependencies
@@ -726,7 +741,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}-x64
+          key: windows-latest-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}-x64
          evict-old-files: 1d

      - name: Install ROCm
@@ -0,0 +1,105 @@
+name: Server (sanitize)
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      slow_tests:
+        description: 'Run slow tests'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/server-sanitize.yml',
+      '**/CMakeLists.txt',
+      '**/Makefile',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      'tools/server/**.*'
+    ]
+
+env:
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  server:
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
+        build_type: [RelWithDebInfo]
+      fail-fast: false
+
+    steps:
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get -y install \
+            build-essential \
+            xxd \
+            git \
+            cmake \
+            curl \
+            wget \
+            language-pack-en \
+            libssl-dev
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
+            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Python setup
+        id: setup_python
+        uses: actions/setup-python@v6
+        with:
+          python-version: '3.11'
+          pip-install: -r tools/server/tests/requirements.txt
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+        run: |
+          cd tools/server/tests
+          export ${{ matrix.extra_args }}
+          pytest -v -x -m "not slow"
+
+      - name: Slow tests
+        id: server_integration_tests_slow
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
+        run: |
+          cd tools/server/tests
+          export ${{ matrix.extra_args }}
+          SLOW_TESTS=1 pytest -v -x
@@ -1,4 +1,4 @@
-name: Server-Metal
+name: Server (self-hosted)

 on:
  workflow_dispatch: # allows manual triggering
@@ -14,7 +14,19 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+    paths: [
+      '.github/workflows/server-self-hosted.yml',
+      '**/CMakeLists.txt',
+      '**/Makefile',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.swift',
+      '**/*.m',
+      'tools/server/**.*'
+    ]

 env:
  LLAMA_LOG_COLORS: 1
@@ -28,7 +40,7 @@ concurrency:

 jobs:
  server-metal:
-    runs-on: [self-hosted, macOS, ARM64]
+    runs-on: [self-hosted, llama-server, macOS, ARM64]

    name: server-metal (${{ matrix.wf_name }})
    strategy:
@@ -71,3 +83,42 @@ jobs:
          pip install -r requirements.txt
          export ${{ matrix.extra_args }}
          pytest -v -x -m "not slow"
+
+  server-cuda:
+    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
+
+    name: server-cuda (${{ matrix.wf_name }})
+    strategy:
+      matrix:
+        build_type: [Release]
+        wf_name: ["GPUx1"]
+        include:
+          - build_type: Release
+            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx1, backend-sampling"
+      fail-fast: false
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
+          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+        run: |
+          cd tools/server/tests
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+          export ${{ matrix.extra_args }}
+          pytest -v -x -m "not slow"
@@ -1,4 +1,3 @@
-# Server WebUI build and tests
 name: Server WebUI

 on:
@@ -11,10 +10,20 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
+    paths: [
+      '.github/workflows/server-webui.yml',
+      'tools/server/webui/**.*',
+      'tools/server/tests/**.*',
+      'tools/server/public/**'
+    ]
  pull_request:
    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
+    paths: [
+      '.github/workflows/server-webui.yml',
+      'tools/server/webui/**.*',
+      'tools/server/tests/**.*',
+      'tools/server/public/**'
+    ]

 env:
  LLAMA_LOG_COLORS: 1
@@ -1,4 +1,3 @@
-# Server build and tests
 name: Server

 on:
@@ -15,10 +14,34 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+    paths: [
+      '.github/workflows/server.yml',
+      '**/CMakeLists.txt',
+      '**/Makefile',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.swift',
+      '**/*.m',
+      'tools/server/**.*'
+    ]
  pull_request:
    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+    paths: [
+      '.github/workflows/server.yml',
+      '**/CMakeLists.txt',
+      '**/Makefile',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.swift',
+      '**/*.m',
+      'tools/server/**.*'
+    ]

 env:
  LLAMA_LOG_COLORS: 1
@@ -34,17 +57,18 @@ jobs:
  server:
    runs-on: ubuntu-latest

+    name: server (${{ matrix.wf_name }})
    strategy:
      matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
-        build_type: [RelWithDebInfo]
+        build_type: [Release]
+        wf_name: ["default"]
        include:
          - build_type: Release
-            sanitizer: ""
            extra_args: ""
+            wf_name:    "default"
          - build_type: Release
-            sanitizer: ""
            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "backend-sampling"
      fail-fast: false

    steps:
@@ -74,13 +98,7 @@ jobs:
        run: |
          cmake -B build \
            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
-            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
+            -DGGML_SCHED_NO_REALLOC=ON
          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server

      - name: Python setup
@@ -2,29 +2,13 @@
 # multiplie collaborators per item can be specified

 /.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @CISC
-/.github/workflows/                     @CISC
+/.github/actions/                       @ggml-org/ci
+/.github/workflows/                     @ggml-org/ci
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
-/common/CMakeLists.txt                  @ggerganov
-/common/arg.*                           @ggerganov
-/common/base64.hpp.*                    @ggerganov
-/common/build-info.*                    @ggerganov
-/common/chat.*                          @pwilkin
-/common/chat-auto*.*                    @pwilkin
-/common/chat-diff-analyzer.*            @pwilkin
-/common/chat-peg-parser.*               @aldehir
-/common/common.*                        @ggerganov
-/common/console.*                       @ggerganov
-/common/http.*                          @angt
-/common/jinja/                          @ngxson @CISC @aldehir
-/common/llguidance.*                    @ggerganov
-/common/log.*                           @ggerganov
+/common/                                @ggml-org/llama-common
+/common/jinja/                          @CISC
 /common/ngram-map.*                     @srogmann
-/common/peg-parser.*                    @aldehir
-/common/sampling.*                      @ggerganov
-/common/speculative.*                   @ggerganov
-/common/unicode.*                       @aldehir
 /convert_*.py                           @CISC
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
@@ -51,29 +35,27 @@
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov
 /ggml/include/                          @ggerganov
+/ggml/src/ggml-cann/                    @ggml-org/ggml-cann
 /ggml/src/ggml-common.h                 @ggerganov
 /ggml/src/ggml-cpu/                     @ggerganov
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
-/ggml/src/ggml-cuda/fattn*              @JohannesGaessler
-/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
-/ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
-/ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
-/ggml/src/ggml-cuda/mmvq.*              @JohannesGaessler
+/ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
 /ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
 /ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
 /ggml/src/ggml-impl.h                   @ggerganov
-/ggml/src/ggml-metal/                   @ggerganov
-/ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
-/ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
+/ggml/src/ggml-metal/                   @ggml-org/ggml-metal
+/ggml/src/ggml-opencl/                  @ggml-org/ggml-opencl
+/ggml/src/ggml-hexagon/                 @ggml-org/ggml-hexagon
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
-/ggml/src/ggml-rpc/                     @rgerganov
+/ggml/src/ggml-rpc/                     @ggml-org/ggml-rpc
+/ggml/src/ggml-sycl/                    @ggml-org/ggml-sycl
 /ggml/src/ggml-threading.*              @ggerganov
-/ggml/src/ggml-vulkan/                  @0cc4m
+/ggml/src/ggml-vulkan/                  @ggml-org/ggml-vulkan
 /ggml/src/ggml-virtgpu/                 @kpouget
-/ggml/src/ggml-webgpu/                  @reeselevine
-/ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
+/ggml/src/ggml-webgpu/                  @ggml-org/ggml-webgpu
+/ggml/src/ggml-zdnn/                    @ggml-org/ggml-zdnn @Andreas-Krebbel @AlekseiNikiforovIBM
 /ggml/src/ggml-openvino/                @cavusmustafa @wine99
 /ggml/src/ggml.c                        @ggerganov
 /ggml/src/ggml.cpp                      @ggerganov
@@ -93,16 +75,18 @@
 /src/models/                            @CISC
 /tests/                                 @ggerganov
 /tests/test-chat.*                      @pwilkin
+/tests/test-llama-archs.cpp             @JohannesGaessler
 /tools/batched-bench/                   @ggerganov
 /tools/cli/                             @ngxson
 /tools/completion/                      @ggerganov
-/tools/mtmd/                            @ngxson
+/tools/mtmd/                            @ggml-org/llama-mtmd
 /tools/perplexity/                      @ggerganov
 /tools/parser/                          @pwilkin
 /tools/quantize/                        @ggerganov
-/tools/rpc/                             @rgerganov
-/tools/server/*                         @ngxson @ggerganov # no subdir
-/tools/server/webui/                    @allozaur
+/tools/rpc/                             @ggml-org/ggml-rpc
+/tools/server/*                         @ggml-org/llama-server # no subdir
+/tools/server/tests/                    @ggml-org/llama-server
+/tools/server/webui/                    @ggml-org/llama-webui
 /tools/tokenize/                        @ggerganov
 /tools/tts/                             @ggerganov
 /vendor/                                @ggerganov
@@ -24,9 +24,9 @@ Fri Mar  6 11:39:45 2026
 +-----------------------------------------+------------------------+----------------------+
 ```

-## ggml-org/nemotron-3-super-120b-GGUF
+## ggml-org/Nemotron-3-Super-120B-GGUF

-Model: https://huggingface.co/ggml-org/nemotron-3-super-120b-GGUF
+Model: https://huggingface.co/ggml-org/Nemotron-3-Super-120B-GGUF

 - `llama-batched-bench`

@@ -53,7 +53,6 @@ main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_
 |  8192 |     32 |   16 | 131584 |  171.066 |   766.21 |   10.774 |    47.52 |  181.840 |   723.62 |
 |  8192 |     32 |   32 | 263168 |  342.140 |   766.19 |   18.969 |    53.98 |  361.109 |   728.78 |

-
 - `llama-bench`

 | model                   |       size |     params | backend    | n_ubatch | fa |            test |                  t/s |
@@ -70,3 +69,49 @@ main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_
 | nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |   tg32 @ d32768 |         19.45 ± 0.18 |

 build: 04a65daab (8268)
+
+## ggml-org/Nemotron-3-Nano-4B-GGUF
+
+Model: https://huggingface.co/ggml-org/Nemotron-3-Nano-4B-GGUF
+
+- `llama-batched-bench`
+
+main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = 99, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.152 |  3371.61 |    0.597 |    53.64 |    0.748 |   726.90 |
+|   512 |     32 |    2 |   1088 |    0.319 |  3208.68 |    0.857 |    74.66 |    1.176 |   924.89 |
+|   512 |     32 |    4 |   2176 |    0.720 |  2843.56 |    1.323 |    96.78 |    2.043 |  1065.18 |
+|   512 |     32 |    8 |   4352 |    1.428 |  2867.96 |    2.311 |   110.76 |    3.739 |  1163.82 |
+|   512 |     32 |   16 |   8704 |    2.857 |  2866.94 |    4.203 |   121.82 |    7.060 |  1232.82 |
+|   512 |     32 |   32 |  17408 |    5.709 |  2869.76 |    7.964 |   128.58 |   13.673 |  1273.14 |
+|  4096 |     32 |    1 |   4128 |    1.458 |  2809.76 |    0.605 |    52.92 |    2.062 |  2001.52 |
+|  4096 |     32 |    2 |   8256 |    2.905 |  2819.95 |    0.875 |    73.12 |    3.780 |  2183.95 |
+|  4096 |     32 |    4 |  16512 |    5.790 |  2829.74 |    1.361 |    94.07 |    7.151 |  2309.17 |
+|  4096 |     32 |    8 |  33024 |   11.598 |  2825.32 |    2.378 |   107.65 |   13.976 |  2362.89 |
+|  4096 |     32 |   16 |  66048 |   23.208 |  2823.88 |    4.348 |   117.76 |   27.556 |  2396.89 |
+|  4096 |     32 |   32 | 132096 |   46.515 |  2817.85 |    8.279 |   123.69 |   54.794 |  2410.79 |
+|  8192 |     32 |    1 |   8224 |    2.950 |  2776.95 |    0.617 |    51.89 |    3.567 |  2305.75 |
+|  8192 |     32 |    2 |  16448 |    5.921 |  2767.32 |    0.896 |    71.45 |    6.816 |  2413.05 |
+|  8192 |     32 |    4 |  32896 |   11.842 |  2767.21 |    1.401 |    91.34 |   13.243 |  2484.03 |
+|  8192 |     32 |    8 |  65792 |   23.726 |  2762.17 |    2.461 |   104.03 |   26.187 |  2512.38 |
+|  8192 |     32 |   16 | 131584 |   47.777 |  2743.43 |    4.577 |   111.86 |   52.354 |  2513.36 |
+|  8192 |     32 |   32 | 263168 |   96.691 |  2711.16 |    8.772 |   116.73 |  105.463 |  2495.36 |
+
+- `llama-bench`
+
+| model                   |       size |     params | backend    | n_ubatch | fa |            test |                  t/s |
+| ----------------------- | ---------: | ---------: | ---------- | -------: | -: | --------------: | -------------------: |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |          pp2048 |      2761.90 ± 19.31 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |            tg32 |         52.85 ± 0.12 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |  pp2048 @ d4096 |      2687.07 ± 21.84 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |    tg32 @ d4096 |         52.32 ± 0.23 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |  pp2048 @ d8192 |      2564.52 ± 57.69 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |    tg32 @ d8192 |         51.27 ± 0.34 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 | pp2048 @ d16384 |      2334.02 ± 37.83 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |   tg32 @ d16384 |         49.71 ± 0.14 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 | pp2048 @ d32768 |      2041.46 ± 40.45 |
+| nemotron 4B Q8_0        |   3.94 GiB |     3.97 B | CUDA       |     2048 |  1 |   tg32 @ d32768 |         46.71 ± 0.13 |
+
+build: 1bbec6a75 (8382)
@@ -479,6 +479,7 @@ analyze_content::analyze_content(const common_chat_template & tmpl, const analyz

    if (!comparison_with_tools || !comparison_with_reasoning) {
        LOG_DBG(ANSI_ORANGE "%s: Template application failed\n" ANSI_RESET, __func__);
+        return;
    }

    const auto & diff_tools     = comparison_with_tools->diff;
@@ -911,8 +912,10 @@ void analyze_tools::extract_function_markers() {
            // we'll have to rely on an extra diff with no-calls version
            auto notool_comp = compare_variants(
                *tmpl, params, [&](template_params & p) { p.messages = json::array({ user_msg, assistant_nocall }); });
-            auto nt_diff  = notool_comp->diff;
-            closer_suffix = nt_diff.left.substr(nt_diff.left.find("YYYY") + 4);
+            if (notool_comp) {
+                auto nt_diff  = notool_comp->diff;
+                closer_suffix = nt_diff.left.substr(nt_diff.left.find("YYYY") + 4);
+            }
        } else {
            closer_suffix = diff.suffix.substr(0, diff.suffix.find(suffix_marker));
        }
@@ -1519,7 +1519,6 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        // map developer to system for all models except for GPT-OSS
        workaround::map_developer_role_to_system(params.messages);
    }
-    workaround::func_args_not_string(params.messages);

    if (!tmpl.original_caps().supports_system_role) {
        workaround::system_message_not_supported(params.messages);
@@ -1532,6 +1531,10 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        workaround::requires_non_null_content(params.messages);
    }

+    if (tmpl.original_caps().supports_object_arguments) {
+        workaround::func_args_not_string(params.messages);
+    }
+
    params.extra_context = common_chat_extra_context();
    for (auto el : inputs.chat_template_kwargs) {
        params.extra_context[el.first] = json::parse(el.second);
@@ -75,6 +75,7 @@ std::map<std::string, bool> caps::to_map() const {
        {"supports_parallel_tool_calls", supports_parallel_tool_calls},
        {"supports_system_role", supports_system_role},
        {"supports_preserve_reasoning", supports_preserve_reasoning},
+        {"supports_object_arguments", supports_object_arguments},
    };
 }

@@ -158,9 +159,9 @@ caps caps_get(jinja::program & prog) {
        }
    );

-    JJ_DEBUG("%s\n", ">>> Running capability check: single tool support");
+    JJ_DEBUG("%s\n", ">>> Running capability check: single tool with object arguments support");

-    // case: tools support: single call
+    // case: tools support: single call with object arguments
    caps_try_execute(
        prog,
        [&]() {
@@ -226,9 +227,7 @@ caps caps_get(jinja::program & prog) {
        },
        [&](bool success, value & messages, value & tools) {
            if (!success) {
-                result.supports_tool_calls = false;
-                result.supports_tools = false;
-                return;
+                return; // Nothing can be inferred
            }

            auto & tool_name = tools->at(0)->at("function")->at("name");
@@ -242,16 +241,117 @@ caps caps_get(jinja::program & prog) {
            caps_print_stats(tool_calls, "messages[1].tool_calls");
            if (!tool_calls->stats.used) {
                result.supports_tool_calls = false;
+                return;
+            }
+
+            auto & tool_arg = tool_calls->at(0)->at("function")->at("arguments")->at("arg");
+            caps_print_stats(tool_arg, "messages[1].tool_calls[0].function.arguments.arg");
+            if (tool_arg->stats.used) {
+                result.supports_object_arguments = true;
            }
        }
    );

+    if (!result.supports_object_arguments) {
+        JJ_DEBUG("%s\n", ">>> Running capability check: single tool with string arguments support");
+
+        // case: tools support: single call with string arguments
+        caps_try_execute(
+            prog,
+            [&]() {
+                // messages
+                return json::array({
+                    {
+                        {"role", "user"},
+                        {"content", "User message"},
+                    },
+                    {
+                        {"role", "assistant"},
+                        {"content", ""}, // Some templates expect content to be empty with tool calls
+                        {"tool_calls", json::array({
+                            {
+                                {"id", "call00001"},
+                                {"type", "function"},
+                                {"function", {
+                                    {"name", "tool1"},
+                                    {"arguments", R"({"arg": "value"})"}
+                                }}
+                            }
+                        })}
+                    },
+                    {
+                        {"role", "tool"},
+                        {"content", "Tool response"},
+                        {"tool_call_id", "call00001"}
+                    },
+                    {
+                        {"role", "assistant"},
+                        {"content", "The tool response was 'tool response'"}
+                    },
+                    {
+                        {"role", "user"},
+                        {"content", "User message"},
+                    },
+                });
+            },
+            [&]() {
+                // tools
+                return json::array({
+                    {
+                        {"name", "tool"},
+                        {"type", "function"},
+                        {"function", {
+                            {"name", "tool1"},
+                            {"description", "Tool description"},
+                            {"parameters", {
+                                {"type", "object"},
+                                {"properties", {
+                                    {"arg", {
+                                        {"type", "string"},
+                                        {"description", "Arg description"},
+                                    }},
+                                }},
+                                {"required", json::array({ "arg" })},
+                            }},
+                        }},
+                    },
+                });
+            },
+            [&](bool success, value & messages, value & tools) {
+                if (!success) {
+                    result.supports_tool_calls = false;
+                    result.supports_tools = false;
+                    return;
+                }
+
+                auto & tool_name = tools->at(0)->at("function")->at("name");
+                caps_print_stats(tool_name, "tools[0].function.name");
+                caps_print_stats(tools, "tools");
+                if (!tool_name->stats.used) {
+                    result.supports_tools = false;
+                }
+
+                auto & tool_calls = messages->at(1)->at("tool_calls");
+                caps_print_stats(tool_calls, "messages[1].tool_calls");
+                if (!tool_calls->stats.used) {
+                    result.supports_tool_calls = false;
+                    return;
+                }
+            }
+        );
+    }
+
    JJ_DEBUG("%s\n", ">>> Running capability check: parallel tool support");

    // case: tools support: parallel calls
    caps_try_execute(
        prog,
        [&]() {
+            json args = json(R"({"arg": "value"})");
+            if (result.supports_object_arguments) {
+                args = json{{"arg", "value"}};
+            }
+
            // messages
            return json::array({
                {
@@ -267,9 +367,7 @@ caps caps_get(jinja::program & prog) {
                            {"type", "function"},
                            {"function", {
                                {"name", "tool1"},
-                                {"arguments", {
-                                    {"arg", "value"}
-                                }}
+                                {"arguments", args}
                            }}
                        },
                        {
@@ -277,9 +375,7 @@ caps caps_get(jinja::program & prog) {
                            {"type", "function"},
                            {"function", {
                                {"name", "tool1"},
-                                {"arguments", {
-                                    {"arg", "value"}
-                                }}
+                                {"arguments", args}
                            }}
                        }
                    })}
@@ -328,7 +424,7 @@ caps caps_get(jinja::program & prog) {
                return;
            }

-            auto & tool_calls = messages->at(1)->at("tool_calls");;
+            auto & tool_calls = messages->at(1)->at("tool_calls");
            caps_print_stats(tool_calls, "messages[1].tool_calls");

            // check for second tool call usage
@@ -18,6 +18,8 @@ struct caps {
    bool supports_string_content = true;
    bool supports_typed_content = false;

+    bool supports_object_arguments = false;
+
    // for reporting on server
    std::map<std::string, bool> to_map() const;

@@ -102,7 +102,7 @@ std::string regex_to_reversed_partial_regex(const std::string & pattern) {
                auto is_star = *it == '*';
                ++it;
                if (is_star) {
-                    if (*it == '?') {
+                    if (it != end && *it == '?') {
                        ++it;
                    }
                }
@@ -272,8 +272,9 @@ class ModelBase:
        return tensors

    def dequant_model(self):
-        if self._is_nvfp4:
-            return  # NVFP4 weights are repacked in _generate_nvfp4_tensors
+        # If all quantized tensors were already handled (e.g. pure NVFP4), skip
+        if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv")) for k in self.model_tensors):
+            return

        tensors_to_remove: list[str] = []
        new_tensors: dict[str, Callable[[], Tensor]] = {}
@@ -297,11 +298,16 @@ class ModelBase:
                scale = scale.float()

                if block_size is not None:
+                    dim_offset = scale.ndim - len(block_size)
                    for i, size in enumerate(block_size):
-                        scale = scale.repeat_interleave(size, i)
+                        scale = scale.repeat_interleave(size, dim_offset + i)
                    # unpad the scale (e.g. when the tensor size isn't a multiple of the block size)
                    scale = scale[tuple(slice(0, size) for size in weight.shape)]

+                # align scale dims to weight for correct broadcasting (e.g. [128] -> [128, 1, 1])
+                while scale.ndim < weight.ndim:
+                    scale = scale.unsqueeze(-1)
+
                return weight.float() * scale

            # ref: https://github.com/ModelCloud/GPTQModel/blob/037c5c0f6c9e33c500d975b038d02e7ca437546d/gptqmodel/nn_modules/qlinear/__init__.py#L437-L476
@@ -392,7 +398,7 @@ class ModelBase:
            elif quant_method == "fp8":
                block_size = quant_config.get("weight_block_size")
                for name in self.model_tensors.keys():
-                    if name.endswith(".weight_scale_inv"):
+                    if name.endswith("_scale_inv"):
                        weight_name = name.removesuffix("_scale_inv")
                        w = self.model_tensors[weight_name]
                        s = self.model_tensors[name]
@@ -400,6 +406,8 @@ class ModelBase:
                        tensors_to_remove.append(name)
                    if name.endswith(".activation_scale"):  # unused
                        tensors_to_remove.append(name)
+                    if name.endswith("_activation_scale"):  # Mistral-Small-4-119B-2602, unused
+                        tensors_to_remove.append(name)
                    # mistral format
                    if name.endswith(".qscale_weight"):
                        weight_name = name.removesuffix("qscale_weight") + "weight"
@@ -474,7 +482,20 @@ class ModelBase:
                                tensors_to_remove.append(base_name + "_zero_point")
                else:
                    raise NotImplementedError(f"Quant format {quant_format!r} for method {quant_method!r} is not yet supported")
-            else:
+            elif quant_method == "modelopt":
+                # Mixed-precision ModelOpt models: NVFP4 tensors are handled by
+                # _generate_nvfp4_tensors; FP8 tensors have 1D weight_scale and
+                # are dequantized here. input_scale tensors are unused.
+                for name in self.model_tensors.keys():
+                    if name.endswith(".weight_scale"):
+                        weight_name = name.removesuffix("_scale")
+                        w = self.model_tensors[weight_name]
+                        s = self.model_tensors[name]
+                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
+                        tensors_to_remove.append(name)
+                    if name.endswith((".input_scale", ".k_scale", ".v_scale")):
+                        tensors_to_remove.append(name)
+            elif quant_method is not None:
                raise NotImplementedError(f"Quant method is not yet supported: {quant_method!r}")

        for name in tensors_to_remove:
@@ -520,12 +541,6 @@ class ModelBase:
        raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # skip NVFP4 auxiliary tensors (handled in _generate_nvfp4_tensors)
-        if self._is_nvfp4:
-            if name.endswith((".weight_scale", ".weight_scale_2", ".input_scale", ".k_scale", ".v_scale")):
-                return []
-            if name.endswith(".weight") and name.replace(".weight", ".weight_scale") in self.model_tensors:
-                return []

        new_name = self.map_tensor_name(name)

@@ -609,6 +624,7 @@ class ModelBase:
        expert_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
        expert_shapes: dict[tuple[int, str], list[int]] = {}
        n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=True) or 0
+        consumed: list[str] = []

        for name in list(self.model_tensors.keys()):
            if not name.endswith(".weight"):
@@ -620,8 +636,18 @@ class ModelBase:
            # Force eager materialization of lazy tensors
            weight = LazyTorchTensor.to_eager(self.model_tensors[name]())
            scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
+
+            # Skip non-NVFP4 tensors (e.g. FP8 with per-channel 1D scales)
+            if scale.ndim < 2:
+                continue
+
            scale2 = LazyTorchTensor.to_eager(self.model_tensors.get(scale2_name, lambda: torch.tensor(1.0))())

+            # Mark tensors for removal from model_tensors (already written to gguf)
+            consumed.extend([name, scale_name])
+            if scale2_name in self.model_tensors:
+                consumed.append(scale2_name)
+
            # Check if this is a per-expert tensor
            m = re.search(r'\.experts\.(\d+)\.(gate_proj|up_proj|down_proj)\.weight$', name)
            if m:
@@ -652,6 +678,15 @@ class ModelBase:
        for (bid, proj_type) in list(expert_blocks.keys()):
            self._flush_nvfp4_experts((bid, proj_type), expert_blocks, expert_scales, expert_shapes, bid, proj_type)

+        # Remove consumed tensors so get_tensors/modify_tensors won't see them
+        for name in consumed:
+            self.model_tensors.pop(name, None)
+
+        # Remove unused auxiliary tensors (input_scale, k_scale, v_scale)
+        for name in list(self.model_tensors.keys()):
+            if name.endswith((".input_scale", ".k_scale", ".v_scale")):
+                del self.model_tensors[name]
+
    def _flush_nvfp4_experts(self, key, expert_blocks, expert_scales, expert_shapes, bid, proj_type):
        experts = expert_blocks.pop(key)
        scales = expert_scales.pop(key)
@@ -677,20 +712,31 @@ class ModelBase:
    def prepare_tensors(self):
        # detect NVFP4 quantization (ModelOpt format)
        quant_algo = (self.hparams.get("quantization_config") or {}).get("quant_algo")
+        quant_layers = (self.hparams.get("quantization_config") or {}).get("quantized_layers") or {}
        quant_config_file = self.dir_model / "hf_quant_config.json"

-        if not quant_algo and quant_config_file.is_file():
+        if (not quant_algo or not quant_layers) and quant_config_file.is_file():
            with open(quant_config_file, "r", encoding="utf-8") as f:
-                quant_algo = (json.load(f).get("quantization") or {}).get("quant_algo")
+                quant_config = json.load(f).get("quantization") or {}
+                quant_algo = quant_config.get("quant_algo", quant_algo)
+                quant_layers = quant_config.get("quantized_layers", quant_layers) or {}
+
+        # Some models use per-tensor quant_algo (e.g. "MIXED_PRECISION" with
+        # per-layer NVFP4/FP8) instead of a single global "NVFP4" value.
+        if quant_algo != "NVFP4":
+            if any(v.get("quant_algo") == "NVFP4" for v in quant_layers.values() if isinstance(v, dict)):
+                quant_algo = "NVFP4"

        self._is_nvfp4 = quant_algo == "NVFP4"

-        self.dequant_model()
-
-        # NVFP4 weights are repacked and written directly to gguf_writer
+        # NVFP4 weights are repacked and written directly to gguf_writer.
+        # This must run before dequant_model so NVFP4 tensors are removed
+        # from model_tensors, leaving only non-NVFP4 (e.g. FP8) for dequant.
        if self._is_nvfp4:
            self._generate_nvfp4_tensors()

+        self.dequant_model()
+
        # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
        if self.tensor_map.mapping:
            max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -2992,10 +3038,16 @@ class LlavaVisionModel(MmprojModel):
    def get_token_id(self, token: str) -> int:
        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-            added_tokens_decoder = json.load(f)['added_tokens_decoder']
+            added_tokens_decoder = json.load(f).get('added_tokens_decoder') or {}
            for id_, token_data in added_tokens_decoder.items():
-                if token_data["content"] == token:
+                if token_data.get("content") == token:
                    return int(id_)
+            # fallthrough to tokenizer.json
+        with open(self.dir_model / "tokenizer.json", "r", encoding="utf-8") as f:
+            tokenizer_json = json.load(f)
+            for token_data in tokenizer_json["added_tokens"]:
+                if token_data["content"] == token:
+                    return int(token_data["id"])
        raise ValueError(f"Token '{token}' not found in tokenizer config.")

    def set_gguf_parameters(self):
@@ -3159,40 +3211,6 @@ class Llama4VisionModel(MmprojModel):
                yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register(
-    "Mistral3ForConditionalGeneration",
-    "Ministral3ForCausalLM",
-)
-class Mistral3Model(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.MISTRAL3
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # for compatibility, we use LLAMA arch for older models
-        # TODO: remove this once everyone has migrated to newer version of llama.cpp
-        if self.hparams.get("model_type") != "ministral3":
-            self.model_arch = gguf.MODEL_ARCH.LLAMA
-            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
-            self.gguf_writer.add_architecture()
-            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        rope_params = self.rope_parameters
-        if self.hparams.get("model_type") == "ministral3":
-            assert rope_params, "ministral3 must have 'rope_parameters' config"
-            assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
-            self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
-            self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
-        name = name.replace("language_model.", "")
-        if "multi_modal_projector" in name or "vision_tower" in name:
-            return
-
-        yield from super().modify_tensors(data_torch, name, bid)
-
-
@ModelBase.register("DeciLMForCausalLM")
 class DeciModel(TextModel):
    model_arch = gguf.MODEL_ARCH.DECI
@@ -8232,6 +8250,8 @@ class DeepseekV2Model(TextModel):
    # TODO @ngxson : remove this when we support MTP for deepseek models
    skip_mtp = True

+    merge_expert = True
+
    def set_vocab(self):
        try:
            self._set_vocab_gpt2()
@@ -8370,7 +8390,7 @@ class DeepseekV2Model(TextModel):
                return

        # process the experts separately
-        if name.find("mlp.experts") != -1:
+        if self.merge_expert and name.find("mlp.experts") != -1:
            n_experts = self.hparams["n_routed_experts"]
            assert bid is not None

@@ -8429,6 +8449,69 @@ class DeepseekV2Model(TextModel):
                raise ValueError(f"Unprocessed experts: {experts}")


+@ModelBase.register(
+    "Mistral3ForConditionalGeneration",
+    "Ministral3ForCausalLM",
+)
+class Mistral3Model(TextModel):
+    class Ministral3Model(LlamaModel):
+        model_arch = gguf.MODEL_ARCH.MISTRAL3
+
+        def set_gguf_parameters(self):
+            super().set_gguf_parameters()
+            rope_params = self.rope_parameters
+            if self.hparams.get("model_type") == "ministral3":
+                assert rope_params, "ministral3 must have 'rope_parameters' config"
+                assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
+                self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
+                self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
+
+        def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+            name = name.replace("language_model.", "")
+            if "multi_modal_projector" in name or "vision_tower" in name:
+                return
+
+            yield from super().modify_tensors(data_torch, name, bid)
+
+    class Mistral4Model(DeepseekV2Model):
+        model_arch = gguf.MODEL_ARCH.MISTRAL4
+        skip_mtp = False # model contains no MTP layers, so no need to skip
+        merge_expert = False # experts are already stacked as 3D
+
+        def modify_tensors(self, data_torch, name, bid):
+            if name.endswith(".down_proj") or name.endswith(".gate_up_proj"):
+                name = name + ".weight"
+            yield from super().modify_tensors(data_torch, name, bid)
+
+    model_arch = gguf.MODEL_ARCH.MISTRAL3 # unused
+    impl: TextModel
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.hparams.get("model_type") == "mistral4":
+            self.impl = Mistral3Model.Mistral4Model(*args, **kwargs)
+        else:
+            self.impl = Mistral3Model.Ministral3Model(*args, **kwargs)
+
+    def set_vocab(self):
+        self.impl.set_vocab()
+
+    def set_gguf_parameters(self):
+        self.impl.set_gguf_parameters()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        yield from self.impl.modify_tensors(data_torch, name, bid)
+
+    def prepare_tensors(self):
+        self.impl.prepare_tensors()
+
+    def write_vocab(self):
+        self.impl.write_vocab()
+
+    def write(self):
+        self.impl.write()
+
+
@ModelBase.register("MiniMaxM2ForCausalLM")
 class MiniMaxM2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MINIMAXM2
@@ -128,6 +128,12 @@ class LoraTorchTensor:
        assert dim is None
        return self.shape

+    def contiguous(self) -> LoraTorchTensor:
+        return LoraTorchTensor(
+            self._lora_A.contiguous(),
+            self._lora_B.contiguous(),
+        )
+
    def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
        if isinstance(shape[0], tuple):
            new_shape: tuple[int, ...] = shape[0]
@@ -117,5 +117,5 @@ Legend:
 |                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
@@ -5937,6 +5937,20 @@
 "SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.100000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=0.100000,v=0","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=0.100000,v=1","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=10.000000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=10.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=10.000000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=10.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=10.000000","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[64,5,4,3],eps=10.000000,v=0","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[64,5,4,3],eps=10.000000,v=1","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=10.000000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=10.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=10.000000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=10.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=10.000000","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=10.000000,v=0","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=10.000000,v=1","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[3,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[6,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
@@ -10209,24 +10223,24 @@
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest","support","1","yes","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","1","yes","SYCL"
 "SYCL0","SUM","type=f32,ne=[10,5,4,3]","support","1","yes","SYCL"
 "SYCL0","SUM","type=f32,ne=[11,5,6,3],permute=[0,2,1,3]","support","0","no","SYCL"
 "SYCL0","SUM","type=f32,ne=[11,5,6,3],permute=[0,3,2,1]","support","0","no","SYCL"
@@ -13325,6 +13339,262 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -666,7 +666,7 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo

    float sumf = 0;

-#if defined __ARM_NEON
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
    const int8x16_t values = vld1q_s8(kvalues_mxfp4);
    const uint8x16_t m4b = vdupq_n_u8(0x0f);
    float32x4_t acc = vdupq_n_f32(0.0f);
@@ -1473,10 +1473,12 @@ 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 (op->src[0]->type != GGML_TYPE_F16) {
+                    return nullptr;
+                }
                std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
                const int slot_total = kleidiai_collect_kernel_chain(op, kernel_chain);
-                const bool has_kernel = slot_total > 0;
-                if (has_kernel && op->src[1]->ne[1] > 1) {
+                if (slot_total > 0 && op->src[1]->ne[1] > 1) {
                    if ((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;
@@ -6205,7 +6205,7 @@ static void ggml_compute_forward_im2col_f16(
    const ggml_tensor * src1 = dst->src[1];

    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
    GGML_ASSERT( dst->type == GGML_TYPE_F16);

    GGML_TENSOR_BINARY_OP_LOCALS;
@@ -6236,7 +6236,7 @@ static void ggml_compute_forward_im2col_f16(
    int ofs1 = is_2D ? nb12 : nb11;

    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));

    // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
    {
@@ -6249,7 +6249,12 @@ static void ggml_compute_forward_im2col_f16(

                        // micro kernel
                        ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
-                        const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW]
+                        const float * const src_data_f32 = src1->type == GGML_TYPE_F32
+                            ? (const float *)((const char *) src1->data + in*ofs0 + iic*ofs1)
+                            : nullptr; // [IH, IW]
+                        const ggml_fp16_t * const src_data_f16 = src1->type == GGML_TYPE_F16
+                            ? (const ggml_fp16_t *)((const char *) src1->data + in*ofs0 + iic*ofs1)
+                            : nullptr; // [IH, IW]

                        for (int64_t ikh = 0; ikh < KH; ikh++) {  // 1
                            for (int64_t ikw = 0; ikw < KW; ikw++) {
@@ -6259,7 +6264,11 @@ static void ggml_compute_forward_im2col_f16(
                                if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
                                } else {
-                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]);
+                                    if (src_data_f32 != nullptr) {
+                                        dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data_f32[iih*IW + iiw]);
+                                    } else {
+                                        dst_data[iic*(KH*KW) + ikh*KW + ikw] = src_data_f16[iih*IW + iiw];
+                                    }
                                }
                            }
                        }
@@ -892,7 +892,7 @@ void launch_fattn(
    const int ntiles_x     = ((Q->ne[1] + ncols1 - 1) / ncols1);
    const int gqa_ratio    = Q->ne[2] / K->ne[2];
    const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2);
-    const int ntiles_total = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];
+    const int ntiles_dst   = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];

    // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
    // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
@@ -919,37 +919,37 @@ void launch_fattn(
    GGML_ASSERT(max_blocks_per_sm > 0);
    int parallel_blocks = max_blocks_per_sm;

+    const int ntiles_KV = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by KV cache length.
+
    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 = 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);
+        const int tiles_nwaves = (ntiles_dst + max_blocks - 1) / max_blocks;
+        const int tiles_efficiency_percent = 100 * ntiles_dst / (max_blocks*tiles_nwaves);

-        const int nblocks_stream_k = max_blocks;
+        const int nblocks_stream_k = std::min(max_blocks, ntiles_KV*ntiles_dst);

        const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || amd_wmma_available(cc) || tiles_efficiency_percent < 75;

-        blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total;
+        blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_dst;
        blocks_num.y = 1;
        blocks_num.z = 1;

-        if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
+        if (ntiles_dst % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
            dst_tmp_meta.alloc((size_t(blocks_num.x) * ncols * (2 + DV/2)));
        }
    } else {
-        const int ntiles_KQ = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by tensor size.
-
        // parallel_blocks must not be larger than what the tensor size allows:
-        parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
+        parallel_blocks = std::min(parallel_blocks, ntiles_KV);

        // If ntiles_total % blocks_per_wave != 0 then some efficiency is lost due to tail effects.
        // Test whether parallel_blocks can be set to a higher value for better efficiency.
        const int blocks_per_wave = nsm * max_blocks_per_sm;
        int nwaves_best = 0;
        int efficiency_percent_best = 0;
-        for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KQ; ++parallel_blocks_test) {
-            const int nblocks_total = ntiles_total * parallel_blocks_test;
+        for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KV; ++parallel_blocks_test) {
+            const int nblocks_total = ntiles_dst * parallel_blocks_test;
            const int nwaves = (nblocks_total + blocks_per_wave - 1) / blocks_per_wave;
            const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave);

@@ -1015,7 +1015,7 @@ void launch_fattn(
    CUDA_CHECK(cudaGetLastError());

    if (stream_k) {
-        if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
+        if (ntiles_dst % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
            const dim3 block_dim_combine(DV, 1, 1);
            const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};

@@ -1,7 +1,8 @@
 #include "gated_delta_net.cuh"

 template <int S_v, bool KDA>
-__global__ void gated_delta_net_cuda(const float * q,
+__global__ void __launch_bounds__((ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v) * 4, 2)
+gated_delta_net_cuda(const float * q,
                                     const float * k,
                                     const float * v,
                                     const float * g,
@@ -38,7 +39,7 @@ __global__ void gated_delta_net_cuda(const float * q,

    const int64_t state_offset = (sequence * H + h_idx) * S_v * S_v;
    state += state_offset;
-    curr_state += state_offset;
+    curr_state += state_offset + col * S_v;
    attn_data += (sequence * n_tokens * H + h_idx) * S_v;

    constexpr int warp_size = ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v;
@@ -46,10 +47,11 @@ __global__ void gated_delta_net_cuda(const float * q,
    constexpr int rows_per_lane = (S_v + warp_size - 1) / warp_size;
    float         s_shard[rows_per_lane];
    // state is stored transposed: M[col][i] = S[i][col], row col is contiguous
+
 #pragma unroll
    for (int r = 0; r < rows_per_lane; r++) {
        const int i = r * warp_size + lane;
-        s_shard[r]  = curr_state[col * S_v + i];
+        s_shard[r]  = curr_state[i];
    }

    for (int t = 0; t < n_tokens; t++) {
@@ -63,6 +65,16 @@ __global__ void gated_delta_net_cuda(const float * q,

        const float beta_val = *beta_t;

+        // Cache k and q in registers
+        float k_reg[rows_per_lane];
+        float q_reg[rows_per_lane];
+#pragma unroll
+        for (int r = 0; r < rows_per_lane; r++) {
+            const int i = r * warp_size + lane;
+            k_reg[r] = k_t[i];
+            q_reg[r] = q_t[i];
+        }
+
        if constexpr (!KDA) {
            const float g_val = expf(*g_t);

@@ -70,8 +82,7 @@ __global__ void gated_delta_net_cuda(const float * q,
            float kv_shard = 0.0f;
 #pragma unroll
            for (int r = 0; r < rows_per_lane; r++) {
-                const int i = r * warp_size + lane;
-                kv_shard += s_shard[r] * k_t[i];
+                kv_shard += s_shard[r] * k_reg[r];
            }
            float kv_col = warp_reduce_sum<warp_size>(kv_shard);

@@ -83,9 +94,8 @@ __global__ void gated_delta_net_cuda(const float * q,
            float attn_partial = 0.0f;
 #pragma unroll
            for (int r = 0; r < rows_per_lane; r++) {
-                const int i = r * warp_size + lane;
-                s_shard[r]  = g_val * s_shard[r] + k_t[i] * delta_col;
-                attn_partial += s_shard[r] * q_t[i];
+                s_shard[r]  = g_val * s_shard[r] + k_reg[r] * delta_col;
+                attn_partial += s_shard[r] * q_reg[r];
            }

            float attn_col = warp_reduce_sum<warp_size>(attn_partial);
@@ -99,7 +109,7 @@ __global__ void gated_delta_net_cuda(const float * q,
 #pragma unroll
            for (int r = 0; r < rows_per_lane; r++) {
                const int i = r * warp_size + lane;
-                kv_shard += expf(g_t[i]) * s_shard[r] * k_t[i];
+                kv_shard += expf(g_t[i]) * s_shard[r] * k_reg[r];
            }

            float kv_col = warp_reduce_sum<warp_size>(kv_shard);
@@ -113,8 +123,8 @@ __global__ void gated_delta_net_cuda(const float * q,
 #pragma unroll
            for (int r = 0; r < rows_per_lane; r++) {
                const int i = r * warp_size + lane;
-                s_shard[r]  = expf(g_t[i]) * s_shard[r] + k_t[i] * delta_col;
-                attn_partial += s_shard[r] * q_t[i];
+                s_shard[r]  = expf(g_t[i]) * s_shard[r] + k_reg[r] * delta_col;
+                attn_partial += s_shard[r] * q_reg[r];
            }

            float attn_col = warp_reduce_sum<warp_size>(attn_partial);
@@ -124,7 +124,10 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
        err = cudaMallocManaged(ptr, size);
 #if defined(GGML_USE_HIP)
        if (err == hipSuccess) {
-            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+            // hipMemAdviseSetCoarseGrain is an optional performance hint;
+            // ignore errors (e.g. hipErrorInvalidValue on some APU/iGPU configs).
+            cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device);
+            (void)hipGetLastError(); // clear any error
        }

        // fall back to cudaMalloc if not supported (e.g. on Windows)
@@ -251,11 +254,6 @@ static ggml_cuda_device_info ggml_cuda_init() {
        info.devices[id].supports_cooperative_launch = false;
 #endif // !(GGML_USE_MUSA)

-        // cudaMemGetInfo returns info for the current device
-        size_t free_mem;
-        CUDA_CHECK(cudaSetDevice(id));
-        CUDA_CHECK(cudaMemGetInfo(&free_mem, NULL));
-
 #if defined(GGML_USE_HIP)
        info.devices[id].smpbo = prop.sharedMemPerBlock;

@@ -270,25 +268,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
                info.devices[id].cc += prop.minor * 0x10;
            }
        }
-        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d, VRAM: %zu MiB (%zu MiB free)\n",
+        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d, VRAM: %zu MiB\n",
                      id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
                      device_vmm ? "yes" : "no", prop.warpSize,
-                      (size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)));
 #elif defined(GGML_USE_MUSA)
        // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
        info.devices[id].warp_size = 32;
        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
        info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100;
        info.devices[id].cc += prop.minor * 0x10;
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB (%zu MiB free)\n",
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB\n",
                      id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
-                      (size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)));
 #else
        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
        info.devices[id].cc = 100*prop.major + 10*prop.minor;
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB (%zu MiB free)\n",
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB\n",
                      id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
-                      (size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)));
        std::string device_name(prop.name);
        if (device_name == "NVIDIA GeForce MX450") {
            turing_devices_without_mma.push_back({ id, device_name });
@@ -303,6 +301,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
        // TODO: Check for future drivers the default scheduling strategy and
        // remove this call again when cudaDeviceScheduleSpin is default.
        if (prop.major == 12 && prop.minor == 1) {
+            CUDA_CHECK(cudaSetDevice(id));
            CUDA_CHECK(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
        }

@@ -60,11 +60,17 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
 enum mmvq_parameter_table_id {
    MMVQ_PARAMETERS_GENERIC = 0,
    MMVQ_PARAMETERS_GCN,
-    MMVQ_PARAMETERS_RDNA2
+    MMVQ_PARAMETERS_RDNA2,
+    MMVQ_PARAMETERS_RDNA3_0,
+    MMVQ_PARAMETERS_RDNA4
 };

 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
-#if defined(RDNA2) || defined(RDNA3) || defined(RDNA4)
+#if defined(RDNA4)
+    return MMVQ_PARAMETERS_RDNA4;
+#elif defined(RDNA3_0)
+    return MMVQ_PARAMETERS_RDNA3_0;
+#elif defined(RDNA2) || defined(RDNA3_5)
    return MMVQ_PARAMETERS_RDNA2;
 #elif defined(GCN) || defined(CDNA)
    return MMVQ_PARAMETERS_GCN;
@@ -74,7 +80,13 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
 }

 static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
-    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+    if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+        return MMVQ_PARAMETERS_RDNA4;
+    }
+    if (GGML_CUDA_CC_IS_RDNA3_0(cc)) {
+        return MMVQ_PARAMETERS_RDNA3_0;
+    }
+    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc)) {
        return MMVQ_PARAMETERS_RDNA2;
    }
    if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
@@ -83,7 +95,7 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
    return MMVQ_PARAMETERS_GENERIC;
 }

-static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
    if (table_id == MMVQ_PARAMETERS_GENERIC) {
        switch (ncols_dst) {
            case 1:
@@ -114,6 +126,50 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_paramet
                return 1;
        }
    }
+    if (table_id == MMVQ_PARAMETERS_RDNA4) {
+        // nwarps=8 benefits types with simple vec_dot on RDNA4 (ncols_dst=1).
+        // Types with complex vec_dot (Q3_K, IQ2_*, IQ3_*) regress due to register
+        // pressure and lookup table contention at higher thread counts.
+        if (ncols_dst == 1) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                case GGML_TYPE_IQ4_XS:
+                    return 8;
+                default:
+                    return 1;
+            }
+        }
+        return 1;
+    }
+    if (table_id == MMVQ_PARAMETERS_RDNA3_0) {
+        // RDNA3 (W7900): stricter whitelist than RDNA4.
+        // Q2_K / Q5_K / IQ4_XS regress in full quant sweeps.
+        if (ncols_dst == 1) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                    return 8;
+                default:
+                    return 1;
+            }
+        }
+        return 1;
+    }
    return 1;
 }

@@ -138,7 +194,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
 }

 template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false>
-__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -151,7 +207,7 @@ static __global__ void mul_mat_vec_q(
    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
    constexpr int vdr = get_vdr_mmvq(type);
    constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
+    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();

@@ -355,12 +411,13 @@ static __global__ void mul_mat_vec_q(
    }
 }

+template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
        const int warp_size, const mmvq_parameter_table_id table_id) {
    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
-    const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
+    const dim3 block_dims(warp_size, calc_nwarps(type, ncols_dst, table_id), 1);
    return {block_nums, block_dims};
 }

@@ -420,7 +477,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
    if (has_ids && ncols_dst > 1) {
        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
        constexpr int c_ncols_dst = 1;
-        std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
+        std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -431,7 +488,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
    switch (ncols_dst) {
        case 1: {
            constexpr int c_ncols_dst = 1;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -439,7 +496,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 2: {
            constexpr int c_ncols_dst = 2;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -447,7 +504,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 3: {
            constexpr int c_ncols_dst = 3;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -455,7 +512,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 4: {
            constexpr int c_ncols_dst = 4;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -463,7 +520,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 5: {
            constexpr int c_ncols_dst = 5;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -471,7 +528,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 6: {
            constexpr int c_ncols_dst = 6;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -479,7 +536,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 7: {
            constexpr int c_ncols_dst = 7;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -487,7 +544,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        } break;
        case 8: {
            constexpr int c_ncols_dst = 8;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -207,6 +207,14 @@
 #define RDNA3
 #endif // defined(__GFX11__)

+#if defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3_5
+#endif // defined(__gfx1150__) || defined(__gfx1151__)
+
+#if defined(RDNA3) && !defined(RDNA3_5)
+#define RDNA3_0
+#endif // defined(RDNA3) && !defined(RDNA3_5)
+
 #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
 #define RDNA2
@@ -1142,6 +1142,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                op->src[0]->ne[0] != 128 &&
                op->src[0]->ne[0] != 192 &&
                op->src[0]->ne[0] != 256 &&
+                op->src[0]->ne[0] != 320 &&
                op->src[0]->ne[0] != 576) {
                return false;
            }
@@ -6176,6 +6176,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk128_dv128")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f32_dk192_dv192")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  192, 192>;
 template [[host_name("kernel_flash_attn_ext_f32_dk192_dv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f32_dk256_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  256, 256>;
+template [[host_name("kernel_flash_attn_ext_f32_dk320_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  320, 256>;
 template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  576, 512>;

 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
@@ -6190,6 +6191,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk128_dv128")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_dk192_dv192")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 192>;
 template [[host_name("kernel_flash_attn_ext_f16_dk192_dv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f16_dk256_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256, 256>;
+template [[host_name("kernel_flash_attn_ext_f16_dk320_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  320, 256>;
 template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;

 #if defined(GGML_METAL_HAS_BF16)
@@ -6205,6 +6207,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk128_dv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 320, 256>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif

@@ -6220,6 +6223,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk128_dv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 320, 256>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;

 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
@@ -6234,6 +6238,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk128_dv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 320, 256>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;

 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
@@ -6248,6 +6253,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk128_dv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 320, 256>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;

 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
@@ -6262,6 +6268,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk128_dv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 320, 256>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;

 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
@@ -6276,6 +6283,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_dk128_dv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 320, 256>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;

 #undef FA_TYPES
@@ -6846,6 +6854,17 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_dk256_dv256")]] kernel flas
 template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk256_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 256, 256, 1>;
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk256_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 256, 256, 1>;

+template [[host_name("kernel_flash_attn_ext_vec_f32_dk320_dv256")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES_F32, float4,     1, dequantize_f32_t4,  float4,      1, dequantize_f32_t4,  320, 256, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_f16_dk320_dv256")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  320, 256, 2>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     bfloat4,    1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 320, 256, 2>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q4_0, 8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 320, 256, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q4_1, 8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 320, 256, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q5_0, 8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 320, 256, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 320, 256, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 320, 256, 2>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f32_dk576_dv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES_F32, float4,     1, dequantize_f32_t4,  float4,      1, dequantize_f32_t4,  576, 512, 2>;
 template [[host_name("kernel_flash_attn_ext_vec_f16_dk576_dv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  576, 512, 2>;
 #if defined(GGML_METAL_HAS_BF16)
@@ -4767,7 +4767,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
            sumqx += w*q*xb[j];
            sumq2 += w*q*q;
        }
-        d = sumqx/sumq2;
+        d = sumq2 > 0 ? sumqx/sumq2 : 0.f;
        float best = d*sumqx;
        for (int itry = -ntry; itry <= ntry; ++itry) {
            id = (itry + values[0])/max;
@@ -24,6 +24,7 @@
 #include "dmmv.hpp"
 #include "element_wise.hpp"
 #include "fattn.hpp"
+#include "gated_delta_net.hpp"
 #include "gla.hpp"
 #include "im2col.hpp"
 #include "mmq.hpp"
@@ -31,6 +32,7 @@
 #include "norm.hpp"
 #include "outprod.hpp"
 #include "pad.hpp"
+#include "pad_reflect_1d.hpp"
 #include "quantize.hpp"
 #include "quants.hpp"
 #include "roll.hpp"
@@ -39,8 +41,8 @@
 #include "ssm_conv.hpp"
 #include "softmax.hpp"
 #include "tsembd.hpp"
+#include "upscale.hpp"
 #include "wkv.hpp"
-#include "pad_reflect_1d.hpp"


 #endif  // GGML_SYCL_BACKEND_HPP
@@ -216,7 +216,8 @@ struct sycl_device_info {
                       // cudaOccupancyMaxActiveBlocksPerMultiprocessor
    bool    vmm;                // virtual memory support
    size_t  total_vram;
-    //sycl_hw_info hw_info;     \\ device id and aarch, currently not used
+    int max_work_group_sizes;
+    sycl_hw_info hw_info;
    optimize_feature opt_feature;
 };

@@ -692,15 +693,20 @@ static __dpct_inline__ float get_alibi_slope(const float    max_bias,
    return dpct::pow(base, exph);
 }

-static const sycl::uint3 init_fastdiv_values(uint32_t d) {
-    GGML_ASSERT(d != 0);
+static const sycl::uint3 init_fastdiv_values(uint64_t d_64) {
+    GGML_ASSERT(d_64 != 0);
+    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());

+    uint32_t d = (uint32_t)d_64;
+
+    // compute L = ceil(log2(d));
    uint32_t L = 0;
    while (L < 32 && (uint32_t{ 1 } << L) < d) {
        L++;
    }

    uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);
+    // pack divisor as well to reduce error surface
    return sycl::uint3(mp, L, d);
 }

@@ -965,4 +971,23 @@ static T block_reduce(T val, T * shared_vals, int block_size_template) {
    return val;
 }

+/*
+ * for Debug
+ */
+static int get_thread_id() {
+  auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+  int blockId = item_ct1.get_group(2) +
+      item_ct1.get_group(1) * item_ct1.get_group_range(2) +
+      item_ct1.get_group(0) * item_ct1.get_group_range(2) *
+          item_ct1.get_group_range(1);
+  int threadsPerBlock = item_ct1.get_local_range(2) *
+      item_ct1.get_local_range(1) * item_ct1.get_local_range(0);
+  int threadInBlockId = item_ct1.get_local_id(2) +
+      item_ct1.get_local_id(1) * item_ct1.get_local_range(2) +
+      item_ct1.get_local_id(0) * item_ct1.get_local_range(2) *
+          item_ct1.get_local_range(1);
+  int id = blockId * threadsPerBlock + threadInBlockId;
+  return id;
+}
+
 #endif // GGML_SYCL_COMMON_HPP
@@ -294,30 +294,6 @@ static void unary_op_trunc_kernel(const T * x, T * dst, const int k, const sycl:
    }
 }

-template<typename  T>
-static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
-                        const int nb02, const int nb03, const int ne10, const int ne11,
-                        const int ne12, const int ne13, const float sf0, const float sf1,
-                        const float sf2, const float sf3, const sycl::nd_item<1> &item_ct1) {
-    int index = item_ct1.get_local_id(0) +
-               item_ct1.get_group(0) * item_ct1.get_local_range(0);
-    if (index >= ne10 * ne11 * ne12 * ne13) {
-        return;
-    }
-    // operation
-    int i10 = index % ne10;
-    int i11 = (index / ne10) % ne11;
-    int i12 = (index / (ne10 * ne11)) % ne12;
-    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
-
-    int i00 = static_cast<int>(i10 / sf0);
-    int i01 = static_cast<int>(i11 / sf1);
-    int i02 = static_cast<int>(i12 / sf2);
-    int i03 = static_cast<int>(i13 / sf3);
-
-    dst[index] = *(const T *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
-}
-
 template<typename T>
 static void clamp(const T * x, T * dst, const float min, const float max, const int k,
                      const sycl::nd_item<1> &item_ct1) {
@@ -392,20 +368,6 @@ static void arange_kernel(T * dst, const int k, T start, T step,
    }
 }

-template<typename T>
-static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
-                             const int nb02, const int nb03, const int ne10, const int ne11,
-                             const int ne12, const int ne13, const float sf0, const float sf1,
-                             const float sf2, const float sf3, queue_ptr stream) {
-    int dst_size = ne10 * ne11 * ne12 * ne13;
-    int num_blocks = ceil_div(dst_size, SYCL_UPSCALE_BLOCK_SIZE);
-    sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
-            upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
-        });
-}
-
 template<typename KernelInvoker, typename... Args>
 static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
@@ -505,42 +467,6 @@ static inline void dispatch_ggml_sycl_op_fused_glu(ggml_backend_sycl_context & c
    }
 }

-template<typename KernelInvoker, typename... Args>
-static inline void dispatch_ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-
-    const float sf0 = (float) dst->ne[0] / dst->src[0]->ne[0];
-    const float sf1 = (float) dst->ne[1] / dst->src[0]->ne[1];
-    const float sf2 = (float) dst->ne[2] / dst->src[0]->ne[2];
-    const float sf3 = (float) dst->ne[3] / dst->src[0]->ne[3];
-    switch (dst->type) {
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->nb[0], (int)dst->src[0]->nb[1], (int)dst->src[0]->nb[2],
-                               (int)dst->src[0]->nb[3], (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], sf0, sf1, sf2, sf3,
-                               main_stream, std::forward<Args>(args)...);
-                break;
-            }
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->nb[0], (int)dst->src[0]->nb[1], (int)dst->src[0]->nb[2],
-                               (int)dst->src[0]->nb[3], (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], sf0, sf1, sf2, sf3,
-                               main_stream, std::forward<Args>(args)...);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
 template<typename F>
 static inline void ggml_sycl_op_unary(
        ggml_backend_sycl_context & ctx, ggml_tensor * dst, F func) {
@@ -784,15 +710,6 @@ static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor
        });
 }

-static inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_upscale(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int nb00, int nb01, int nb02, int nb03,
-           int ne10, int ne11, int ne12, int ne13, float sf0, float sf1, float sf2, float sf3,
-           queue_ptr stream) {
-            ggml_sycl_detail::upscale_sycl(src, dst_ptr, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, stream);
-        });
-}
-
 static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    float min_val;
    float max_val;
@@ -1131,12 +1048,6 @@ void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    ggml_sycl_op_sqr(ctx, dst);
 }

-void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
-    ggml_sycl_op_upscale(ctx, dst);
-}
-
-
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
    ggml_sycl_op_clamp(ctx, dst);
@@ -71,8 +71,6 @@ void ggml_sycl_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

 void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

-void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

 void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
@@ -70,6 +70,7 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp16(const int DKQ, co
    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2,  64,  64)

    return 0;
 }
@@ -130,134 +131,6 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp32(const int DKQ, co
    return 0;
 }

-static constexpr uint32_t ggml_sycl_fattn_tile_get_config_amd(const int DKQ, const int DV, const int ncols) {
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  2,  64, 3,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  4, 128, 3,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  8, 128, 2,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  2,  64, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  4, 128, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  8, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 16, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 32, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 64, 256, 2,  32,  48)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  2,  64, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  4, 128, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  8, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2,  32,  56)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  2, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  4, 128, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  8, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 2,  64,  32)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  2, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  4, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  8, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
-
-    return 0;
-}
-
-static constexpr uint32_t ggml_sycl_fattn_tile_get_config_amd_rdna(const int DKQ, const int DV, const int ncols) {
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  4,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  8, 128, 5, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 128, 5, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 128, 4,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 128, 5,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  2,  64, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  4, 128, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  8, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 16, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 32, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 64, 256, 2,  32,  48)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  2,  64, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  4, 128, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  8, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2,  32,  56)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  4, 128, 8,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  8, 128, 8,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 3, 128, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 3, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 3,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  4, 128, 6,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  8, 128, 6,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
-
-    return 0;
-}
-
 static constexpr uint32_t ggml_sycl_fattn_tile_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
    if(fast_fp16_available(cc))
        return ggml_sycl_fattn_tile_get_config_fp16(DKQ, DV, ncols);
@@ -455,7 +328,7 @@ static __dpct_inline__ void flash_attn_tile_iter_KQ(T_vec_dot * const Q_tmp,

    flash_attn_tile_load_tile<warp_size, nwarps, nbatch_fa, nbatch_K, cpy_ne, oob_check>
        (K_h2 + int64_t(k_VKQ_0)*stride_K2 + k_KQ_0/2, KV_tmp, stride_K2, k_VKQ_sup);
-    item_ct1.barrier();
+    item_ct1.barrier(sycl::access::fence_space::local_space);

 #ifdef SYCL_FAST_FP16
    static_assert((nbatch_K/2) % cpy_ne == 0, "bad nbatch_K");
@@ -505,7 +378,7 @@ static __dpct_inline__ void flash_attn_tile_iter_KQ(T_vec_dot * const Q_tmp,
    }

    if (k_KQ_0 + nbatch_K < DKQ) {
-        item_ct1.barrier();  // Sync not needed on last iteration.
+        item_ct1.barrier(sycl::access::fence_space::local_space);  // Sync not needed on last iteration.
    }
 }

@@ -545,7 +418,7 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
                                                 const int         k_VKQ_max,
                                                 const int         col_Q_0,
                                                 float *           KQ_max_new_shared) {
-    auto          item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
    constexpr int cpy_nb   = ggml_sycl_get_max_cpy_bytes();
    constexpr int cpy_ne = cpy_nb / 4;

@@ -620,14 +493,14 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
    }

    if constexpr (np == 1) {
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
    } else {
        static_assert(cpw == 1, "bad cpw");

        if (item_ct1.get_local_id(2) == 0) {
            KQ_max_new_shared[item_ct1.get_local_id(1)] = KQ_max_new[0];
        }
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
        KQ_max_new[0] = KQ_max_new_shared[(item_ct1.get_local_id(1) & ~(np - 1)) + item_ct1.get_local_id(2) % np];
        KQ_max_new[0] = warp_reduce_max<np>(KQ_max_new[0]);
    }
@@ -697,7 +570,7 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
    for (int k0 = 0; k0 < nbatch_fa; k0 += nbatch_V) {
        flash_attn_tile_load_tile<warp_size, nwarps, nbatch_V, DV, 0, oob_check>
            (V_h2 + int64_t(k_VKQ_0 + k0)*stride_V2, KV_tmp, stride_V2, k_VKQ_sup - k0);
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);

 #ifdef SYCL_FAST_FP16
 #pragma unroll
@@ -765,7 +638,7 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
            }
        }
 #endif // SYCL_FAST_FP16
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
    }
 }

@@ -825,6 +698,9 @@ static void flash_attn_tile(const char *  Q,
                  nb31, nb32, nb33);
        return;
    }
+    // int id =get_thread_id();
+    // if(id==0) sycl::ext::oneapi::experimental::printf("DKQ=%d, DV=%d, ncols1=%d, ncols2=%d\n",
+    //     DKQ, DV, ncols1, ncols2);

    static_assert(ggml_sycl_fattn_tile_get_config(DKQ, DV, ncols1*ncols2) != 0, "kernel config not defined");

@@ -932,7 +808,9 @@ static void flash_attn_tile(const char *  Q,

        constexpr int cpy_ne_D = cpy_ne < DKQp/warp_size ? cpy_ne : DKQp/warp_size;

+
 #pragma unroll
+
        for (int i0 = 0; i0 < DKQp; i0 += np*warp_size*cpy_ne_D) {
            if (i0 + np * warp_size * cpy_ne_D <= DKQ ||
                i0 + (item_ct1.get_local_id(1) % np) * (warp_size * cpy_ne_D) + item_ct1.get_local_id(2) * cpy_ne_D <
@@ -972,7 +850,7 @@ static void flash_attn_tile(const char *  Q,
        }
    }

-    item_ct1.barrier();
+    item_ct1.barrier(sycl::access::fence_space::local_space);

    // Main loop over KV cache:
    const int k_VKQ_max = KV_max ? KV_max[sequence * item_ct1.get_group_range(2) + item_ct1.get_group(2)] : ne11;
@@ -1051,7 +929,7 @@ static void flash_attn_tile(const char *  Q,
            return;
        }

-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);

 #pragma unroll
        for (int ip = 1; ip < np; ++ip) {
@@ -1189,30 +1067,58 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_sycl_context & ctx, ggm

    const int id        = ggml_sycl_get_device();
    const int cc        = ggml_sycl_info().devices[id].cc;
+    const auto arch     = ggml_sycl_info().devices[id].hw_info.arch;
+    const auto gpu_name     = ggml_sycl_info().devices[id].hw_info.name;
    const int warp_size = WARP_32_SIZE; //can't support WARP_16_SIZE

    constexpr size_t nbytes_shared = 0;

-    if constexpr (DV <= 256) {
-        if (Q->ne[1] > 16/ncols2) {
-            constexpr int cols_per_block = 32;
-            const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
-            const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
-            launch_fattn<DV, cols_per_block/ncols2, ncols2,
-                flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
-                (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
-            return;
+    // printf("zjy DKQ=%d, DV=%d, ncols2=%d, use_logit_softca=%d Q->ne[1]=%d 4/ncols2=%d\n",
+    //     DKQ, DV, ncols2, use_logit_softcap, Q->ne[1], 4/ncols2);
+    // printf("zjy name=%s\n", gpu_name.data());
+    // std::cout << "GPU name "<<gpu_name<< " arch "<< ggml_sycl_info().devices[id].hw_info.arch_name
+    //    <<std::endl;
+
+    //same or older iGPU are weak too.
+    if (arch <= gpu_arch::intel_gpu_adl_s) {
+        if (Q->ne[1]>=75) {
+            {
+                constexpr int cols_per_block = ncols2*2;
+                const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+                const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+                launch_fattn<DV, cols_per_block/ncols2, ncols2,
+                    flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+                    (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+                return;
+            }
        }
    }

-    if (Q->ne[1] > 8/ncols2) {
-        constexpr int cols_per_block = 16;
-        const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
-        const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
-        launch_fattn<DV, cols_per_block/ncols2, ncols2,
-            flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
-            (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
-        return;
+    if (arch > gpu_arch::intel_gpu_acm_g10 || Q->ne[1]<75) {
+        if constexpr (ncols2 <= 32) {
+            if (Q->ne[1] > 16/ncols2) {
+                constexpr int cols_per_block = 32;
+                const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+                const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+                launch_fattn<DV, cols_per_block/ncols2, ncols2,
+                    flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+                    (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+                return;
+            }
+        }
+
+
+        if constexpr (ncols2 <= 16) {
+            if (Q->ne[1] > 8/ncols2) {
+                constexpr int cols_per_block = 16;
+                const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+                const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+                launch_fattn<DV, cols_per_block/ncols2, ncols2,
+                    flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+                    (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+                return;
+            }
+        }
    }

    if constexpr (ncols2 <= 8) {
@@ -1237,6 +1143,8 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_sycl_context & ctx, ggm
                (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
            return;
        }
+
+
    }

    if constexpr (ncols2 <= 2) {
@@ -55,7 +55,7 @@ void gated_delta_net_sycl(const float *     q,
 #pragma unroll
    for (int r = 0; r < rows_per_lane; r++) {
        const int i = r * warp_size + lane;
-        s_shard[r]  = curr_state[i * S_v + col];
+        s_shard[r]  = curr_state[col * S_v + i];
    }

    for (int t = 0; t < n_tokens; t++) {
@@ -137,7 +137,7 @@ void gated_delta_net_sycl(const float *     q,
 #pragma unroll
    for (int r = 0; r < rows_per_lane; r++) {
        const int i          = r * warp_size + lane;
-        state[i * S_v + col] = s_shard[r];
+        state[col * S_v + i] = s_shard[r];
    }
 }

@@ -44,7 +44,6 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
-#include "ggml-sycl/gated_delta_net.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml-sycl/norm.hpp"
@@ -105,6 +104,7 @@ static ggml_sycl_device_info ggml_sycl_init() {

        info.max_work_group_sizes[i] = prop.get_max_work_group_size();
        info.devices[i].max_wg_per_cu = info.max_work_group_sizes[i] / prop.get_max_compute_units();
+        info.devices[i].hw_info = get_device_hw_info(&device);

    }

@@ -4863,9 +4863,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_ROPE:
        case GGML_OP_ROPE_BACK:
        case GGML_OP_IM2COL:
-            return true;
        case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS);
+            return true;
        case GGML_OP_SUM:
        case GGML_OP_SUM_ROWS:
        case GGML_OP_MEAN:
@@ -1,15 +1,67 @@
 #include "sycl_hw.hpp"

-// TODO: currently not used
-/*
-sycl_hw_info get_device_hw_info(sycl::device *device_ptr) {
-  sycl_hw_info res;
-  int32_t id = device_ptr->get_info<sycl::ext::intel::info::device::device_id>();
-  res.device_id = id;
+using namespace std;

-  syclex::architecture arch = device_ptr->get_info<syclex::info::device::architecture>();
-  res.arch = arch;
-
-  return res;
-}
+/*defined in
+* /opt/intel/oneapi/compiler/latest/include/sycl/ext/oneapi/experimental/device_architecture.def
 */
+static map<gpu_arch, std::pair<const char*, sycl_intel_gpu_family>> arch2name = {
+    {gpu_arch::intel_gpu_bdw,     {"intel_gpu_bdw",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_skl,     {"intel_gpu_skl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_kbl,     {"intel_gpu_kbl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_cfl,     {"intel_gpu_cfl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_apl,     {"intel_gpu_apl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_glk,     {"intel_gpu_glk",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_whl,     {"intel_gpu_whl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_aml,     {"intel_gpu_aml",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_cml,     {"intel_gpu_cml",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_icllp,   {"intel_gpu_icllp",   GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_ehl,     {"intel_gpu_ehl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_tgllp,   {"intel_gpu_tgllp",   GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_rkl,     {"intel_gpu_rkl",     GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_adl_s,   {"intel_gpu_adl_s",   GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_adl_p,   {"intel_gpu_adl_p",   GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_adl_n,   {"intel_gpu_adl_n",   GPU_FAMILY_IGPU_NON_XE}},
+    {gpu_arch::intel_gpu_dg1,     {"intel_gpu_dg1",     GPU_FAMILY_DGPU_CLIENT_GAME}},
+    {gpu_arch::intel_gpu_acm_g10, {"intel_gpu_acm_g10", GPU_FAMILY_DGPU_CLIENT_GAME}},
+    {gpu_arch::intel_gpu_acm_g11, {"intel_gpu_acm_g11", GPU_FAMILY_DGPU_CLIENT_GAME}},
+    {gpu_arch::intel_gpu_acm_g12, {"intel_gpu_acm_g12", GPU_FAMILY_DGPU_CLIENT_GAME}},
+    {gpu_arch::intel_gpu_pvc,     {"intel_gpu_pvc",     GPU_FAMILY_DGPU_CLOUD}},
+    {gpu_arch::intel_gpu_pvc_vg,  {"intel_gpu_pvc_vg",  GPU_FAMILY_DGPU_CLOUD}},
+    {gpu_arch::intel_gpu_mtl_u,   {"intel_gpu_mtl_u",   GPU_FAMILY_IGPU_XE}},
+    {gpu_arch::intel_gpu_mtl_h,   {"intel_gpu_mtl_h",   GPU_FAMILY_IGPU_XE}},
+    {gpu_arch::intel_gpu_arl_h,   {"intel_gpu_arl_h",   GPU_FAMILY_IGPU_XE}},
+    {gpu_arch::intel_gpu_bmg_g21, {"intel_gpu_bmg_g21", GPU_FAMILY_DGPU_CLIENT_GAME}},
+    {gpu_arch::intel_gpu_bmg_g31, {"intel_gpu_bmg_g31", GPU_FAMILY_DGPU_CLIENT_GAME}},
+    {gpu_arch::intel_gpu_lnl_m,   {"intel_gpu_lnl_m",   GPU_FAMILY_IGPU_XE}},
+    {gpu_arch::intel_gpu_ptl_h,   {"intel_gpu_ptl_h",   GPU_FAMILY_IGPU_XE}},
+    {gpu_arch::intel_gpu_ptl_u,   {"intel_gpu_ptl_u",   GPU_FAMILY_IGPU_XE}},
+    {gpu_arch::intel_gpu_wcl,     {"intel_gpu_wcl",     GPU_FAMILY_IGPU_XE}}
+};
+
+
+sycl_hw_info get_device_hw_info(sycl::device* device_ptr) {
+    sycl_hw_info res;
+    int32_t id =
+        device_ptr->get_info<sycl::ext::intel::info::device::device_id>();
+    res.device_id = id;
+
+    res.name = device_ptr->get_info<sycl::info::device::name>();
+
+    syclex::architecture arch =
+        device_ptr->get_info<syclex::info::device::architecture>();
+    res.arch = arch;
+
+    map<syclex::architecture,
+        std::pair<const char*, sycl_intel_gpu_family>>::iterator it =
+        arch2name.find(res.arch);
+    if (it != arch2name.end()) {
+        res.arch_name = it->second.first;
+        res.gpu_family = it->second.second;
+    } else {
+        res.arch_name = "unknown";
+        res.gpu_family = GPU_FAMILY_UKNOWN;
+    }
+
+    return res;
+}
@@ -9,18 +9,30 @@
 #include <sycl/sycl.hpp>

 namespace syclex = sycl::ext::oneapi::experimental;
+using gpu_arch = sycl::ext::oneapi::experimental::architecture;

-// TODO: currently not used
-/*
-struct sycl_hw_info {
-  syclex::architecture arch;
-  int32_t device_id;
+// It's used to mark the GPU computing capacity in Flash-attention OP
+// The value must flow the order of performance.
+enum sycl_intel_gpu_family {
+  GPU_FAMILY_UKNOWN = -1,
+  // Normal iGPU in Intel Core CPU, before Meteor Lake iGPU(Xe)
+  GPU_FAMILY_IGPU_NON_XE = 0,
+  // iGPU with Xe core, Meteor Lake iGPU or newer.
+  GPU_FAMILY_IGPU_XE = 1,
+  // dGPU for gaming in client/data center, DG1/FLex 140 or newer.
+  GPU_FAMILY_DGPU_CLIENT_GAME = 2,
+  // dGPU for AI in cloud, PVC or newer.
+  GPU_FAMILY_DGPU_CLOUD = 3
 };

-bool is_in_vector(std::vector<int> &vec, int item);
+struct sycl_hw_info {
+  syclex::architecture arch;
+  const char* arch_name;
+  int32_t device_id;
+  std::string name;
+  sycl_intel_gpu_family gpu_family;
+};

 sycl_hw_info get_device_hw_info(sycl::device *device_ptr);
-*/
-

 #endif // SYCL_HW_HPP
@@ -0,0 +1,410 @@
+#include "upscale.hpp"
+
+static void upscale_f32(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3) {
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    int  index    = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (index >= ne10 * ne11 * ne12 * ne13) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
+
+    int i00 = i10 / sf0;
+    int i01 = i11 / sf1;
+    int i02 = i12 / sf2;
+    int i03 = i13 / sf3;
+
+    dst[index] = *((const float*)((const char*)x + i03 * nb03 + i02 * nb02 +
+                                  i01 * nb01 + i00 * nb00));
+}
+
+static void upscale_f32_bilinear(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset) {
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t index = item_ct1.get_local_id(2) +
+        item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int64_t dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+  if (index >= dst_total_elements) {
+    return;
+  }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    int         y0_src    = (int) sycl::floor((float) y_src_f);
+    int y1_src    = y0_src + 1;
+
+    y0_src = sycl::max(0, sycl::min(y0_src, ne01_src - 1));
+    y1_src = sycl::max(0, sycl::min(y1_src, ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy       = sycl::max(0.0f, sycl::min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    int   x0_src    = (int) sycl::floor(x_src_f);
+    int x1_src    = x0_src + 1;
+
+    x0_src = sycl::max(0, sycl::min(x0_src, ne00_src - 1));
+    x1_src = sycl::max(0, sycl::min(x1_src, ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx       = sycl::max(0.0f, sycl::min(dx, 1.0f));
+
+    const float* p_a =
+        (const float*)((const char*)x + (int64_t)x0_src * nb00 +
+                       (int64_t)y0_src * nb01 + (int64_t)i02_src * nb02 +
+                       (int64_t)i03_src * nb03);
+    const float* p_b =
+        (const float*)((const char*)x + (int64_t)x1_src * nb00 +
+                       (int64_t)y0_src * nb01 + (int64_t)i02_src * nb02 +
+                       (int64_t)i03_src * nb03);
+    const float* p_c =
+        (const float*)((const char*)x + (int64_t)x0_src * nb00 +
+                       (int64_t)y1_src * nb01 + (int64_t)i02_src * nb02 +
+                       (int64_t)i03_src * nb03);
+    const float* p_d =
+        (const float*)((const char*)x + (int64_t)x1_src * nb00 +
+                       (int64_t)y1_src * nb01 + (int64_t)i02_src * nb02 +
+                       (int64_t)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst[index] = result;
+}
+
+// Similar to F.interpolate(..., mode="bilinear", align_corners=False, antialias=True)
+// https://github.com/pytorch/pytorch/blob/8871ff29b743948d1225389d5b7068f37b22750b/aten/src/ATen/native/cpu/UpSampleKernel.cpp
+static void upscale_f32_bilinear_antialias(const float * src0,
+                                           float *       dst,
+                                           const int     nb00,
+                                           const int     nb01,
+                                           const int     nb02,
+                                           const int     nb03,
+                                           const int     ne00_src,
+                                           const int     ne01_src,
+                                           const int     ne10_dst,
+                                           const int     ne11_dst,
+                                           const int     ne12_dst,
+                                           const int     ne13_dst,
+                                           const float   sf0,
+                                           const float   sf1,
+                                           const float   sf2,
+                                           const float   sf3,
+                                           const float   pixel_offset) {
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t index = item_ct1.get_local_id(2) +
+        item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int64_t dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y = ((float)i11_dst + pixel_offset) / sf1;
+    const float x = ((float)i10_dst + pixel_offset) / sf0;
+
+    // support and invscale, minimum 1 pixel for bilinear
+    const float support1  = sycl::max(1.0f / sf1, 1.0f);
+    const float invscale1 = 1.0f / support1;
+    const float support0  = sycl::max(1.0f / sf0, 1.0f);
+    const float invscale0 = 1.0f / support0;
+
+    // the range of source pixels that contribute
+    const int64_t x_min = sycl::max(int64_t(0), int64_t(x - support0 + pixel_offset));
+    const int64_t x_max = sycl::min(int64_t(ne00_src), int64_t(x + support0 + pixel_offset));
+    const int64_t y_min = sycl::max(int64_t(0), int64_t(y - support1 + pixel_offset));
+    const int64_t y_max = sycl::min(int64_t(ne01_src), int64_t(y + support1 + pixel_offset));
+
+    // bilinear filter with antialiasing
+    float val = 0.0f;
+    float total_weight = 0.0f;
+
+    auto triangle_filter = [](float x) -> float {
+        return sycl::max(1.0f - sycl::fabs(x), 0.0f);
+    };
+
+    for (int64_t sy = y_min; sy < y_max; sy++) {
+        const float weight_y = triangle_filter((sy - y + pixel_offset) * invscale1);
+
+        for (int64_t sx = x_min; sx < x_max; sx++) {
+            const float weight_x = triangle_filter((sx - x + pixel_offset) * invscale0);
+            const float weight = weight_x * weight_y;
+
+            if (weight <= 0.0f) {
+                continue;
+            }
+
+            const float pixel =
+                *(const float*)((const char*)src0 + sx * nb00 + sy * nb01 +
+                                i02_src * nb02 + i03_src * nb03);
+            val += pixel * weight;
+            total_weight += weight;
+        }
+    }
+
+    if (total_weight > 0.0f) {
+        val /= total_weight;
+    }
+
+    dst[index] = val;
+}
+
+namespace bicubic_interpolation {
+static float weight1(float x, const float &a) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+static float weight2(float x, const float &a) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+
+static float bicubic(float p0, float p1, float p2, float p3, float x, float a) {
+    const float w0 = weight2(x + 1, a);
+    const float w1 = weight1(x + 0, a);
+    const float w2 = weight1(1 - x, a);
+    const float w3 = weight2(2 - x, a);
+    return p0 * w0 + p1 * w1 + p2 * w2 + p3 * w3;
+};
+
+}
+
+static void upscale_f32_bicubic(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset) {
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const float a = -0.75f;
+    using bicubic_interpolation::bicubic;
+
+    const int64_t index = item_ct1.get_local_id(2) +
+        item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int64_t dst_total_elements =
+        ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    const int   y0_src  = (int) sycl::floor((float) y_src_f);
+    const float dy      = y_src_f - (float)y0_src;
+
+    const float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    const int   x0_src  = (int) sycl::floor((float) x_src_f);
+    const float dx      = x_src_f - (float)x0_src;
+
+    const char * x_base = (const char *)x + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03;
+
+    auto load = [=](int x_off, int y_off) -> float {
+        int i00_src = sycl::max(0, sycl::min(x0_src + x_off, ne00_src - 1));
+        int i01_src = sycl::max(0, sycl::min(y0_src + y_off, ne01_src - 1));
+        return *(const float *)(x_base + (int64_t)i00_src * nb00 + (int64_t)i01_src * nb01);
+    };
+
+    const float result = bicubic(
+        bicubic(load(-1, -1), load(0, -1), load(1, -1), load(2, -1), dx, a),
+        bicubic(load(-1, 0), load(0, 0), load(1, 0), load(2, 0), dx, a),
+        bicubic(load(-1, 1), load(0, 1), load(1, 1), load(2, 1), dx, a),
+        bicubic(load(-1, 2), load(0, 2), load(1, 2), load(2, 2), dx, a),
+        dy,
+        a);
+
+    dst[index] = result;
+}
+
+static void upscale_f32_sycl(const float *   x,
+                             float *         dst,
+                             const int       nb00,
+                             const int       nb01,
+                             const int       nb02,
+                             const int       nb03,
+                             const int       ne10,
+                             const int       ne11,
+                             const int       ne12,
+                             const int       ne13,
+                             const float     sf0,
+                             const float     sf1,
+                             const float     sf2,
+                             const float     sf3,
+                             dpct::queue_ptr stream) {
+    const int64_t dst_size   = ne10 * ne11 * ne12 * ne13;
+    const int64_t num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
+
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE),
+             sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            upscale_f32(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
+        });
+}
+
+static void upscale_f32_bilinear_sycl(const float *   x,
+                                      float *         dst,
+                                      const int       nb00,
+                                      const int       nb01,
+                                      const int       nb02,
+                                      const int       nb03,
+                                      const int       ne00_src,
+                                      const int       ne01_src,
+                                      const int       ne10_dst,
+                                      const int       ne11_dst,
+                                      const int       ne12_dst,
+                                      const int       ne13_dst,
+                                      const float     sf0,
+                                      const float     sf1,
+                                      const float     sf2,
+                                      const float     sf3,
+                                      const float     pixel_offset,
+                                      bool            antialias,
+                                      dpct::queue_ptr stream) {
+    const int64_t dst_size   = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    const int64_t num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
+
+    if (antialias) {
+        stream->parallel_for(
+            sycl::nd_range<3>(
+                sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE),
+                sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                upscale_f32_bilinear_antialias(
+                    x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst,
+                    ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+            });
+    } else {
+        stream->parallel_for(
+            sycl::nd_range<3>(
+                sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE),
+                sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                upscale_f32_bilinear(
+                    x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst,
+                    ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+            });
+    }
+}
+
+static void upscale_f32_bicubic_sycl(const float *   x,
+                                     float *         dst,
+                                     const int       nb00,
+                                     const int       nb01,
+                                     const int       nb02,
+                                     const int       nb03,
+                                     const int       ne00_src,
+                                     const int       ne01_src,
+                                     const int       ne10_dst,
+                                     const int       ne11_dst,
+                                     const int       ne12_dst,
+                                     const int       ne13_dst,
+                                     const float     sf0,
+                                     const float     sf1,
+                                     const float     sf2,
+                                     const float     sf3,
+                                     const float     pixel_offset,
+                                     dpct::queue_ptr stream) {
+    const int64_t dst_size   = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    const int64_t num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
+
+    {
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(
+                    sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE),
+                    sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    upscale_f32_bicubic(
+                        x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst,
+                        ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+                });
+        });
+    }
+}
+
+void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr     stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int mode_flags = dst->op_params[0];
+    const ggml_scale_mode mode = (ggml_scale_mode)(mode_flags & 0xFF);
+
+    float sf0 = (float)dst->ne[0]/src0->ne[0];
+    float sf1 = (float)dst->ne[1]/src0->ne[1];
+    float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const float sf3 = (float)dst->ne[3]/src0->ne[3];
+
+    float pixel_offset = 0.5f;
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        sf0 = dst->ne[0] > 1 && src0->ne[0] > 1
+            ? (float)(dst->ne[0] - 1) / (src0->ne[0] - 1)
+            : sf0;
+        sf1 = dst->ne[1] > 1 && src0->ne[1] > 1
+            ? (float)(dst->ne[1] - 1) / (src0->ne[1] - 1)
+            : sf1;
+        pixel_offset = 0.0f;
+    }
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        upscale_f32_sycl(
+            src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        const bool antialias = (mode_flags & GGML_SCALE_FLAG_ANTIALIAS);
+        upscale_f32_bilinear_sycl(
+            src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+            sf0, sf1, sf2, sf3, pixel_offset, antialias, stream);
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        upscale_f32_bicubic_sycl(
+            src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+            sf0, sf1, sf2, sf3, pixel_offset, stream);
+    }
+}
+
+void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_upscale(ctx, dst);
+}
@@ -0,0 +1,9 @@
+#pragma once
+
+#include <sycl/sycl.hpp>
+#include "dpct/helper.hpp"
+#include "common.hpp"
+
+#define SYCL_UPSCALE_BLOCK_SIZE 256
+
+void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
@@ -4981,8 +4981,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
        std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties();

        // Try to find a non-graphics compute queue and transfer-focused queues
-        const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1);
-        const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | vk::QueueFlagBits::eGraphics, compute_queue_family_index, 1);
+        // On AMD, the graphics queue seems to be faster, so don't avoid it
+        const vk::QueueFlagBits graphics_flag = device->vendor_id == VK_VENDOR_ID_AMD ? (vk::QueueFlagBits)0 : vk::QueueFlagBits::eGraphics;
+        const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, graphics_flag, -1, 1);
+        const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | graphics_flag, compute_queue_family_index, 1);

        const float priorities[] = { 1.0f, 1.0f };
        device->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1;
@@ -5441,13 +5443,11 @@ static vk_device ggml_vk_get_device(size_t idx) {

        ggml_vk_load_shaders(device);

-        const bool prefers_transfer_queue = device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != AMD_GCN;
-
        if (!device->single_queue) {
            const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
            ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);

-            device->async_use_transfer_queue = prefers_transfer_queue || (getenv("GGML_VK_ASYNC_USE_TRANSFER_QUEUE") != nullptr);
+            device->async_use_transfer_queue = (getenv("GGML_VK_ASYNC_USE_TRANSFER_QUEUE") != nullptr);
        } else {
            // TODO: Use pointer or reference to avoid copy
            device->transfer_queue.copyFrom(device->compute_queue);
@@ -245,7 +245,7 @@ void main() {
 #endif
                    }
                    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                        Sf[r][c] += ACC_TYPE(dot(Q_cache[r], K_Tf));
+                        Sf[r][c] += dot(ACC_TYPEV4(Q_cache[r]), ACC_TYPEV4(K_Tf));
                    }
                }
            }
@@ -270,7 +270,7 @@ void main() {
 #endif
                    }
                    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                        Sf[r][c] += ACC_TYPE(dot(Qf[tile_row(r) * qf_stride + d * D_split + d_tid], K_Tf));
+                        Sf[r][c] += dot(ACC_TYPEV4(Qf[tile_row(r) * qf_stride + d * D_split + d_tid]), ACC_TYPEV4(K_Tf));
                    }
                }
            }
@@ -478,6 +478,7 @@ class MODEL_ARCH(IntEnum):
    RND1             = auto()
    PANGU_EMBED      = auto()
    MISTRAL3         = auto()
+    MISTRAL4         = auto()
    PADDLEOCR        = auto()
    MIMO2            = auto()
    STEP35           = auto()
@@ -924,6 +925,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.RND1:             "rnd1",
    MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
    MODEL_ARCH.MISTRAL3:         "mistral3",
+    MODEL_ARCH.MISTRAL4:         "mistral4",
    MODEL_ARCH.PADDLEOCR:        "paddleocr",
    MODEL_ARCH.MIMO2:            "mimo2",
    MODEL_ARCH.STEP35:           "step35",
@@ -3538,6 +3540,37 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN_EXP,
        MODEL_TENSOR.FFN_UP_EXP,
    ],
+    MODEL_ARCH.MISTRAL4: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
    MODEL_ARCH.MIMO2: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -1 +1 @@
-d6754f3d0e6d0acd21c12442353c9fd2f94188e7
+553552e1d88be2b214b85e5159eedd39a63e2c34
@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess

-HTTPLIB_VERSION = "refs/tags/v0.37.2"
+HTTPLIB_VERSION = "refs/tags/v0.38.0"

 vendor = {
    "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
@@ -123,6 +123,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_RND1,             "rnd1"             },
    { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
    { LLM_ARCH_MISTRAL3,         "mistral3"         },
+    { LLM_ARCH_MISTRAL4,         "mistral4"         },
    { LLM_ARCH_PADDLEOCR,        "paddleocr"        },
    { LLM_ARCH_MIMO2,            "mimo2"            },
    { LLM_ARCH_STEP35,           "step35"           },
@@ -1589,6 +1590,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                LLM_TENSOR_FFN_UP_SHEXP,
            };
        case LLM_ARCH_DEEPSEEK2:
+        case LLM_ARCH_MISTRAL4:
            return {
                LLM_TENSOR_TOKEN_EMBD,
                LLM_TENSOR_OUTPUT_NORM,
@@ -127,6 +127,7 @@ enum llm_arch {
    LLM_ARCH_RND1,
    LLM_ARCH_PANGU_EMBED,
    LLM_ARCH_MISTRAL3,
+    LLM_ARCH_MISTRAL4,
    LLM_ARCH_PADDLEOCR,
    LLM_ARCH_MIMO2,
    LLM_ARCH_STEP35,
@@ -1953,6 +1953,12 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32

            cells.pos_set(i, pos);

+            if (hparams.n_pos_per_embd() > 1) {
+                llama_kv_cell_ext ext;
+                io.read_to(&ext, sizeof(ext));
+                cells.ext_set(i, ext);
+            }
+
            for (uint32_t j = 0; j < n_seq_id; ++j) {
                llama_seq_id seq_id;
                io.read_to(&seq_id, sizeof(seq_id));
@@ -1587,6 +1587,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                }
            } break;
        case LLM_ARCH_DEEPSEEK2:
+        case LLM_ARCH_MISTRAL4:
            {
                // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B, Kanana-2-30B-A3B
                const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26 || (hparams.n_layer == 48 && n_vocab == 128256));
@@ -4883,6 +4884,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                    }
                } break;
            case LLM_ARCH_DEEPSEEK2:
+            case LLM_ARCH_MISTRAL4:
                {
                    const bool is_mla = hparams.is_mla();

@@ -7462,6 +7464,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
            if (!layer.wo_s && layer.wo) {
                layer.wo_s = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}, TENSOR_NOT_REQUIRED);
            }
+            if (!layer.wqkv_s && layer.wqkv) {
+                layer.wqkv_s = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.wqkv_gate_s && layer.wqkv_gate) {
+                layer.wqkv_gate_s = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }

            // dense FFN weight scales (per-tensor, shape {1})
            if (!layer.ffn_gate_s && layer.ffn_gate) {
@@ -7473,6 +7481,15 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
            if (!layer.ffn_up_s && layer.ffn_up) {
                layer.ffn_up_s = create_tensor(tn(LLM_TENSOR_FFN_UP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
            }
+            if (!layer.ffn_gate_shexp_s && layer.ffn_gate_shexp) {
+                layer.ffn_gate_shexp_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ffn_down_shexp_s && layer.ffn_down_shexp) {
+                layer.ffn_down_shexp_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ffn_up_shexp_s && layer.ffn_up_shexp) {
+                layer.ffn_up_shexp_s = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }

            // MoE expert weight scales (per-expert, shape {n_expert})
            if (!layer.ffn_gate_exps_s && layer.ffn_gate_exps) {
@@ -7484,6 +7501,20 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
            if (!layer.ffn_up_exps_s && layer.ffn_up_exps) {
                layer.ffn_up_exps_s = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
            }
+
+            // recurrent / linear-attention weight scales (per-tensor, shape {1})
+            if (!layer.ssm_in_s && layer.ssm_in) {
+                layer.ssm_in_s = create_tensor(tn(LLM_TENSOR_SSM_IN, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ssm_out_s && layer.ssm_out) {
+                layer.ssm_out_s = create_tensor(tn(LLM_TENSOR_SSM_OUT, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ssm_alpha_s && layer.ssm_alpha) {
+                layer.ssm_alpha_s = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ssm_beta_s && layer.ssm_beta) {
+                layer.ssm_beta_s = create_tensor(tn(LLM_TENSOR_SSM_BETA, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
        }
    }

@@ -7821,7 +7852,7 @@ void llama_model::print_info() const {
        LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
    }

-    if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA) {
+    if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
        LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
        LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
        LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
@@ -8399,6 +8430,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
            } break;
        case LLM_ARCH_DEEPSEEK2:
        case LLM_ARCH_GLM_DSA:
+        case LLM_ARCH_MISTRAL4:
            {
                llm = std::make_unique<llm_build_deepseek2>(*this, params);
            } break;
@@ -8810,6 +8842,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_ERNIE4_5:
        case LLM_ARCH_ERNIE4_5_MOE:
        case LLM_ARCH_MISTRAL3:
+        case LLM_ARCH_MISTRAL4:
        case LLM_ARCH_LLAMA_EMBED:
        case LLM_ARCH_MAINCODER:
        case LLM_ARCH_GLM_DSA:
@@ -401,9 +401,18 @@ struct llama_layer {
    struct ggml_tensor * wk_s       = nullptr;
    struct ggml_tensor * wv_s       = nullptr;
    struct ggml_tensor * wo_s       = nullptr;
+    struct ggml_tensor * wqkv_s     = nullptr;
+    struct ggml_tensor * wqkv_gate_s = nullptr;
    struct ggml_tensor * ffn_gate_s = nullptr;
    struct ggml_tensor * ffn_up_s   = nullptr;
    struct ggml_tensor * ffn_down_s = nullptr;
+    struct ggml_tensor * ffn_gate_shexp_s = nullptr;
+    struct ggml_tensor * ffn_up_shexp_s   = nullptr;
+    struct ggml_tensor * ffn_down_shexp_s = nullptr;
+    struct ggml_tensor * ssm_in_s    = nullptr;
+    struct ggml_tensor * ssm_out_s   = nullptr;
+    struct ggml_tensor * ssm_alpha_s = nullptr;
+    struct ggml_tensor * ssm_beta_s  = nullptr;

    // altup & laurel
    struct ggml_tensor * per_layer_inp_gate   = nullptr;
@@ -42,7 +42,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba_layer(llm_graph_input_rs * inp,
    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);

    // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
-    ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur);
+    ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur, layer.ssm_in_s);
    // split the above in two
    // => {d_inner, n_seq_tokens, n_seqs}
    ggml_tensor * x  = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0);
@@ -137,7 +137,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba_layer(llm_graph_input_rs * inp,
        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);

        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
-        cur = build_lora_mm(layer.ssm_out, y);
+        cur = build_lora_mm(layer.ssm_out, y, layer.ssm_out_s);
    }

    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
@@ -184,7 +184,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
    // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads

    // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
-    ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
+    ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur, model.layers[il].ssm_in_s);

    // split the above in three
    ggml_tensor * z   = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0],
@@ -278,7 +278,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
        y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);

        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
-        cur = build_lora_mm(model.layers[il].ssm_out, y);
+        cur = build_lora_mm(model.layers[il].ssm_out, y, model.layers[il].ssm_out_s);
    }

    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
@@ -107,9 +107,9 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
 ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il) {
    if (model.layers[il].ffn_gate_inp == nullptr) {
        cur = build_ffn(cur,
-                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,
                NULL,                      NULL,                        NULL,
-                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
                NULL,
                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
        cb(cur, "ffn_out", il);
@@ -136,7 +136,10 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
                    hparams.expert_weights_scale,
                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
                    il,
-                    router_logits);
+                    router_logits, nullptr,
+                    model.layers[il].ffn_up_exps_s,
+                    nullptr, // no gate
+                    model.layers[il].ffn_down_exps_s);
        cb(moe_out, "ffn_moe_out", il);

        if (model.layers[il].ffn_latent_up) {
@@ -144,9 +147,9 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
        }

        ggml_tensor * ffn_shexp = build_ffn(inp_emb,
-                    model.layers[il].ffn_up_shexp,  NULL, NULL,
-                    NULL /* no gate */           ,  NULL, NULL,
-                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    model.layers[il].ffn_up_shexp,   NULL, model.layers[il].ffn_up_shexp_s,
+                    NULL /* no gate */           ,   NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
                    NULL,
                    LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
        cb(ffn_shexp, "ffn_shexp", il);
@@ -90,11 +90,11 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_qkvz(
    const int64_t n_seqs       = ubatch.n_seqs;
    const int64_t n_seq_tokens = ubatch.n_seq_tokens;

-    ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);
+    ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input, model.layers[il].wqkv_s);
    qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);
    cb(qkv_mixed, "linear_attn_qkv_mixed", il);

-    ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);
+    ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input, model.layers[il].wqkv_gate_s);
    cb(z, "z", il);

    return { qkv_mixed, z };
@@ -123,7 +123,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn(
    // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention

    // Qwen3Next uses a single Q projection that outputs query + gate
-    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ]
+    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s); // [ (n_embd_head * 2) * n_head, n_tokens ]
    cb(Qcur_full, "Qcur_full", il);

    ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
@@ -135,10 +135,10 @@ ggml_tensor * llm_build_qwen35::build_layer_attn(
    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
    cb(Qcur, "Qcur_normed", il);

-    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
    cb(Kcur, "Kcur", il);

-    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
    cb(Vcur, "Vcur", il);

    // Apply K normalization
@@ -186,7 +186,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn(
    cur = ggml_mul(ctx0, cur, gate_sigmoid);
    cb(cur, "attn_gated", il);

-    cur = build_lora_mm(model.layers[il].wo, cur);
+    cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
    cb(cur, "attn_output", il);

    return cur;
@@ -217,14 +217,14 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
    ggml_tensor * qkv_mixed = qkvz.first;
    ggml_tensor * z         = qkvz.second;

-    ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
+    ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur, model.layers[il].ssm_beta_s);
    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
    cb(beta, "beta", il);

    beta = ggml_sigmoid(ctx0, beta);

-    ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
-    alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur, model.layers[il].ssm_alpha_s);
+    alpha = ggml_reshape_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
    cb(alpha, "alpha", il);

    ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
@@ -356,7 +356,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
    cb(final_output, "final_output", il);

    // Output projection
-    cur = build_lora_mm(model.layers[il].ssm_out, final_output);
+    cur = build_lora_mm(model.layers[il].ssm_out, final_output, model.layers[il].ssm_out_s);
    cb(cur, "linear_attn_out", il);

    // Reshape back to original dimensions
@@ -370,9 +370,9 @@ ggml_tensor * llm_build_qwen35::build_layer_ffn(ggml_tensor * cur, const int il)
    GGML_ASSERT(model.layers[il].ffn_gate_inp == nullptr);

    cur = build_ffn(cur,
-        model.layers[il].ffn_up, NULL, NULL,
-        model.layers[il].ffn_gate, NULL, NULL,
-        model.layers[il].ffn_down, NULL, NULL,
+        model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
+        model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+        model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
        NULL,
        LLM_FFN_SILU, LLM_FFN_PAR, il);
    cb(cur, "ffn_out", il);
@@ -90,11 +90,11 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
    const int64_t n_seqs       = ubatch.n_seqs;
    const int64_t n_seq_tokens = ubatch.n_seq_tokens;

-    ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);
+    ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input, model.layers[il].wqkv_s);
    qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);
    cb(qkv_mixed, "linear_attn_qkv_mixed", il);

-    ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);
+    ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input, model.layers[il].wqkv_gate_s);
    cb(z, "z", il);

    return { qkv_mixed, z };
@@ -123,7 +123,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
    // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention

    // Qwen3Next uses a single Q projection that outputs query + gate
-    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ]
+    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s); // [ (n_embd_head * 2) * n_head, n_tokens ]
    cb(Qcur_full, "Qcur_full", il);

    ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
@@ -135,10 +135,10 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
    cb(Qcur, "Qcur_normed", il);

-    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
    cb(Kcur, "Kcur", il);

-    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
    cb(Vcur, "Vcur", il);

    // Apply K normalization
@@ -186,7 +186,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
    cur = ggml_mul(ctx0, cur, gate_sigmoid);
    cb(cur, "attn_gated", il);

-    cur = build_lora_mm(model.layers[il].wo, cur);
+    cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
    cb(cur, "attn_output", il);

    return cur;
@@ -217,14 +217,14 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
    ggml_tensor * qkv_mixed = qkvz.first;
    ggml_tensor * z         = qkvz.second;

-    ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
+    ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur, model.layers[il].ssm_beta_s);
    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
    cb(beta, "beta", il);

    beta = ggml_sigmoid(ctx0, beta);

-    ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
-    alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur, model.layers[il].ssm_alpha_s);
+    alpha = ggml_reshape_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
    cb(alpha, "alpha", il);

    ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
@@ -356,7 +356,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
    cb(final_output, "final_output", il);

    // Output projection
-    cur = build_lora_mm(model.layers[il].ssm_out, final_output);
+    cur = build_lora_mm(model.layers[il].ssm_out, final_output, model.layers[il].ssm_out_s);
    cb(cur, "linear_attn_out", il);

    // Reshape back to original dimensions
@@ -380,16 +380,19 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int
            LLM_FFN_SILU, true,
            hparams.expert_weights_scale,
            LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
-            nullptr, model.layers[il].ffn_gate_up_exps);
+            nullptr, model.layers[il].ffn_gate_up_exps,
+            model.layers[il].ffn_up_exps_s,
+            model.layers[il].ffn_gate_exps_s,
+            model.layers[il].ffn_down_exps_s);
    cb(moe_out, "ffn_moe_out", il);

    // Add shared experts if present - following Qwen3Next reference implementation
    if (model.layers[il].ffn_up_shexp != nullptr) {
        ggml_tensor * ffn_shexp =
            build_ffn(cur,
-                model.layers[il].ffn_up_shexp, NULL, NULL,
-                model.layers[il].ffn_gate_shexp, NULL, NULL,
-                model.layers[il].ffn_down_shexp, NULL, NULL,
+                model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,
+                model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
+                model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
                NULL,
                LLM_FFN_SILU, LLM_FFN_PAR, il);
        cb(ffn_shexp, "ffn_shexp", il);
@@ -8576,11 +8576,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
        }
    }

-    for (int hsk : { 40, 64, 72, 80, 96, 128, 192, 256, 576 }) {
+    for (int hsk : { 40, 64, 72, 80, 96, 128, 192, 256, 320, 576 }) {
        for (int hsv : { 40, 64, 72, 80, 96, 128, 192, 256, 512 }) {
-            if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
+            if (hsk != 192 && hsk != 320 && hsk != 576 && hsk != hsv) continue;
            if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
            if (hsk == 576 && hsv != 512) continue; // DeepSeek MLA
+            if (hsk == 320 && hsv != 256) continue; // MLA

            for (bool mask : { true, false } ) {
                for (bool sinks : { true, false } ) {
@@ -8589,12 +8590,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                        for (float logit_softcap : {0.0f, 10.0f}) {
                            if (hsk != 128 && logit_softcap != 0.0f) continue;
                            for (int nh : { 1, 4 }) {
-                                if (nh == 1 && hsk != 576) continue; // GLM 4.7 Flash
+                                if (nh == 1 && hsk != 320 && hsk != 576) continue; // GLM 4.7 Flash
                                for (int nr3 : { 1, 3, }) {
                                    if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes
-                                    for (int nr2 : { 1, 4, 12, 20 }) {
+                                    for (int nr2 : { 1, 4, 12, 20, 32 }) {
                                        if (nr2 == 12 && hsk != 128) continue;
                                        if (nr2 == 20 && (nh != 1 || hsk != 576)) continue;
+                                        if (nr2 == 32 && (nh != 1 || hsk != 320)) continue;
                                        //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) {
                                        for (int kv : { 113, 512, 1024, }) {
                                            if (nr2 != 1 && kv != 512) continue;
@@ -1915,7 +1915,7 @@ env.globals["raise_exception"] = raise_exception

 template = env.from_string(tmpl)
 result = template.render(**vars_json)
-print(result, end='')
+sys.stdout.buffer.write(result.encode())
 )";

 static void test_template_py(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect) {
@@ -90,7 +90,10 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
        n_embd = 64;
        n_head = 1;
        n_ff   = 96;
-    } else if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_KIMI_LINEAR) {
+    } else if (arch == LLM_ARCH_DEEPSEEK2
+            || arch == LLM_ARCH_GLM_DSA
+            || arch == LLM_ARCH_KIMI_LINEAR
+            || arch == LLM_ARCH_MISTRAL4) {
        n_embd = 128;
        n_head = 1;
        n_ff   = 192;
@@ -145,7 +148,10 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
    }

    ms.add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, 8.0f);
-    if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA  || arch == LLM_ARCH_KIMI_LINEAR) {
+    if (arch == LLM_ARCH_DEEPSEEK2
+            || arch == LLM_ARCH_GLM_DSA
+            || arch == LLM_ARCH_KIMI_LINEAR
+            || arch == LLM_ARCH_MISTRAL4) {
        ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH,       uint32_t(576));
        ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,     uint32_t(512));
        ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,       uint32_t(64));
@@ -319,6 +325,7 @@ static bool moe_mandatory(const llm_arch arch) {
        case LLM_ARCH_MIMO2:
        case LLM_ARCH_KIMI_LINEAR:
        case LLM_ARCH_STEP35:
+        case LLM_ARCH_MISTRAL4:
            return true;
        default:
            return false;
@@ -215,7 +215,7 @@ struct cli_context {
        inputs.parallel_tool_calls   = false;
        inputs.add_generation_prompt = true;
        inputs.reasoning_format      = COMMON_REASONING_FORMAT_DEEPSEEK;
-        inputs.enable_thinking       = common_chat_templates_support_enable_thinking(chat_params.tmpls.get());
+        inputs.enable_thinking       = chat_params.enable_thinking ? common_chat_templates_support_enable_thinking(chat_params.tmpls.get()) : false;

        // Apply chat template to the list of messages
        return common_chat_templates_apply(chat_params.tmpls.get(), inputs);
@@ -1273,17 +1273,27 @@ json convert_responses_to_chatcmpl(const json & response_body) {

                for (const auto & output_text : item.at("content")) {
                    const std::string type = json_value(output_text, "type", std::string());
-                    if (type != "output_text") {
-                        throw std::invalid_argument("'type' must be 'output_text'");
+                    if (type == "output_text") {
+                        if (!exists_and_is_string(output_text, "text")) {
+                            throw std::invalid_argument("'Output text' requires 'text'");
+                            // Ignore annotations and logprobs for now
+                            chatcmpl_content.push_back({
+                                {"text", output_text.at("text")},
+                                {"type", "text"},
+                            });
+                        }
+                    } else if (type == "refusal") {
+                        if (!exists_and_is_string(output_text, "refusal")) {
+                            throw std::invalid_argument("'Refusal' requires 'refusal'");
+                            // Ignore annotations and logprobs for now
+                            chatcmpl_content.push_back({
+                                {"refusal", output_text.at("refusal")},
+                                {"type", "refusal"},
+                            });
+                        }
+                    } else {
+                        throw std::invalid_argument("'type' must be one of 'output_text' or 'refusal'");
                    }
-                    if (!exists_and_is_string(output_text, "text")) {
-                        throw std::invalid_argument("'Output text' requires 'text'");
-                    }
-                    // Ignore annotations and logprobs for now
-                    chatcmpl_content.push_back({
-                        {"text", output_text.at("text")},
-                        {"type", "text"},
-                    });
                }

                if (merge_prev) {
@@ -563,7 +563,7 @@ def test_cancel_request():
    except requests.exceptions.ReadTimeout:
        pass # expected
    # make sure the slot is free
-    time.sleep(1) # wait for HTTP_POLLING_SECONDS
+    time.sleep(2)
    res = server.make_request("GET", "/slots")
    assert res.body[0]["is_processing"] == False

@@ -939,7 +939,6 @@
 			"integrity": "sha512-oJrXtQiAXLvT9clCf1K4kxp3eKsQhIaZqxEyowkBcsvZDdZkbWrVmnGknxs5flTD0VGsxrxKgBCZty1EzoiMzA==",
 			"dev": true,
 			"license": "Apache-2.0",
-			"peer": true,
 			"dependencies": {
 				"@swc/helpers": "^0.5.0"
 			}
@@ -2161,7 +2160,6 @@
 			"integrity": "sha512-W9R51zUCd2iHOQBg/D93+bdpYv6kbtFx+kft5X8lPKQl6yEu0aKs9i5N5GyCASOhIApgx/tkqZIJ7vgM4cqrHA==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"ts-dedent": "^2.0.0",
 				"type-fest": "~2.19"
@@ -2245,7 +2243,6 @@
 			"integrity": "sha512-875hTUkEbz+MyJIxWbQjfMaekqdmEKUUfR7JyKcpfMRZqcGyrO9Gd+iS1D/Dx8LpE5FEtutWGOtlAh4ReSAiOA==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@standard-schema/spec": "^1.0.0",
 				"@sveltejs/acorn-typescript": "^1.0.5",
@@ -2289,7 +2286,6 @@
 			"integrity": "sha512-YZs/OSKOQAQCnJvM/P+F1URotNnYNeU3P2s4oIpzm1uFaqUEqRxUB0g5ejMjEb5Gjb9/PiBI5Ktrq4rUUF8UVQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@sveltejs/vite-plugin-svelte-inspector": "^5.0.0",
 				"debug": "^4.4.1",
@@ -2705,7 +2701,6 @@
 			"integrity": "sha512-pemlzrSESWbdAloYml3bAJMEfNh1Z7EduzqPKprCH5S341frlpYnUEW0H72dLxa6IsYr+mPno20GiSm+h9dEdQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@babel/code-frame": "^7.10.4",
 				"@babel/runtime": "^7.12.5",
@@ -2873,7 +2868,6 @@
 			"integrity": "sha512-+0/4J266CBGPUq/ELg7QUHhN25WYjE0wYTPSQJn1xeu8DOlIOPxXxrNGiLmfAWl7HMMgWFWXpt9IDjMWrF5Iow==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"undici-types": "~7.16.0"
 			}
@@ -2940,7 +2934,6 @@
 			"integrity": "sha512-IgSWvLobTDOjnaxAfDTIHaECbkNlAlKv2j5SjpB2v7QHKv1FIfjwMy8FsDbVfDX/KjmCmYICcw7uGaXLhtsLNg==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@typescript-eslint/scope-manager": "8.56.0",
 				"@typescript-eslint/types": "8.56.0",
@@ -3177,7 +3170,6 @@
 			"integrity": "sha512-tJxiPrWmzH8a+w9nLKlQMzAKX/7VjFs50MWgcAj7p9XQ7AQ9/35fByFYptgPELyLw+0aixTnC4pUWV+APcZ/kw==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@testing-library/dom": "^10.4.0",
 				"@testing-library/user-event": "^14.6.1",
@@ -3305,7 +3297,6 @@
 			"integrity": "sha512-oukfKT9Mk41LreEW09vt45f8wx7DordoWUZMYdY/cyAk7w5TWkTRCNZYF7sX7n2wB7jyGAl74OxgwhPgKaqDMQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@vitest/utils": "3.2.4",
 				"pathe": "^2.0.3",
@@ -3376,7 +3367,6 @@
 			"resolved": "https://registry.npmjs.org/acorn/-/acorn-8.15.0.tgz",
 			"integrity": "sha512-NZyJarBfL7nWwIq+FDL6Zp/yHEhePMNnnJ0y3qfieCrmNvYct8uvtiV41UvlSe6apAfk0fY1FbWx+NwfmpvtTg==",
 			"license": "MIT",
-			"peer": true,
 			"bin": {
 				"acorn": "bin/acorn"
 			},
@@ -4094,7 +4084,8 @@
 			"resolved": "https://registry.npmjs.org/csstype/-/csstype-3.1.3.tgz",
 			"integrity": "sha512-M1uQkMl8rQK/szD0LNhtqxIPLpimGm8sOBwU7lLnCpSbTyY3yeU1Vc7l4KT5zT4s/yOxHH5O7tIuuLOCnLADRw==",
 			"dev": true,
-			"license": "MIT"
+			"license": "MIT",
+			"peer": true
 		},
 		"node_modules/debug": {
 			"version": "4.4.3",
@@ -4404,7 +4395,6 @@
 			"dev": true,
 			"hasInstallScript": true,
 			"license": "MIT",
-			"peer": true,
 			"bin": {
 				"esbuild": "bin/esbuild"
 			},
@@ -4465,7 +4455,6 @@
 			"integrity": "sha512-LEyamqS7W5HB3ujJyvi0HQK/dtVINZvd5mAAp9eT5S/ujByGjiZLCzPcHVzuXbpJDJF/cxwHlfceVUDZ2lnSTw==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@eslint-community/eslint-utils": "^4.8.0",
 				"@eslint-community/regexpp": "^4.12.1",
@@ -5672,7 +5661,6 @@
 			"resolved": "https://registry.npmjs.org/hono/-/hono-4.11.7.tgz",
 			"integrity": "sha512-l7qMiNee7t82bH3SeyUCt9UF15EVmaBvsppY2zQtrbIhl/yzBTny+YUxsVjSjQ6gaqaeVtZmGocom8TzBlA4Yw==",
 			"license": "MIT",
-			"peer": true,
 			"engines": {
 				"node": ">=16.9.0"
 			}
@@ -8097,7 +8085,6 @@
 				}
 			],
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"nanoid": "^3.3.11",
 				"picocolors": "^1.1.1",
@@ -8231,7 +8218,6 @@
 			"integrity": "sha512-I7AIg5boAr5R0FFtJ6rCfD+LFsWHp81dolrFD8S79U9tb8Az2nGrJncnMSnys+bpQJfRUzqs9hnA81OAA3hCuQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"bin": {
 				"prettier": "bin/prettier.cjs"
 			},
@@ -8248,7 +8234,6 @@
 			"integrity": "sha512-pn1ra/0mPObzqoIQn/vUTR3ZZI6UuZ0sHqMK5x2jMLGrs53h0sXhkVuDcrlssHwIMk7FYrMjHBPoUSyyEEDlBQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"peerDependencies": {
 				"prettier": "^3.0.0",
 				"svelte": "^3.2.0 || ^4.0.0-next.0 || ^5.0.0-next.0"
@@ -8480,7 +8465,6 @@
 			"integrity": "sha512-FS+XFBNvn3GTAWq26joslQgWNoFu08F4kl0J4CgdNKADkdSGXQyTCnKteIAJy96Br6YbpEU1LSzV5dYtjMkMDg==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"engines": {
 				"node": ">=0.10.0"
 			}
@@ -8491,7 +8475,6 @@
 			"integrity": "sha512-Xs1hdnE+DyKgeHJeJznQmYMIBG3TKIHJJT95Q58nHLSrElKlGQqDTR2HQ9fx5CN/Gk6Vh/kupBTDLU11/nDk/g==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"scheduler": "^0.26.0"
 			},
@@ -8766,7 +8749,6 @@
 			"integrity": "sha512-4iya7Jb76fVpQyLoiVpzUrsjQ12r3dM7fIVz+4NwoYvZOShknRmiv+iu9CClZml5ZLGb0XMcYLutK6w9tgxHDw==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@types/estree": "1.0.8"
 			},
@@ -8877,7 +8859,6 @@
 			"integrity": "sha512-elOcIZRTM76dvxNAjqYrucTSI0teAF/L2Lv0s6f6b7FOwcwIuA357bIE871580AjHJuSvLIRUosgV+lIWx6Rgg==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"chokidar": "^4.0.0",
 				"immutable": "^5.0.2",
@@ -9172,7 +9153,6 @@
 			"integrity": "sha512-LwF0VZsT4qkgx66Ad/q0QgZZrU2a5WftaADDEcJ3bGq3O2fHvwWPlSZjM1HiXD4vqP9U5JiMqQkV1gkyH0XJkw==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@storybook/global": "^5.0.0",
 				"@storybook/icons": "^2.0.1",
@@ -9387,7 +9367,6 @@
 			"resolved": "https://registry.npmjs.org/svelte/-/svelte-5.48.3.tgz",
 			"integrity": "sha512-w7QZ398cdNherTdiQ/v3SYLLGOO4948Jgjh04PYqtTYVohmBvbmFwLmo7pp8gp4/1tceRWfSTjHgjtfpCVNJmQ==",
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@jridgewell/remapping": "^2.3.4",
 				"@jridgewell/sourcemap-codec": "^1.5.0",
@@ -9633,7 +9612,6 @@
 			"integrity": "sha512-gBXpgUm/3rp1lMZZrM/w7D8GKqshif0zAymAhbCyIt8KMe+0v9DQ7cdYLR4FHH/cKpdTXb+A/tKKU3eolfsI+g==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"funding": {
 				"type": "github",
 				"url": "https://github.com/sponsors/dcastil"
@@ -9664,8 +9642,7 @@
 			"resolved": "https://registry.npmjs.org/tailwindcss/-/tailwindcss-4.1.11.tgz",
 			"integrity": "sha512-2E9TBm6MDD/xKYe+dvJZAmg3yxIEDNRc0jwlNyDg/4Fil2QcSLjFKGVff0lAf1jjeaArlG/M75Ey/EYr/OJtBA==",
 			"dev": true,
-			"license": "MIT",
-			"peer": true
+			"license": "MIT"
 		},
 		"node_modules/tapable": {
 			"version": "2.2.2",
@@ -9942,7 +9919,6 @@
 			"integrity": "sha512-p1diW6TqL9L07nNxvRMM7hMMw4c5XOo/1ibL4aAIGmSAt9slTE1Xgw5KWuof2uTOvCg9BY7ZRi+GaF+7sfgPeQ==",
 			"dev": true,
 			"license": "Apache-2.0",
-			"peer": true,
 			"bin": {
 				"tsc": "bin/tsc",
 				"tsserver": "bin/tsserver"
@@ -10336,7 +10312,6 @@
 			"integrity": "sha512-BxAKBWmIbrDgrokdGZH1IgkIk/5mMHDreLDmCJ0qpyJaAteP8NvMhkwr/ZCQNqNH97bw/dANTE9PDzqwJghfMQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"esbuild": "^0.25.0",
 				"fdir": "^6.5.0",
@@ -10497,7 +10472,6 @@
 			"integrity": "sha512-LUCP5ev3GURDysTWiP47wRRUpLKMOfPh+yKTx3kVIEiu5KOMeqzpnYNsKyOoVrULivR8tLcks4+lga33Whn90A==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@types/chai": "^5.2.2",
 				"@vitest/expect": "3.2.4",
@@ -10819,7 +10793,6 @@
 			"resolved": "https://registry.npmjs.org/zod/-/zod-4.2.1.tgz",
 			"integrity": "sha512-0wZ1IRqGGhMP76gLqz8EyfBXKk0J2qo2+H3fi4mcUP/KtTocoX08nmIAHl1Z2kJIZbZee8KOpBCSNPRgauucjw==",
 			"license": "MIT",
-			"peer": true,
 			"funding": {
 				"url": "https://github.com/sponsors/colinhacks"
 			}
@@ -11,7 +11,7 @@
 		iconSize?: string;
 		class?: string;
 		disabled?: boolean;
-		onclick: () => void;
+		onclick: (e?: MouseEvent) => void;
 		'aria-label'?: string;
 	}

@@ -65,7 +65,8 @@
 	$effect(() => {
 		if (conversationModel) {
 			modelsStore.selectModelByName(conversationModel);
-		} else if (isRouter && modelsStore.loadedModelIds.length > 0) {
+		} else if (isRouter && !modelsStore.selectedModelId && modelsStore.loadedModelIds.length > 0) {
+			// auto-select the first loaded model only when nothing is selected yet
 			const first = modelOptions().find((m) => modelsStore.loadedModelIds.includes(m.model));
 			if (first) modelsStore.selectModelById(first.id);
 		}
@@ -3,6 +3,7 @@
 	import { Button } from '$lib/components/ui/button';
 	import { DialogConversationSelection, DialogConfirmation } from '$lib/components/app';
 	import { createMessageCountMap } from '$lib/utils';
+	import { ISO_DATE_TIME_SEPARATOR } from '$lib/constants';
 	import { conversationsStore, conversations } from '$lib/stores/conversations.svelte';
 	import { toast } from 'svelte-sonner';

@@ -55,18 +56,10 @@
 				})
 			);

-			const blob = new Blob([JSON.stringify(allData, null, 2)], {
-				type: 'application/json'
-			});
-			const url = URL.createObjectURL(blob);
-			const a = document.createElement('a');
-
-			a.href = url;
-			a.download = `conversations_${new Date().toISOString().split('T')[0]}.json`;
-			document.body.appendChild(a);
-			a.click();
-			document.body.removeChild(a);
-			URL.revokeObjectURL(url);
+			conversationsStore.downloadConversationFile(
+				allData,
+				`${new Date().toISOString().split(ISO_DATE_TIME_SEPARATOR)[0]}_conversations.json`
+			);

 			exportedConversations = selectedConversations;
 			showExportSummary = true;
@@ -5,21 +5,38 @@
 	import { serverStore } from '$lib/stores/server.svelte';
 	import { modelsStore, modelOptions, modelsLoading } from '$lib/stores/models.svelte';
 	import { formatFileSize, formatParameters, formatNumber } from '$lib/utils';
+	import type { ApiLlamaCppServerProps } from '$lib/types';

 	interface Props {
 		open?: boolean;
 		onOpenChange?: (open: boolean) => void;
+		// when set, fetch props from the child process (router mode)
+		modelId?: string | null;
 	}

-	let { open = $bindable(), onOpenChange }: Props = $props();
+	let { open = $bindable(), onOpenChange, modelId = null }: Props = $props();

-	let serverProps = $derived(serverStore.props);
-	let modelName = $derived(modelsStore.singleModelName);
+	let isRouter = $derived(serverStore.isRouterMode);
+
+	// per-model props fetched from the child process
+	let routerModelProps = $state<ApiLlamaCppServerProps | null>(null);
+	let isLoadingRouterProps = $state(false);
+
+	// in router mode use per-model props, otherwise use global props
+	let serverProps = $derived(isRouter && modelId ? routerModelProps : serverStore.props);
+
+	let modelName = $derived(isRouter && modelId ? modelId : modelsStore.singleModelName);
 	let models = $derived(modelOptions());
 	let isLoadingModels = $derived(modelsLoading());

-	// Get the first model for single-model mode display
-	let firstModel = $derived(models[0] ?? null);
+	// in router mode, find the model option matching modelId
+	// in single mode, use the first model as before
+	let firstModel = $derived.by(() => {
+		if (isRouter && modelId) {
+			return models.find((m) => m.model === modelId) ?? null;
+		}
+		return models[0] ?? null;
+	});

 	// Get modalities from modelStore using the model ID from the first model
 	let modalities = $derived.by(() => {
@@ -33,10 +50,31 @@
 			modelsStore.fetch();
 		}
 	});
+
+	// fetch per-model props from child process when dialog opens in router mode
+	$effect(() => {
+		if (open && isRouter && modelId) {
+			isLoadingRouterProps = true;
+			modelsStore
+				.fetchModelProps(modelId)
+				.then((props) => {
+					routerModelProps = props;
+				})
+				.catch(() => {
+					routerModelProps = null;
+				})
+				.finally(() => {
+					isLoadingRouterProps = false;
+				});
+		}
+		if (!open) {
+			routerModelProps = null;
+		}
+	});
 </script>

 <Dialog.Root bind:open {onOpenChange}>
-	<Dialog.Content class="@container z-9999 !max-w-[60rem] max-w-full">
+	<Dialog.Content class="@container z-9999 !max-h-[80dvh] !max-w-[60rem] max-w-full">
 		<style>
 			@container (max-width: 56rem) {
 				.resizable-text-container {
@@ -52,7 +90,7 @@
 		</Dialog.Header>

 		<div class="space-y-6 py-4">
-			{#if isLoadingModels}
+			{#if isLoadingModels || isLoadingRouterProps}
 				<div class="flex items-center justify-center py-8">
 					<div class="text-sm text-muted-foreground">Loading model information...</div>
 				</div>
@@ -212,7 +250,7 @@
 									<Table.Cell class="align-middle font-medium">Chat Template</Table.Cell>

 									<Table.Cell class="py-10">
-										<div class="max-h-120 overflow-y-auto rounded-md bg-muted p-4">
+										<div class="rounded-md bg-muted p-4">
 											<pre
 												class="font-mono text-xs whitespace-pre-wrap">{serverProps.chat_template}</pre>
 										</div>
@@ -6,6 +6,7 @@
 	import { parseHeadersToArray, serializeHeaders } from '$lib/utils';
 	import { UrlProtocol } from '$lib/enums';
 	import { MCP_SERVER_URL_PLACEHOLDER } from '$lib/constants';
+	import { mcpStore } from '$lib/stores/mcp.svelte';

 	interface Props {
 		url: string;
@@ -62,14 +63,33 @@
 		{/if}

 		{#if !isWebSocket && onUseProxyChange}
-			<label class="mt-3 flex cursor-pointer items-center gap-2">
+			<label
+				class="mt-3 flex items-start gap-2"
+				class:cursor-pointer={mcpStore.isProxyAvailable}
+				class:opacity-80={!mcpStore.isProxyAvailable}
+			>
 				<Switch
+					class="mt-1"
 					id="use-proxy-{id}"
 					checked={useProxy}
+					disabled={!mcpStore.isProxyAvailable}
 					onCheckedChange={(checked) => onUseProxyChange?.(checked)}
 				/>

-				<span class="text-xs text-muted-foreground">Use llama-server proxy</span>
+				<span>
+					<span class="text-xs text-muted-foreground">Use llama-server proxy</span>
+
+					<br />
+
+					{#if !mcpStore.isProxyAvailable}
+						<span class="inline-flex gap-0.75 text-xs text-muted-foreground/60"
+							>(Run <pre>llama-server</pre>
+							with
+							<pre>--webui-mcp-proxy</pre>
+							flag)</span
+						>
+					{/if}
+				</span>
 			</label>
 		{/if}
 	</div>
@@ -1,6 +1,5 @@
 <script lang="ts">
 	import { onMount } from 'svelte';
-	import { SvelteMap } from 'svelte/reactivity';
 	import { ChevronDown, Loader2, Package } from '@lucide/svelte';
 	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
 	import * as Tooltip from '$lib/components/ui/tooltip';
@@ -19,9 +18,11 @@
 		DialogModelInformation,
 		DropdownMenuSearchable,
 		ModelId,
+		ModelsSelectorList,
 		ModelsSelectorOption
 	} from '$lib/components/app';
 	import type { ModelOption } from '$lib/types/models';
+	import { filterModelOptions, groupModelOptions, type ModelItem } from './utils';

 	interface Props {
 		class?: string;
@@ -73,89 +74,13 @@
 	let searchTerm = $state('');
 	let highlightedIndex = $state<number>(-1);

-	let filteredOptions: ModelOption[] = $derived.by(() => {
-		const term = searchTerm.trim().toLowerCase();
-		if (!term) return options;
+	let filteredOptions = $derived(filterModelOptions(options, searchTerm));

-		return options.filter(
-			(option) =>
-				option.model.toLowerCase().includes(term) ||
-				option.name?.toLowerCase().includes(term) ||
-				option.aliases?.some((alias: string) => alias.toLowerCase().includes(term)) ||
-				option.tags?.some((tag: string) => tag.toLowerCase().includes(term))
-		);
-	});
-
-	let groupedFilteredOptions = $derived.by(() => {
-		const favIds = modelsStore.favouriteModelIds;
-		const result: {
-			orgName: string | null;
-			isFavouritesGroup: boolean;
-			isLoadedGroup: boolean;
-			items: { option: ModelOption; flatIndex: number }[];
-		}[] = [];
-
-		// Loaded models group (top)
-		const loadedItems: { option: ModelOption; flatIndex: number }[] = [];
-		for (let i = 0; i < filteredOptions.length; i++) {
-			if (modelsStore.isModelLoaded(filteredOptions[i].model)) {
-				loadedItems.push({ option: filteredOptions[i], flatIndex: i });
-			}
-		}
-
-		if (loadedItems.length > 0) {
-			result.push({
-				orgName: null,
-				isFavouritesGroup: false,
-				isLoadedGroup: true,
-				items: loadedItems
-			});
-		}
-
-		// Favourites group
-		const loadedModelIds = new Set(loadedItems.map((item) => item.option.model));
-		const favItems: { option: ModelOption; flatIndex: number }[] = [];
-		for (let i = 0; i < filteredOptions.length; i++) {
-			if (favIds.has(filteredOptions[i].model) && !loadedModelIds.has(filteredOptions[i].model)) {
-				favItems.push({ option: filteredOptions[i], flatIndex: i });
-			}
-		}
-
-		if (favItems.length > 0) {
-			result.push({
-				orgName: null,
-				isFavouritesGroup: true,
-				isLoadedGroup: false,
-				items: favItems
-			});
-		}
-
-		// Org groups (excluding loaded and favourites)
-		const orgGroups = new SvelteMap<string, { option: ModelOption; flatIndex: number }[]>();
-		for (let i = 0; i < filteredOptions.length; i++) {
-			const option = filteredOptions[i];
-
-			if (loadedModelIds.has(option.model) || favIds.has(option.model)) continue;
-
-			const orgName = option.parsedId?.orgName ?? null;
-			const key = orgName ?? '';
-
-			if (!orgGroups.has(key)) orgGroups.set(key, []);
-
-			orgGroups.get(key)!.push({ option, flatIndex: i });
-		}
-
-		for (const [orgName, items] of orgGroups) {
-			result.push({
-				orgName: orgName || null,
-				isFavouritesGroup: false,
-				isLoadedGroup: false,
-				items
-			});
-		}
-
-		return result;
-	});
+	let groupedFilteredOptions = $derived(
+		groupModelOptions(filteredOptions, modelsStore.favouriteModelIds, (m) =>
+			modelsStore.isModelLoaded(m)
+		)
+	);

 	$effect(() => {
 		void searchTerm;
@@ -164,6 +89,12 @@

 	let isOpen = $state(false);
 	let showModelDialog = $state(false);
+	let infoModelId = $state<string | null>(null);
+
+	function handleInfoClick(modelName: string) {
+		infoModelId = modelName;
+		showModelDialog = true;
+	}

 	onMount(() => {
 		modelsStore.fetch().catch((error) => {
@@ -418,45 +349,39 @@
 								<p class="px-4 py-3 text-sm text-muted-foreground">No models found.</p>
 							{/if}

-							{#each groupedFilteredOptions as group (group.isLoadedGroup ? '__loaded__' : group.isFavouritesGroup ? '__favourites__' : group.orgName)}
-								{#if group.isLoadedGroup}
-									<p class="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none">
-										Loaded models
-									</p>
-								{:else if group.isFavouritesGroup}
-									<p class="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none">
-										Favourite models
-									</p>
-								{:else if group.orgName}
-									<p
-										class="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none [&:not(:first-child)]:mt-2"
-									>
-										{group.orgName}
-									</p>
-								{/if}
+							{#snippet modelOption(item: ModelItem, showOrgName: boolean)}
+								{@const { option, flatIndex } = item}
+								{@const isSelected = currentModel === option.model || activeId === option.id}
+								{@const isHighlighted = flatIndex === highlightedIndex}
+								{@const isFav = modelsStore.favouriteModelIds.has(option.model)}

-								{#each group.items as { option, flatIndex } (group.isLoadedGroup ? `loaded-${option.id}` : group.isFavouritesGroup ? `fav-${option.id}` : option.id)}
-									{@const isSelected = currentModel === option.model || activeId === option.id}
-									{@const isHighlighted = flatIndex === highlightedIndex}
-									{@const isFav = modelsStore.favouriteModelIds.has(option.model)}
+								<ModelsSelectorOption
+									{option}
+									{isSelected}
+									{isHighlighted}
+									{isFav}
+									{showOrgName}
+									onSelect={handleSelect}
+									onInfoClick={handleInfoClick}
+									onMouseEnter={() => (highlightedIndex = flatIndex)}
+									onKeyDown={(e) => {
+										if (e.key === KeyboardKey.ENTER || e.key === KeyboardKey.SPACE) {
+											e.preventDefault();
+											handleSelect(option.id);
+										}
+									}}
+								/>
+							{/snippet}

-									<ModelsSelectorOption
-										{option}
-										{isSelected}
-										{isHighlighted}
-										{isFav}
-										showOrgName={group.isFavouritesGroup || group.isLoadedGroup}
-										onSelect={handleSelect}
-										onMouseEnter={() => (highlightedIndex = flatIndex)}
-										onKeyDown={(e) => {
-											if (e.key === KeyboardKey.ENTER || e.key === KeyboardKey.SPACE) {
-												e.preventDefault();
-												handleSelect(option.id);
-											}
-										}}
-									/>
-								{/each}
-							{/each}
+							<ModelsSelectorList
+								groups={groupedFilteredOptions}
+								{currentModel}
+								{activeId}
+								sectionHeaderClass="my-1.5 px-2 py-2 text-[13px] font-semibold text-muted-foreground/70 select-none"
+								onSelect={handleSelect}
+								onInfoClick={handleInfoClick}
+								renderOption={modelOption}
+							/>
 						</div>
 					</DropdownMenuSearchable>
 				</DropdownMenu.Content>
@@ -500,6 +425,6 @@
 	{/if}
 </div>

-{#if showModelDialog && !isRouter}
-	<DialogModelInformation bind:open={showModelDialog} />
+{#if showModelDialog}
+	<DialogModelInformation bind:open={showModelDialog} modelId={infoModelId} />
 {/if}
@@ -0,0 +1,72 @@
+<script lang="ts">
+	import { modelsStore } from '$lib/stores/models.svelte';
+	import { ModelsSelectorOption } from '$lib/components/app';
+	import type { GroupedModelOptions, ModelItem } from './utils';
+
+	interface Props {
+		groups: GroupedModelOptions;
+		currentModel: string | null;
+		activeId: string | null;
+		sectionHeaderClass?: string;
+		orgHeaderClass?: string;
+		onSelect: (modelId: string) => void;
+		onInfoClick: (modelName: string) => void;
+		renderOption?: import('svelte').Snippet<[ModelItem, boolean]>;
+	}
+
+	let {
+		groups,
+		currentModel,
+		activeId,
+		sectionHeaderClass = 'my-1 px-2 py-2 text-[13px] font-semibold text-muted-foreground/70 select-none',
+		orgHeaderClass = 'px-2 py-2 text-[11px] font-semibold text-muted-foreground/50 select-none [&:not(:first-child)]:mt-1',
+		onSelect,
+		onInfoClick,
+		renderOption
+	}: Props = $props();
+	let render = $derived(renderOption ?? defaultOption);
+</script>
+
+{#snippet defaultOption(item: ModelItem, showOrgName: boolean)}
+	{@const { option } = item}
+	{@const isSelected = currentModel === option.model || activeId === option.id}
+	{@const isFav = modelsStore.favouriteModelIds.has(option.model)}
+
+	<ModelsSelectorOption
+		{option}
+		{isSelected}
+		isHighlighted={false}
+		{isFav}
+		{showOrgName}
+		{onSelect}
+		{onInfoClick}
+		onMouseEnter={() => {}}
+		onKeyDown={() => {}}
+	/>
+{/snippet}
+
+{#if groups.loaded.length > 0}
+	<p class={sectionHeaderClass}>Loaded models</p>
+	{#each groups.loaded as item (`loaded-${item.option.id}`)}
+		{@render render(item, true)}
+	{/each}
+{/if}
+
+{#if groups.favourites.length > 0}
+	<p class={sectionHeaderClass}>Favourite models</p>
+	{#each groups.favourites as item (`fav-${item.option.id}`)}
+		{@render render(item, true)}
+	{/each}
+{/if}
+
+{#if groups.available.length > 0}
+	<p class={sectionHeaderClass}>Available models</p>
+	{#each groups.available as group (group.orgName)}
+		{#if group.orgName}
+			<p class={orgHeaderClass}>{group.orgName}</p>
+		{/if}
+		{#each group.items as item (item.option.id)}
+			{@render render(item, false)}
+		{/each}
+	{/each}
+{/if}
@@ -1,5 +1,14 @@
 <script lang="ts">
-	import { CircleAlert, Heart, HeartOff, Loader2, Power, PowerOff, RotateCw } from '@lucide/svelte';
+	import {
+		CircleAlert,
+		Heart,
+		HeartOff,
+		Info,
+		Loader2,
+		Power,
+		PowerOff,
+		RotateCw
+	} from '@lucide/svelte';
 	import { cn } from '$lib/components/ui/utils';
 	import { ActionIcon, ModelId } from '$lib/components/app';
 	import type { ModelOption } from '$lib/types/models';
@@ -15,6 +24,7 @@
 		onSelect: (modelId: string) => void;
 		onMouseEnter: () => void;
 		onKeyDown: (e: KeyboardEvent) => void;
+		onInfoClick?: (modelName: string) => void;
 	}

 	let {
@@ -25,7 +35,8 @@
 		showOrgName = false,
 		onSelect,
 		onMouseEnter,
-		onKeyDown
+		onKeyDown,
+		onInfoClick
 	}: Props = $props();

 	let currentRouterModels = $derived(routerModels());
@@ -63,11 +74,11 @@
 		class="flex-1"
 	/>

-	<div class="flex shrink-0 items-center gap-2.5">
+	<div class="flex shrink-0 items-center gap-1">
 		<!-- svelte-ignore a11y_no_static_element_interactions -->
 		<!-- svelte-ignore a11y_click_events_have_key_events -->
 		<div
-			class="pointer-events-none flex w-4 items-center justify-center pl-2 opacity-0 group-hover:pointer-events-auto group-hover:opacity-100"
+			class="pointer-events-none flex items-center justify-center gap-0.75 pl-2 opacity-0 group-hover:pointer-events-auto group-hover:opacity-100"
 			onclick={(e) => e.stopPropagation()}
 		>
 			{#if isFav}
@@ -87,7 +98,19 @@
 					onclick={() => modelsStore.toggleFavourite(option.model)}
 				/>
 			{/if}
+
+			<!-- info button: only shown when model is loaded and callback is provided -->
+			{#if isLoaded && onInfoClick}
+				<ActionIcon
+					iconSize="h-2.5 w-2.5"
+					icon={Info}
+					tooltip="Model information"
+					class="h-3 w-3 hover:text-foreground"
+					onclick={() => onInfoClick(option.model)}
+				/>
+			{/if}
 		</div>
+
 		{#if isLoading}
 			<Loader2 class="h-4 w-4 animate-spin text-muted-foreground" />
 		{:else if isFailed}
@@ -1,6 +1,5 @@
 <script lang="ts">
 	import { onMount } from 'svelte';
-	import { SvelteMap } from 'svelte/reactivity';
 	import { ChevronDown, Loader2, Package } from '@lucide/svelte';
 	import * as Sheet from '$lib/components/ui/sheet';
 	import { cn } from '$lib/components/ui/utils';
@@ -15,11 +14,12 @@
 	import { isRouterMode } from '$lib/stores/server.svelte';
 	import {
 		DialogModelInformation,
+		ModelsSelectorList,
 		SearchInput,
-		TruncatedText,
-		ModelsSelectorOption
+		TruncatedText
 	} from '$lib/components/app';
 	import type { ModelOption } from '$lib/types/models';
+	import { filterModelOptions, groupModelOptions } from './utils';

 	interface Props {
 		class?: string;
@@ -73,85 +73,22 @@

 	let searchTerm = $state('');

-	let filteredOptions: ModelOption[] = $derived.by(() => {
-		const term = searchTerm.trim().toLowerCase();
-		if (!term) return options;
+	let filteredOptions = $derived(filterModelOptions(options, searchTerm));

-		return options.filter(
-			(option) =>
-				option.model.toLowerCase().includes(term) ||
-				option.name?.toLowerCase().includes(term) ||
-				option.aliases?.some((alias: string) => alias.toLowerCase().includes(term)) ||
-				option.tags?.some((tag: string) => tag.toLowerCase().includes(term))
-		);
-	});
-
-	let groupedFilteredOptions = $derived.by(() => {
-		const favIds = modelsStore.favouriteModelIds;
-		const result: {
-			orgName: string | null;
-			isFavouritesGroup: boolean;
-			isLoadedGroup: boolean;
-			items: { option: ModelOption; flatIndex: number }[];
-		}[] = [];
-
-		// Loaded models group (top)
-		const loadedItems: { option: ModelOption; flatIndex: number }[] = [];
-		for (let i = 0; i < filteredOptions.length; i++) {
-			if (modelsStore.isModelLoaded(filteredOptions[i].model)) {
-				loadedItems.push({ option: filteredOptions[i], flatIndex: i });
-			}
-		}
-		if (loadedItems.length > 0) {
-			result.push({
-				orgName: null,
-				isFavouritesGroup: false,
-				isLoadedGroup: true,
-				items: loadedItems
-			});
-		}
-
-		// Favourites group
-		const loadedModelIds = new Set(loadedItems.map((item) => item.option.model));
-		const favItems: { option: ModelOption; flatIndex: number }[] = [];
-		for (let i = 0; i < filteredOptions.length; i++) {
-			if (favIds.has(filteredOptions[i].model) && !loadedModelIds.has(filteredOptions[i].model)) {
-				favItems.push({ option: filteredOptions[i], flatIndex: i });
-			}
-		}
-		if (favItems.length > 0) {
-			result.push({
-				orgName: null,
-				isFavouritesGroup: true,
-				isLoadedGroup: false,
-				items: favItems
-			});
-		}
-
-		// Org groups (excluding loaded and favourites)
-		const orgGroups = new SvelteMap<string, { option: ModelOption; flatIndex: number }[]>();
-		for (let i = 0; i < filteredOptions.length; i++) {
-			const option = filteredOptions[i];
-			if (loadedModelIds.has(option.model) || favIds.has(option.model)) continue;
-			const orgName = option.parsedId?.orgName ?? null;
-			const key = orgName ?? '';
-			if (!orgGroups.has(key)) orgGroups.set(key, []);
-			orgGroups.get(key)!.push({ option, flatIndex: i });
-		}
-		for (const [orgName, items] of orgGroups) {
-			result.push({
-				orgName: orgName || null,
-				isFavouritesGroup: false,
-				isLoadedGroup: false,
-				items
-			});
-		}
-
-		return result;
-	});
+	let groupedFilteredOptions = $derived(
+		groupModelOptions(filteredOptions, modelsStore.favouriteModelIds, (m) =>
+			modelsStore.isModelLoaded(m)
+		)
+	);

 	let sheetOpen = $state(false);
 	let showModelDialog = $state(false);
+	let infoModelId = $state<string | null>(null);
+
+	function handleInfoClick(modelName: string) {
+		infoModelId = modelName;
+		showModelDialog = true;
+	}

 	onMount(() => {
 		modelsStore.fetch().catch((error) => {
@@ -339,38 +276,15 @@
 								<p class="px-3 py-3 text-center text-sm text-muted-foreground">No models found.</p>
 							{/if}

-							{#each groupedFilteredOptions as group (group.isLoadedGroup ? '__loaded__' : group.isFavouritesGroup ? '__favourites__' : group.orgName)}
-								{#if group.isLoadedGroup}
-									<p class="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none">
-										Loaded models
-									</p>
-								{:else if group.isFavouritesGroup}
-									<p class="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none">
-										Favourite models
-									</p>
-								{:else if group.orgName}
-									<p
-										class="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none [&:not(:first-child)]:mt-2"
-									>
-										{group.orgName}
-									</p>
-								{/if}
-
-								{#each group.items as { option } (group.isLoadedGroup ? `loaded-${option.id}` : group.isFavouritesGroup ? `fav-${option.id}` : option.id)}
-									{@const isSelected = currentModel === option.model || activeId === option.id}
-									{@const isFav = modelsStore.favouriteModelIds.has(option.model)}
-									<ModelsSelectorOption
-										{option}
-										{isSelected}
-										isHighlighted={false}
-										{isFav}
-										showOrgName={group.isFavouritesGroup || group.isLoadedGroup}
-										onSelect={handleSelect}
-										onMouseEnter={() => {}}
-										onKeyDown={() => {}}
-									/>
-								{/each}
-							{/each}
+							<ModelsSelectorList
+								groups={groupedFilteredOptions}
+								{currentModel}
+								{activeId}
+								sectionHeaderClass="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none"
+								orgHeaderClass="px-2 py-2 text-xs font-semibold text-muted-foreground/60 select-none [&:not(:first-child)]:mt-2"
+								onSelect={handleSelect}
+								onInfoClick={handleInfoClick}
+							/>
 						</div>
 					</div>
 				</Sheet.Content>
@@ -403,6 +317,6 @@
 	{/if}
 </div>

-{#if showModelDialog && !isRouter}
-	<DialogModelInformation bind:open={showModelDialog} />
+{#if showModelDialog}
+	<DialogModelInformation bind:open={showModelDialog} modelId={infoModelId} />
 {/if}
@@ -44,6 +44,27 @@
 */
 export { default as ModelsSelector } from './ModelsSelector.svelte';

+/**
+ * **ModelsSelectorList** - Grouped model options list
+ *
+ * Renders grouped model options (loaded, favourites, available) with section
+ * headers and org subgroups. Shared between ModelsSelector and ModelsSelectorSheet
+ * to avoid template duplication.
+ *
+ * Accepts an optional `renderOption` snippet to customize how each option is
+ * rendered (e.g., to add keyboard navigation or highlighting).
+ */
+export { default as ModelsSelectorList } from './ModelsSelectorList.svelte';
+
+/**
+ * **ModelsSelectorOption** - Single model option row
+ *
+ * Renders a single model option with selection state, favourite toggle,
+ * load/unload actions, status indicators, and an info button.
+ * Used inside ModelsSelectorList or directly in custom render snippets.
+ */
+export { default as ModelsSelectorOption } from './ModelsSelectorOption.svelte';
+
 /**
 * **ModelsSelectorSheet** - Mobile model selection sheet
 *
@@ -80,5 +101,12 @@ export { default as ModelsSelectorSheet } from './ModelsSelectorSheet.svelte';
 * ```
 */
 export { default as ModelBadge } from './ModelBadge.svelte';
+
+/**
+ * **ModelId** - Parsed model identifier display
+ *
+ * Displays a model ID with optional org name, parameter badges, quantization,
+ * aliases, and tags. Supports raw mode to show the unprocessed model name.
+ * Respects the user's `showRawModelNames` setting.
+ */
 export { default as ModelId } from './ModelId.svelte';
-export { default as ModelsSelectorOption } from './ModelsSelectorOption.svelte';
@@ -0,0 +1,75 @@
+import { SvelteMap } from 'svelte/reactivity';
+import type { ModelOption } from '$lib/types/models';
+
+export interface ModelItem {
+	option: ModelOption;
+	flatIndex: number;
+}
+
+export interface OrgGroup {
+	orgName: string | null;
+	items: ModelItem[];
+}
+
+export interface GroupedModelOptions {
+	loaded: ModelItem[];
+	favourites: ModelItem[];
+	available: OrgGroup[];
+}
+
+export function filterModelOptions(options: ModelOption[], searchTerm: string): ModelOption[] {
+	const term = searchTerm.trim().toLowerCase();
+	if (!term) return options;
+
+	return options.filter(
+		(option) =>
+			option.model.toLowerCase().includes(term) ||
+			option.name?.toLowerCase().includes(term) ||
+			option.aliases?.some((alias: string) => alias.toLowerCase().includes(term)) ||
+			option.tags?.some((tag: string) => tag.toLowerCase().includes(term))
+	);
+}
+
+export function groupModelOptions(
+	filteredOptions: ModelOption[],
+	favouriteIds: Set<string>,
+	isModelLoaded: (model: string) => boolean
+): GroupedModelOptions {
+	// Loaded models
+	const loaded: ModelItem[] = [];
+	for (let i = 0; i < filteredOptions.length; i++) {
+		if (isModelLoaded(filteredOptions[i].model)) {
+			loaded.push({ option: filteredOptions[i], flatIndex: i });
+		}
+	}
+
+	// Favourites (excluding loaded)
+	const loadedModelIds = new Set(loaded.map((item) => item.option.model));
+	const favourites: ModelItem[] = [];
+	for (let i = 0; i < filteredOptions.length; i++) {
+		if (
+			favouriteIds.has(filteredOptions[i].model) &&
+			!loadedModelIds.has(filteredOptions[i].model)
+		) {
+			favourites.push({ option: filteredOptions[i], flatIndex: i });
+		}
+	}
+
+	// Available models grouped by org (excluding loaded and favourites)
+	const available: OrgGroup[] = [];
+	const orgGroups = new SvelteMap<string, ModelItem[]>();
+	for (let i = 0; i < filteredOptions.length; i++) {
+		const option = filteredOptions[i];
+		if (loadedModelIds.has(option.model) || favouriteIds.has(option.model)) continue;
+
+		const key = option.parsedId?.orgName ?? '';
+		if (!orgGroups.has(key)) orgGroups.set(key, []);
+		orgGroups.get(key)!.push({ option, flatIndex: i });
+	}
+
+	for (const [orgName, items] of orgGroups) {
+		available.push({ orgName: orgName || null, items });
+	}
+
+	return { loaded, favourites, available };
+}
@@ -24,6 +24,7 @@ export * from './max-bundle-size';
 export * from './mcp';
 export * from './mcp-form';
 export * from './mcp-resource';
+export * from './message-export';
 export * from './model-id';
 export * from './precision';
 export * from './processing-info';
@@ -0,0 +1,20 @@
+// Conversation filename constants
+
+// Length of the trimmed conversation ID in the filename
+export const EXPORT_CONV_ID_TRIM_LENGTH = 8;
+// Maximum length of the sanitized conversation name snippet
+export const EXPORT_CONV_NAME_SUFFIX_MAX_LENGTH = 20;
+// Characters to keep in the ISO timestamp. 19 keeps 2026-01-01T00:00:00
+export const ISO_TIMESTAMP_SLICE_LENGTH = 19;
+
+// Replacements for making the conversation title filename-friendly
+export const NON_ALPHANUMERIC_REGEX = /[^a-z0-9]/gi;
+export const EXPORT_CONV_NONALNUM_REPLACEMENT = '_';
+export const MULTIPLE_UNDERSCORE_REGEX = /_+/g;
+
+// Replacements to the ISO date for use in the export filename
+export const ISO_DATE_TIME_SEPARATOR = 'T';
+export const ISO_DATE_TIME_SEPARATOR_REPLACEMENT = '_';
+
+export const ISO_TIME_SEPARATOR = ':';
+export const ISO_TIME_SEPARATOR_REPLACEMENT = '-';
@@ -26,6 +26,18 @@ import { config } from '$lib/stores/settings.svelte';
 import { filterByLeafNodeId, findLeafNode } from '$lib/utils';
 import type { McpServerOverride } from '$lib/types/database';
 import { MessageRole } from '$lib/enums';
+import {
+	ISO_DATE_TIME_SEPARATOR,
+	ISO_DATE_TIME_SEPARATOR_REPLACEMENT,
+	ISO_TIMESTAMP_SLICE_LENGTH,
+	EXPORT_CONV_ID_TRIM_LENGTH,
+	EXPORT_CONV_NONALNUM_REPLACEMENT,
+	EXPORT_CONV_NAME_SUFFIX_MAX_LENGTH,
+	ISO_TIME_SEPARATOR,
+	ISO_TIME_SEPARATOR_REPLACEMENT,
+	NON_ALPHANUMERIC_REGEX,
+	MULTIPLE_UNDERSCORE_REGEX
+} from '$lib/constants';

 class ConversationsStore {
 	/**
@@ -619,6 +631,66 @@ class ConversationsStore {
 	 *
 	 */

+	/**
+	 * Generates a sanitized filename for a conversation export
+	 * @param conversation - The conversation metadata
+	 * @param msgs - Optional array of messages belonging to the conversation
+	 * @returns The generated filename string
+	 */
+	generateConversationFilename(
+		conversation: { id?: string; name?: string },
+		msgs?: DatabaseMessage[]
+	): string {
+		const conversationName = (conversation.name ?? '').trim().toLowerCase();
+
+		const sanitizedName = conversationName
+			.replace(NON_ALPHANUMERIC_REGEX, EXPORT_CONV_NONALNUM_REPLACEMENT)
+			.replace(MULTIPLE_UNDERSCORE_REGEX, '_')
+			.substring(0, EXPORT_CONV_NAME_SUFFIX_MAX_LENGTH);
+
+		// If we have messages, use the timestamp of the newest message
+		const referenceDate = msgs?.length
+			? new Date(Math.max(...msgs.map((m) => m.timestamp)))
+			: new Date();
+
+		const iso = referenceDate.toISOString().slice(0, ISO_TIMESTAMP_SLICE_LENGTH);
+		const formattedDate = iso
+			.replace(ISO_DATE_TIME_SEPARATOR, ISO_DATE_TIME_SEPARATOR_REPLACEMENT)
+			.replaceAll(ISO_TIME_SEPARATOR, ISO_TIME_SEPARATOR_REPLACEMENT);
+		const trimmedConvId = conversation.id?.slice(0, EXPORT_CONV_ID_TRIM_LENGTH) ?? '';
+		return `${formattedDate}_conv_${trimmedConvId}_${sanitizedName}.json`;
+	}
+
+	/**
+	 * Triggers a browser download of the provided exported conversation data
+	 * @param data - The exported conversation payload (either a single conversation or array of them)
+	 * @param filename - Filename; if omitted, a deterministic name is generated
+	 */
+	downloadConversationFile(data: ExportedConversations, filename?: string): void {
+		// Choose the first conversation or message
+		const conversation =
+			'conv' in data ? data.conv : Array.isArray(data) ? data[0]?.conv : undefined;
+		const msgs =
+			'messages' in data ? data.messages : Array.isArray(data) ? data[0]?.messages : undefined;
+
+		if (!conversation) {
+			console.error('Invalid data: missing conversation');
+			return;
+		}
+
+		const downloadFilename = filename ?? this.generateConversationFilename(conversation, msgs);
+
+		const blob = new Blob([JSON.stringify(data, null, 2)], { type: 'application/json' });
+		const url = URL.createObjectURL(blob);
+		const a = document.createElement('a');
+		a.href = url;
+		a.download = downloadFilename;
+		document.body.appendChild(a);
+		a.click();
+		document.body.removeChild(a);
+		URL.revokeObjectURL(url);
+	}
+
 	/**
 	 * Downloads a conversation as JSON file.
 	 * @param convId - The conversation ID to download
@@ -636,40 +708,7 @@ class ConversationsStore {
 			messages = await DatabaseService.getConversationMessages(convId);
 		}

-		this.triggerDownload({ conv: conversation, messages });
-	}
-
-	/**
-	 * Exports all conversations with their messages as a JSON file
-	 * @returns The list of exported conversations
-	 */
-	async exportAllConversations(): Promise<DatabaseConversation[]> {
-		const allConversations = await DatabaseService.getAllConversations();
-
-		if (allConversations.length === 0) {
-			throw new Error('No conversations to export');
-		}
-
-		const allData = await Promise.all(
-			allConversations.map(async (conv) => {
-				const messages = await DatabaseService.getConversationMessages(conv.id);
-				return { conv, messages };
-			})
-		);
-
-		const blob = new Blob([JSON.stringify(allData, null, 2)], { type: 'application/json' });
-		const url = URL.createObjectURL(blob);
-		const a = document.createElement('a');
-		a.href = url;
-		a.download = `all_conversations_${new Date().toISOString().split('T')[0]}.json`;
-		document.body.appendChild(a);
-		a.click();
-		document.body.removeChild(a);
-		URL.revokeObjectURL(url);
-
-		toast.success(`All conversations (${allConversations.length}) prepared for download`);
-
-		return allConversations;
+		this.downloadConversationFile({ conv: conversation, messages });
 	}

 	/**
@@ -743,37 +782,6 @@ class ConversationsStore {
 		await this.loadConversations();
 		return result;
 	}
-
-	/**
-	 * Triggers file download in browser
-	 */
-	private triggerDownload(data: ExportedConversations, filename?: string): void {
-		const conversation =
-			'conv' in data ? data.conv : Array.isArray(data) ? data[0]?.conv : undefined;
-
-		if (!conversation) {
-			console.error('Invalid data: missing conversation');
-			return;
-		}
-
-		const conversationName = conversation.name?.trim() || '';
-		const truncatedSuffix = conversationName
-			.toLowerCase()
-			.replace(/[^a-z0-9]/gi, '_')
-			.replace(/_+/g, '_')
-			.substring(0, 20);
-		const downloadFilename = filename || `conversation_${conversation.id}_${truncatedSuffix}.json`;
-
-		const blob = new Blob([JSON.stringify(data, null, 2)], { type: 'application/json' });
-		const url = URL.createObjectURL(blob);
-		const a = document.createElement('a');
-		a.href = url;
-		a.download = downloadFilename;
-		document.body.appendChild(a);
-		a.click();
-		document.body.removeChild(a);
-		URL.revokeObjectURL(url);
-	}
 }

 export const conversationsStore = new ConversationsStore();
@@ -20,6 +20,7 @@
 */

 import { browser } from '$app/environment';
+import { base } from '$app/paths';
 import { MCPService } from '$lib/services/mcp.service';
 import { config, settingsStore } from '$lib/stores/settings.svelte';
 import { mcpResourceStore } from '$lib/stores/mcp-resources.svelte';
@@ -42,6 +43,7 @@ import {
 	ToolCallType
 } from '$lib/enums';
 import {
+	CORS_PROXY_ENDPOINT,
 	DEFAULT_CACHE_TTL_MS,
 	DEFAULT_MCP_CONFIG,
 	EXPECTED_THEMED_ICON_PAIR_COUNT,
@@ -78,165 +80,13 @@ import type { ListChangedHandlers } from '@modelcontextprotocol/sdk/types.js';
 import type { DatabaseMessageExtraMcpResource, McpServerOverride } from '$lib/types/database';
 import type { SettingsConfigType } from '$lib/types/settings';

-export function buildMcpClientConfig(
-	cfg: SettingsConfigType,
-	perChatOverrides?: McpServerOverride[]
-): MCPClientConfig | undefined {
-	return buildMcpClientConfigInternal(cfg, perChatOverrides);
-}
-
-/**
- * Internal helper to build MCP client config.
- * Kept as standalone function for external use and tests.
- */
-export function buildMcpClientConfigInternal(
-	cfg: SettingsConfigType,
-	perChatOverrides?: McpServerOverride[]
-): MCPClientConfig | undefined {
-	const rawServers = parseServerSettings(cfg.mcpServers);
-	if (!rawServers.length) {
-		return undefined;
-	}
-
-	const servers: Record<string, MCPServerConfig> = {};
-
-	for (const [index, entry] of rawServers.entries()) {
-		if (!checkServerEnabled(entry, perChatOverrides)) continue;
-		const normalized = buildServerConfig(entry);
-		if (normalized) servers[generateMcpServerId(entry.id, index)] = normalized;
-	}
-
-	if (Object.keys(servers).length === 0) {
-		return undefined;
-	}
-
-	return {
-		protocolVersion: DEFAULT_MCP_CONFIG.protocolVersion,
-		capabilities: DEFAULT_MCP_CONFIG.capabilities,
-		clientInfo: DEFAULT_MCP_CONFIG.clientInfo,
-		requestTimeoutMs: Math.round(DEFAULT_MCP_CONFIG.requestTimeoutSeconds * 1000),
-		servers
-	};
-}
-
-/**
- * Generates a unique server ID from an optional ID string or index.
- * @deprecated Use MCPStore.#generateServerId instead
- */
-function generateMcpServerId(id: unknown, index: number): string {
-	if (typeof id === 'string' && id.trim()) {
-		return id.trim();
-	}
-
-	return `${MCP_SERVER_ID_PREFIX}-${index + 1}`;
-}
-
-/**
- * Parses raw server settings from config into MCPServerSettingsEntry array.
- * @deprecated Use MCPStore.#parseServerSettings instead
- */
-function parseServerSettings(rawServers: unknown): MCPServerSettingsEntry[] {
-	if (!rawServers) {
-		return [];
-	}
-
-	let parsed: unknown;
-	if (typeof rawServers === 'string') {
-		const trimmed = rawServers.trim();
-		if (!trimmed) {
-			return [];
-		}
-
-		try {
-			parsed = JSON.parse(trimmed);
-		} catch (error) {
-			console.warn('[MCP] Failed to parse mcpServers JSON:', error);
-
-			return [];
-		}
-	} else {
-		parsed = rawServers;
-	}
-	if (!Array.isArray(parsed)) {
-		return [];
-	}
-
-	return parsed.map((entry, index) => {
-		const url = typeof entry?.url === 'string' ? entry.url.trim() : '';
-		const headers = typeof entry?.headers === 'string' ? entry.headers.trim() : undefined;
-
-		return {
-			id: generateMcpServerId((entry as { id?: unknown })?.id, index),
-			enabled: Boolean((entry as { enabled?: unknown })?.enabled),
-			url,
-			name: (entry as { name?: string })?.name,
-			requestTimeoutSeconds: DEFAULT_MCP_CONFIG.requestTimeoutSeconds,
-			headers: headers || undefined,
-			useProxy: Boolean((entry as { useProxy?: unknown })?.useProxy)
-		} satisfies MCPServerSettingsEntry;
-	});
-}
-
-/**
- * Builds server configuration from a settings entry.
- * @deprecated Use MCPStore.#buildServerConfig instead
- */
-function buildServerConfig(
-	entry: MCPServerSettingsEntry,
-	connectionTimeoutMs = DEFAULT_MCP_CONFIG.connectionTimeoutMs
-): MCPServerConfig | undefined {
-	if (!entry?.url) {
-		return undefined;
-	}
-
-	let headers: Record<string, string> | undefined;
-	if (entry.headers) {
-		try {
-			const parsed = JSON.parse(entry.headers);
-			if (typeof parsed === 'object' && parsed !== null && !Array.isArray(parsed))
-				headers = parsed as Record<string, string>;
-		} catch {
-			console.warn('[MCP] Failed to parse custom headers JSON:', entry.headers);
-		}
-	}
-
-	return {
-		url: entry.url,
-		transport: detectMcpTransportFromUrl(entry.url),
-		handshakeTimeoutMs: connectionTimeoutMs,
-		requestTimeoutMs: Math.round(entry.requestTimeoutSeconds * 1000),
-		headers,
-		useProxy: entry.useProxy
-	};
-}
-
-/**
- * Checks if a server is enabled, considering per-chat overrides.
- * @deprecated Use MCPStore.#checkServerEnabled instead
- */
-function checkServerEnabled(
-	server: MCPServerSettingsEntry,
-	perChatOverrides?: McpServerOverride[]
-): boolean {
-	if (!server.enabled) {
-		return false;
-	}
-
-	if (perChatOverrides) {
-		const override = perChatOverrides.find((o) => o.serverId === server.id);
-
-		return override?.enabled ?? false;
-	}
-
-	return false;
-}
-
 class MCPStore {
 	private _isInitializing = $state(false);
 	private _error = $state<string | null>(null);
 	private _toolCount = $state(0);
 	private _connectedServers = $state<string[]>([]);
 	private _healthChecks = $state<Record<string, HealthCheckState>>({});
+	private _proxyAvailable = $state(false);

 	private connections = new Map<string, MCPConnection>();
 	private toolsIndex = new Map<string, string>();
@@ -246,6 +96,29 @@ class MCPStore {
 	private initPromise: Promise<boolean> | null = null;
 	private activeFlowCount = 0;

+	constructor() {
+		if (browser) {
+			this.probeProxy();
+		}
+	}
+
+	/**
+	 * Probes the CORS proxy endpoint to determine availability.
+	 * The endpoint is only registered when llama-server runs with --webui-mcp-proxy.
+	 */
+	async probeProxy(): Promise<void> {
+		try {
+			const response = await fetch(`${base}${CORS_PROXY_ENDPOINT}`, { method: 'HEAD' });
+			this._proxyAvailable = response.status !== 404;
+		} catch {
+			this._proxyAvailable = false;
+		}
+	}
+
+	get isProxyAvailable(): boolean {
+		return this._proxyAvailable;
+	}
+
 	/**
 	 * Generates a unique server ID from an optional ID string or index.
 	 */
@@ -520,6 +393,7 @@ class MCPStore {

 	getServerLabel(server: MCPServerSettingsEntry): string {
 		const healthState = this.getHealthCheckState(server.id);
+
 		if (healthState?.status === HealthCheckStatus.SUCCESS)
 			return (
 				healthState.serverInfo?.title || healthState.serverInfo?.name || server.name || server.url
@@ -603,6 +477,7 @@ class MCPStore {
 	 */
 	#proxyIconSrc(src: string): string {
 		if (src.startsWith('data:')) return src;
+		if (!this._proxyAvailable) return src;

 		return getProxiedUrlString(src);
 	}
@@ -629,7 +504,7 @@ class MCPStore {
 			}
 		}

-		return getFaviconUrl(server.url);
+		return getFaviconUrl(server.url, this._proxyAvailable);
 	}

 	isAnyServerLoading(): boolean {
@@ -2072,6 +1947,7 @@ export const mcpIsInitializing = () => mcpStore.isInitializing;
 export const mcpIsInitialized = () => mcpStore.isInitialized;
 export const mcpError = () => mcpStore.error;
 export const mcpIsEnabled = () => mcpStore.isEnabled;
+export const mcpIsProxyAvailable = () => mcpStore.isProxyAvailable;
 export const mcpAvailableTools = () => mcpStore.availableTools;
 export const mcpConnectedServerCount = () => mcpStore.connectedServerCount;
 export const mcpConnectedServerNames = () => mcpStore.connectedServerNames;
@@ -1,6 +1,7 @@
 /**
 * Utility functions for conversation data manipulation
 */
+import type { DatabaseMessage } from '$lib/types';

 /**
 * Creates a map of conversation IDs to their message counts from exported conversation data
@@ -17,7 +17,7 @@ import {
 * @param urlString - The URL to get the favicon for
 * @returns The favicon URL or null if invalid
 */
-export function getFaviconUrl(urlString: string): string | null {
+export function getFaviconUrl(urlString: string, useProxy = true): string | null {
 	try {
 		const url = new URL(urlString);
 		const hostnameParts = url.hostname.split(DOMAIN_SEPARATOR);
@@ -27,7 +27,7 @@ export function getFaviconUrl(urlString: string): string | null {
 				: url.hostname;

 		const googleFaviconUrl = `${GOOGLE_FAVICON_BASE_URL}?domain=${rootDomain}&sz=${DEFAULT_FAVICON_SIZE}`;
-		return getProxiedUrlString(googleFaviconUrl);
+		return useProxy ? getProxiedUrlString(googleFaviconUrl) : googleFaviconUrl;
 	} catch {
 		return null;
 	}
@@ -231,7 +231,7 @@
 				<Sidebar.Trigger
 					class="transition-left absolute left-0 z-[900] duration-200 ease-linear {sidebarOpen
 						? 'md:left-[var(--sidebar-width)]'
-						: ''}"
+						: 'md:left-0!'}"
 					style="translate: 1rem 1rem;"
 				/>
 			{/if}
@@ -1025,6 +1025,30 @@ bool is_valid_path(const std::string &path) {
  return true;
 }

+bool canonicalize_path(const char *path, std::string &resolved) {
+#if defined(_WIN32)
+  char buf[_MAX_PATH];
+  if (_fullpath(buf, path, _MAX_PATH) == nullptr) { return false; }
+  resolved = buf;
+#else
+  char buf[PATH_MAX];
+  if (realpath(path, buf) == nullptr) { return false; }
+  resolved = buf;
+#endif
+  return true;
+}
+
+bool is_path_within_base(const std::string &resolved_path,
+                                const std::string &resolved_base) {
+#if defined(_WIN32)
+  return _strnicmp(resolved_path.c_str(), resolved_base.c_str(),
+                   resolved_base.size()) == 0;
+#else
+  return strncmp(resolved_path.c_str(), resolved_base.c_str(),
+                 resolved_base.size()) == 0;
+#endif
+}
+
 FileStat::FileStat(const std::string &path) {
 #if defined(_WIN32)
  auto wpath = u8string_to_wstring(path.c_str());
@@ -2627,33 +2651,114 @@ bool can_compress_content_type(const std::string &content_type) {
  }
 }

+bool parse_quality(const char *b, const char *e, std::string &token,
+                          double &quality) {
+  quality = 1.0;
+  token.clear();
+
+  // Split on first ';': left = token name, right = parameters
+  const char *params_b = nullptr;
+  std::size_t params_len = 0;
+
+  divide(
+      b, static_cast<std::size_t>(e - b), ';',
+      [&](const char *lb, std::size_t llen, const char *rb, std::size_t rlen) {
+        auto r = trim(lb, lb + llen, 0, llen);
+        if (r.first < r.second) { token.assign(lb + r.first, lb + r.second); }
+        params_b = rb;
+        params_len = rlen;
+      });
+
+  if (token.empty()) { return false; }
+  if (params_len == 0) { return true; }
+
+  // Scan parameters for q= (stops on first match)
+  bool invalid = false;
+  split_find(params_b, params_b + params_len, ';',
+             (std::numeric_limits<size_t>::max)(),
+             [&](const char *pb, const char *pe) -> bool {
+               // Match exactly "q=" or "Q=" (not "query=" etc.)
+               auto len = static_cast<size_t>(pe - pb);
+               if (len < 2) { return false; }
+               if ((pb[0] != 'q' && pb[0] != 'Q') || pb[1] != '=') {
+                 return false;
+               }
+
+               // Trim the value portion
+               auto r = trim(pb, pe, 2, len);
+               if (r.first >= r.second) {
+                 invalid = true;
+                 return true;
+               }
+
+               double v = 0.0;
+               auto res = from_chars(pb + r.first, pb + r.second, v);
+               if (res.ec != std::errc{} || v < 0.0 || v > 1.0) {
+                 invalid = true;
+                 return true;
+               }
+               quality = v;
+               return true;
+             });
+
+  return !invalid;
+}
+
 EncodingType encoding_type(const Request &req, const Response &res) {
-  auto ret =
-      detail::can_compress_content_type(res.get_header_value("Content-Type"));
-  if (!ret) { return EncodingType::None; }
+  if (!can_compress_content_type(res.get_header_value("Content-Type"))) {
+    return EncodingType::None;
+  }

  const auto &s = req.get_header_value("Accept-Encoding");
-  (void)(s);
+  if (s.empty()) { return EncodingType::None; }

+  // Single-pass: iterate tokens and track the best supported encoding.
+  // Server preference breaks ties (br > gzip > zstd).
+  EncodingType best = EncodingType::None;
+  double best_q = 0.0; // q=0 means "not acceptable"
+
+  // Server preference: Brotli > Gzip > Zstd (lower = more preferred)
+  auto priority = [](EncodingType t) -> int {
+    switch (t) {
+    case EncodingType::Brotli: return 0;
+    case EncodingType::Gzip: return 1;
+    case EncodingType::Zstd: return 2;
+    default: return 3;
+    }
+  };
+
+  std::string name;
+  split(s.data(), s.data() + s.size(), ',', [&](const char *b, const char *e) {
+    double quality = 1.0;
+    if (!parse_quality(b, e, name, quality)) { return; }
+    if (quality <= 0.0) { return; }
+
+    EncodingType type = EncodingType::None;
 #ifdef CPPHTTPLIB_BROTLI_SUPPORT
-  // TODO: 'Accept-Encoding' has br, not br;q=0
-  ret = s.find("br") != std::string::npos;
-  if (ret) { return EncodingType::Brotli; }
+    if (case_ignore::equal(name, "br")) { type = EncodingType::Brotli; }
 #endif
-
 #ifdef CPPHTTPLIB_ZLIB_SUPPORT
-  // TODO: 'Accept-Encoding' has gzip, not gzip;q=0
-  ret = s.find("gzip") != std::string::npos;
-  if (ret) { return EncodingType::Gzip; }
+    if (type == EncodingType::None && case_ignore::equal(name, "gzip")) {
+      type = EncodingType::Gzip;
+    }
 #endif
-
 #ifdef CPPHTTPLIB_ZSTD_SUPPORT
-  // TODO: 'Accept-Encoding' has zstd, not zstd;q=0
-  ret = s.find("zstd") != std::string::npos;
-  if (ret) { return EncodingType::Zstd; }
+    if (type == EncodingType::None && case_ignore::equal(name, "zstd")) {
+      type = EncodingType::Zstd;
+    }
 #endif

-  return EncodingType::None;
+    if (type == EncodingType::None) { return; }
+
+    // Higher q-value wins; for equal q, server preference breaks ties
+    if (quality > best_q ||
+        (quality == best_q && priority(type) < priority(best))) {
+      best_q = quality;
+      best = type;
+    }
+  });
+
+  return best;
 }

 bool nocompressor::compress(const char *data, size_t data_length,
@@ -2937,6 +3042,21 @@ create_decompressor(const std::string &encoding) {
  return decompressor;
 }

+// Returns the best available compressor and its Content-Encoding name.
+// Priority: Brotli > Gzip > Zstd (matches server-side preference).
+std::pair<std::unique_ptr<compressor>, const char *>
+create_compressor() {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+  return {detail::make_unique<brotli_compressor>(), "br"};
+#elif defined(CPPHTTPLIB_ZLIB_SUPPORT)
+  return {detail::make_unique<gzip_compressor>(), "gzip"};
+#elif defined(CPPHTTPLIB_ZSTD_SUPPORT)
+  return {detail::make_unique<zstd_compressor>(), "zstd"};
+#else
+  return {nullptr, nullptr};
+#endif
+}
+
 bool is_prohibited_header_name(const std::string &name) {
  using udl::operator""_t;

@@ -3769,7 +3889,7 @@ bool parse_accept_header(const std::string &s,
  struct AcceptEntry {
    std::string media_type;
    double quality;
-    int order; // Original order in header
+    int order;
  };

  std::vector<AcceptEntry> entries;
@@ -3787,48 +3907,12 @@ bool parse_accept_header(const std::string &s,
    }

    AcceptEntry accept_entry;
-    accept_entry.quality = 1.0; // Default quality
    accept_entry.order = order++;

-    // Find q= parameter
-    auto q_pos = entry.find(";q=");
-    if (q_pos == std::string::npos) { q_pos = entry.find("; q="); }
-
-    if (q_pos != std::string::npos) {
-      // Extract media type (before q parameter)
-      accept_entry.media_type = trim_copy(entry.substr(0, q_pos));
-
-      // Extract quality value
-      auto q_start = entry.find('=', q_pos) + 1;
-      auto q_end = entry.find(';', q_start);
-      if (q_end == std::string::npos) { q_end = entry.length(); }
-
-      std::string quality_str =
-          trim_copy(entry.substr(q_start, q_end - q_start));
-      if (quality_str.empty()) {
-        has_invalid_entry = true;
-        return;
-      }
-
-      {
-        double v = 0.0;
-        auto res = detail::from_chars(
-            quality_str.data(), quality_str.data() + quality_str.size(), v);
-        if (res.ec == std::errc{}) {
-          accept_entry.quality = v;
-        } else {
-          has_invalid_entry = true;
-          return;
-        }
-      }
-      // Check if quality is in valid range [0.0, 1.0]
-      if (accept_entry.quality < 0.0 || accept_entry.quality > 1.0) {
-        has_invalid_entry = true;
-        return;
-      }
-    } else {
-      // No quality parameter, use entire entry as media type
-      accept_entry.media_type = entry;
+    if (!parse_quality(entry.data(), entry.data() + entry.size(),
+                       accept_entry.media_type, accept_entry.quality)) {
+      has_invalid_entry = true;
+      return;
    }

    // Remove additional parameters from media type
@@ -5481,7 +5565,8 @@ std::string decode_path_component(const std::string &component) {
        // Unicode %uXXXX encoding
        auto val = 0;
        if (detail::from_hex_to_i(component, i + 2, 4, val)) {
-          // 4 digits Unicode codes
+          // 4 digits Unicode codes: val is 0x0000-0xFFFF (from 4 hex digits),
+          // so to_utf8 writes at most 3 bytes. buff[4] is safe.
          char buff[4];
          size_t len = detail::to_utf8(val, buff);
          if (len > 0) { result.append(buff, len); }
@@ -5586,6 +5671,30 @@ std::string decode_query_component(const std::string &component,
  return result;
 }

+std::string sanitize_filename(const std::string &filename) {
+  // Extract basename: find the last path separator (/ or \)
+  auto pos = filename.find_last_of("/\\");
+  auto result =
+      (pos != std::string::npos) ? filename.substr(pos + 1) : filename;
+
+  // Strip null bytes
+  result.erase(std::remove(result.begin(), result.end(), '\0'), result.end());
+
+  // Trim whitespace
+  {
+    auto start = result.find_first_not_of(" \t");
+    auto end = result.find_last_not_of(" \t");
+    result = (start == std::string::npos)
+                 ? ""
+                 : result.substr(start, end - start + 1);
+  }
+
+  // Reject . and ..
+  if (result == "." || result == "..") { return ""; }
+
+  return result;
+}
+
 std::string append_query_params(const std::string &path,
                                       const Params &params) {
  std::string path_with_query = path;
@@ -6714,7 +6823,18 @@ bool Server::set_mount_point(const std::string &mount_point,
  if (stat.is_dir()) {
    std::string mnt = !mount_point.empty() ? mount_point : "/";
    if (!mnt.empty() && mnt[0] == '/') {
-      base_dirs_.push_back({std::move(mnt), dir, std::move(headers)});
+      std::string resolved_base;
+      if (detail::canonicalize_path(dir.c_str(), resolved_base)) {
+#if defined(_WIN32)
+        if (resolved_base.back() != '\\' && resolved_base.back() != '/') {
+          resolved_base += '\\';
+        }
+#else
+        if (resolved_base.back() != '/') { resolved_base += '/'; }
+#endif
+      }
+      base_dirs_.push_back(
+          {std::move(mnt), dir, std::move(resolved_base), std::move(headers)});
      return true;
    }
  }
@@ -6874,6 +6994,20 @@ Server &Server::set_payload_max_length(size_t length) {
  return *this;
 }

+Server &Server::set_websocket_ping_interval(time_t sec) {
+  websocket_ping_interval_sec_ = sec;
+  return *this;
+}
+
+template <class Rep, class Period>
+Server &Server::set_websocket_ping_interval(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(duration, [&](time_t sec, time_t /*usec*/) {
+    set_websocket_ping_interval(sec);
+  });
+  return *this;
+}
+
 bool Server::bind_to_port(const std::string &host, int port,
                                 int socket_flags) {
  auto ret = bind_internal(host, port, socket_flags);
@@ -7294,6 +7428,18 @@ bool Server::handle_file_request(Request &req, Response &res) {
        auto path = entry.base_dir + sub_path;
        if (path.back() == '/') { path += "index.html"; }

+        // Defense-in-depth: is_valid_path blocks ".." traversal in the URL,
+        // but symlinks/junctions can still escape the base directory.
+        if (!entry.resolved_base_dir.empty()) {
+          std::string resolved_path;
+          if (detail::canonicalize_path(path.c_str(), resolved_path) &&
+              !detail::is_path_within_base(resolved_path,
+                                           entry.resolved_base_dir)) {
+            res.status = StatusCode::Forbidden_403;
+            return true;
+          }
+        }
+
        detail::FileStat stat(path);

        if (stat.is_dir()) {
@@ -8012,7 +8158,7 @@ Server::process_request(Stream &strm, const std::string &remote_addr,
        {
          // Use WebSocket-specific read timeout instead of HTTP timeout
          strm.set_read_timeout(CPPHTTPLIB_WEBSOCKET_READ_TIMEOUT_SECOND, 0);
-          ws::WebSocket ws(strm, req, true);
+          ws::WebSocket ws(strm, req, true, websocket_ping_interval_sec_);
          entry.handler(req, ws);
        }
        return true;
@@ -8256,6 +8402,13 @@ bool ClientImpl::ensure_socket_connection(Socket &socket, Error &error) {
  return create_and_connect_socket(socket, error);
 }

+bool ClientImpl::setup_proxy_connection(
+    Socket & /*socket*/,
+    std::chrono::time_point<std::chrono::steady_clock> /*start_time*/,
+    Response & /*res*/, bool & /*success*/, Error & /*error*/) {
+  return true;
+}
+
 void ClientImpl::shutdown_ssl(Socket & /*socket*/,
                                     bool /*shutdown_gracefully*/) {
  // If there are any requests in flight from threads other than us, then it's
@@ -8377,27 +8530,14 @@ bool ClientImpl::send_(Request &req, Response &res, Error &error) {
        return false;
      }

-#ifdef CPPHTTPLIB_SSL_ENABLED
-      // TODO: refactoring
-      if (is_ssl()) {
-        auto &scli = static_cast<SSLClient &>(*this);
-        if (!proxy_host_.empty() && proxy_port_ != -1) {
-          auto success = false;
-          if (!scli.connect_with_proxy(socket_, req.start_time_, res, success,
-                                       error)) {
-            if (!success) { output_error_log(error, &req); }
-            return success;
-          }
-        }
-
-        if (!proxy_host_.empty() && proxy_port_ != -1) {
-          if (!scli.initialize_ssl(socket_, error)) {
-            output_error_log(error, &req);
-            return false;
-          }
+      {
+        auto success = true;
+        if (!setup_proxy_connection(socket_, req.start_time_, res, success,
+                                    error)) {
+          if (!success) { output_error_log(error, &req); }
+          return success;
        }
      }
-#endif
    }

    // Mark the current socket as being in use so that it cannot be closed by
@@ -8558,17 +8698,15 @@ ClientImpl::open_stream(const std::string &method, const std::string &path,
        return handle;
      }

-#ifdef CPPHTTPLIB_SSL_ENABLED
-      if (is_ssl()) {
-        auto &scli = static_cast<SSLClient &>(*this);
-        if (!proxy_host_.empty() && proxy_port_ != -1) {
-          if (!scli.initialize_ssl(socket_, handle.error)) {
-            handle.response.reset();
-            return handle;
-          }
+      {
+        auto success = true;
+        auto start_time = std::chrono::steady_clock::now();
+        if (!setup_proxy_connection(socket_, start_time, *handle.response,
+                                    success, handle.error)) {
+          if (!success) { handle.response.reset(); }
+          return handle;
        }
      }
-#endif
    }

    transfer_socket_ownership_to_handle(handle);
@@ -8847,7 +8985,7 @@ bool ClientImpl::handle_request(Stream &strm, Request &req,

  if (res.get_header_value("Connection") == "close" ||
      (res.version == "HTTP/1.0" && res.reason != "Connection established")) {
-    // TODO this requires a not-entirely-obvious chain of calls to be correct
+    // NOTE: this requires a not-entirely-obvious chain of calls to be correct
    // for this to be safe.

    // This is safe to call because handle_request is only called by send_
@@ -9086,14 +9224,9 @@ bool ClientImpl::write_content_with_provider(Stream &strm,
  auto is_shutting_down = []() { return false; };

  if (req.is_chunked_content_provider_) {
-    // TODO: Brotli support
-    std::unique_ptr<detail::compressor> compressor;
-#ifdef CPPHTTPLIB_ZLIB_SUPPORT
-    if (compress_) {
-      compressor = detail::make_unique<detail::gzip_compressor>();
-    } else
-#endif
-    {
+    auto compressor = compress_ ? detail::create_compressor().first
+                                : std::unique_ptr<detail::compressor>();
+    if (!compressor) {
      compressor = detail::make_unique<detail::nocompressor>();
    }

@@ -9324,14 +9457,15 @@ ClientImpl::send_with_content_provider_and_receiver(
    Error &error) {
  if (!content_type.empty()) { req.set_header("Content-Type", content_type); }

-#ifdef CPPHTTPLIB_ZLIB_SUPPORT
-  if (compress_) { req.set_header("Content-Encoding", "gzip"); }
-#endif
+  auto enc = compress_
+                 ? detail::create_compressor()
+                 : std::pair<std::unique_ptr<detail::compressor>, const char *>(
+                       nullptr, nullptr);

-#ifdef CPPHTTPLIB_ZLIB_SUPPORT
-  if (compress_ && !content_provider_without_length) {
-    // TODO: Brotli support
-    detail::gzip_compressor compressor;
+  if (enc.second) { req.set_header("Content-Encoding", enc.second); }
+
+  if (enc.first && !content_provider_without_length) {
+    auto &compressor = enc.first;

    if (content_provider) {
      auto ok = true;
@@ -9342,7 +9476,7 @@ ClientImpl::send_with_content_provider_and_receiver(
        if (ok) {
          auto last = offset + data_len == content_length;

-          auto ret = compressor.compress(
+          auto ret = compressor->compress(
              data, data_len, last,
              [&](const char *compressed_data, size_t compressed_data_len) {
                req.body.append(compressed_data, compressed_data_len);
@@ -9366,19 +9500,17 @@ ClientImpl::send_with_content_provider_and_receiver(
        }
      }
    } else {
-      if (!compressor.compress(body, content_length, true,
-                               [&](const char *data, size_t data_len) {
-                                 req.body.append(data, data_len);
-                                 return true;
-                               })) {
+      if (!compressor->compress(body, content_length, true,
+                                [&](const char *data, size_t data_len) {
+                                  req.body.append(data, data_len);
+                                  return true;
+                                })) {
        error = Error::Compression;
        output_error_log(error, &req);
        return nullptr;
      }
    }
-  } else
-#endif
-  {
+  } else {
    if (content_provider) {
      req.content_length_ = content_length;
      req.content_provider_ = std::move(content_provider);
@@ -11545,6 +11677,24 @@ bool SSLClient::create_and_connect_socket(Socket &socket, Error &error) {
  return ClientImpl::create_and_connect_socket(socket, error);
 }

+bool SSLClient::setup_proxy_connection(
+    Socket &socket,
+    std::chrono::time_point<std::chrono::steady_clock> start_time,
+    Response &res, bool &success, Error &error) {
+  if (proxy_host_.empty() || proxy_port_ == -1) { return true; }
+
+  if (!connect_with_proxy(socket, start_time, res, success, error)) {
+    return false;
+  }
+
+  if (!initialize_ssl(socket, error)) {
+    success = false;
+    return false;
+  }
+
+  return true;
+}
+
 // Assumes that socket_mutex_ is locked and that there are no requests in
 // flight
 bool SSLClient::connect_with_proxy(
@@ -16061,11 +16211,11 @@ WebSocket::~WebSocket() {
 }

 void WebSocket::start_heartbeat() {
+  if (ping_interval_sec_ == 0) { return; }
  ping_thread_ = std::thread([this]() {
    std::unique_lock<std::mutex> lock(ping_mutex_);
    while (!closed_) {
-      ping_cv_.wait_for(lock, std::chrono::seconds(
-                                  CPPHTTPLIB_WEBSOCKET_PING_INTERVAL_SECOND));
+      ping_cv_.wait_for(lock, std::chrono::seconds(ping_interval_sec_));
      if (closed_) { break; }
      lock.unlock();
      if (!send_frame(Opcode::Ping, nullptr, 0)) {
@@ -16203,7 +16353,8 @@ bool WebSocketClient::connect() {
  Request req;
  req.method = "GET";
  req.path = path_;
-  ws_ = std::unique_ptr<WebSocket>(new WebSocket(std::move(strm), req, false));
+  ws_ = std::unique_ptr<WebSocket>(
+      new WebSocket(std::move(strm), req, false, websocket_ping_interval_sec_));
  return true;
 }

@@ -16243,6 +16394,10 @@ void WebSocketClient::set_write_timeout(time_t sec, time_t usec) {
  write_timeout_usec_ = usec;
 }

+void WebSocketClient::set_websocket_ping_interval(time_t sec) {
+  websocket_ping_interval_sec_ = sec;
+}
+
 #ifdef CPPHTTPLIB_SSL_ENABLED

 void WebSocketClient::set_ca_cert_path(const std::string &path) {
@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H

-#define CPPHTTPLIB_VERSION "0.37.2"
-#define CPPHTTPLIB_VERSION_NUM "0x002502"
+#define CPPHTTPLIB_VERSION "0.38.0"
+#define CPPHTTPLIB_VERSION_NUM "0x002600"

 #ifdef _WIN32
 #if defined(_WIN32_WINNT) && _WIN32_WINNT < 0x0A00
@@ -1666,6 +1666,11 @@ public:

  Server &set_payload_max_length(size_t length);

+  Server &set_websocket_ping_interval(time_t sec);
+  template <class Rep, class Period>
+  Server &set_websocket_ping_interval(
+      const std::chrono::duration<Rep, Period> &duration);
+
  bool bind_to_port(const std::string &host, int port, int socket_flags = 0);
  int bind_to_any_port(const std::string &host, int socket_flags = 0);
  bool listen_after_bind();
@@ -1700,6 +1705,8 @@ protected:
  time_t idle_interval_sec_ = CPPHTTPLIB_IDLE_INTERVAL_SECOND;
  time_t idle_interval_usec_ = CPPHTTPLIB_IDLE_INTERVAL_USECOND;
  size_t payload_max_length_ = CPPHTTPLIB_PAYLOAD_MAX_LENGTH;
+  time_t websocket_ping_interval_sec_ =
+      CPPHTTPLIB_WEBSOCKET_PING_INTERVAL_SECOND;

 private:
  using Handlers =
@@ -1769,6 +1776,7 @@ private:
  struct MountPointEntry {
    std::string mount_point;
    std::string base_dir;
+    std::string resolved_base_dir;
    Headers headers;
  };
  std::vector<MountPointEntry> base_dirs_;
@@ -2186,6 +2194,10 @@ protected:

  virtual bool create_and_connect_socket(Socket &socket, Error &error);
  virtual bool ensure_socket_connection(Socket &socket, Error &error);
+  virtual bool setup_proxy_connection(
+      Socket &socket,
+      std::chrono::time_point<std::chrono::steady_clock> start_time,
+      Response &res, bool &success, Error &error);

  // All of:
  //   shutdown_ssl
@@ -2712,6 +2724,10 @@ private:
                 std::function<bool(Stream &strm)> callback) override;
  bool is_ssl() const override;

+  bool setup_proxy_connection(
+      Socket &socket,
+      std::chrono::time_point<std::chrono::steady_clock> start_time,
+      Response &res, bool &success, Error &error) override;
  bool connect_with_proxy(
      Socket &sock,
      std::chrono::time_point<std::chrono::steady_clock> start_time,
@@ -2911,6 +2927,8 @@ std::string encode_query_component(const std::string &component,
 std::string decode_query_component(const std::string &component,
                                   bool plus_as_space = true);

+std::string sanitize_filename(const std::string &filename);
+
 std::string append_query_params(const std::string &path, const Params &params);

 std::pair<std::string, std::string> make_range_header(const Ranges &ranges);
@@ -3714,15 +3732,19 @@ private:
  friend class httplib::Server;
  friend class WebSocketClient;

-  WebSocket(Stream &strm, const Request &req, bool is_server)
-      : strm_(strm), req_(req), is_server_(is_server) {
+  WebSocket(
+      Stream &strm, const Request &req, bool is_server,
+      time_t ping_interval_sec = CPPHTTPLIB_WEBSOCKET_PING_INTERVAL_SECOND)
+      : strm_(strm), req_(req), is_server_(is_server),
+        ping_interval_sec_(ping_interval_sec) {
    start_heartbeat();
  }

-  WebSocket(std::unique_ptr<Stream> &&owned_strm, const Request &req,
-            bool is_server)
+  WebSocket(
+      std::unique_ptr<Stream> &&owned_strm, const Request &req, bool is_server,
+      time_t ping_interval_sec = CPPHTTPLIB_WEBSOCKET_PING_INTERVAL_SECOND)
      : strm_(*owned_strm), owned_strm_(std::move(owned_strm)), req_(req),
-        is_server_(is_server) {
+        is_server_(is_server), ping_interval_sec_(ping_interval_sec) {
    start_heartbeat();
  }

@@ -3733,6 +3755,7 @@ private:
  std::unique_ptr<Stream> owned_strm_;
  Request req_;
  bool is_server_;
+  time_t ping_interval_sec_;
  std::atomic<bool> closed_{false};
  std::mutex write_mutex_;
  std::thread ping_thread_;
@@ -3761,6 +3784,7 @@ public:
  const std::string &subprotocol() const;
  void set_read_timeout(time_t sec, time_t usec = 0);
  void set_write_timeout(time_t sec, time_t usec = 0);
+  void set_websocket_ping_interval(time_t sec);

 #ifdef CPPHTTPLIB_SSL_ENABLED
  void set_ca_cert_path(const std::string &path);
@@ -3784,6 +3808,8 @@ private:
  time_t read_timeout_usec_ = 0;
  time_t write_timeout_sec_ = CPPHTTPLIB_CLIENT_WRITE_TIMEOUT_SECOND;
  time_t write_timeout_usec_ = CPPHTTPLIB_CLIENT_WRITE_TIMEOUT_USECOND;
+  time_t websocket_ping_interval_sec_ =
+      CPPHTTPLIB_WEBSOCKET_PING_INTERVAL_SECOND;

 #ifdef CPPHTTPLIB_SSL_ENABLED
  bool is_ssl_ = false;
Author	SHA1	Message	Date
arthw	2985be3324	update hw info	2026-03-31 09:24:40 +08:00
arthw	8dc96153c3	enhance FA stable in UT	2026-03-17 15:57:02 +08:00
Neo Zhang	b6c83aad55	[SYCL] ehance UPSCALE to support all UT cases (#20637 ) * [SYCL] ehance UPSCALE to support more cases * rm test case result of SYCL1	2026-03-17 10:01:52 +08:00
Piotr Wilkin (ilintar)	2e4a6edd4a	tools/server: support refusal content for Responses API (#20285 ) * Support refusal content for Responses API * Update tools/server/server-common.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update tools/server/server-common.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2026-03-17 01:42:04 +01:00
Xuan-Son Nguyen	d34ff7eb5b	model: mistral small 4 support (#20649 ) * model: mistral small 4 support * fix test * fix test (2) * Apply suggestions from code review Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * change newline --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2026-03-17 00:31:14 +01:00
Georgi Gerganov	45172df4d6	ci : disable AMX jobs (#20654 ) [no ci]	2026-03-16 22:38:59 +02:00
Georgi Gerganov	9b342d0a9f	benches : add Nemotron 3 Nano on DGX Spark (#20652 ) [no ci]	2026-03-16 21:50:43 +02:00
Sigbjørn Skjæret	55e87026f7	tests : write to binary buffer to avoid newline translation in jinja -py [no ci] (#20365 )	2026-03-16 20:40:22 +01:00
Martin Klacer	cf21cdf36c	kleidiai: add data type check to get_tensor_traits (#20639 ) * kleidiai: add data type check to get_tensor_traits * Added check for F16 data type into get_tensor_traits path with input data not in ggml_backend_cpu_kleidiai_buffer_type format (unsupported for Q4/8) Signed-off-by: Martin Klacer <martin.klacer@arm.com> Change-Id: I9aca4b9b8d669d35db6f1dbcc4e080b1919b1de7 * updated ggml/src/ggml-cpu/kleidiai/kleidiai.cpp updated kleidiai.cpp file as per suggestion Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Signed-off-by: Martin Klacer <martin.klacer@arm.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2026-03-16 21:25:54 +02:00
Sigbjørn Skjæret	0ed992973b	ci : update labeler (#20629 )	2026-03-16 20:24:20 +01:00
Aldehir Rojas	1bbec6a75d	jinja : add capability check for object args (#20612 )	2026-03-16 17:43:14 +01:00
Georgi Gerganov	f47a246a08	sync : ggml	2026-03-16 17:22:06 +02:00
Georgi Gerganov	c0ccbd1f86	ggml : try fix arm build (whisper/0)	2026-03-16 17:22:06 +02:00
David366AI	f6da02c3f2	ggml : extend im2col f16 (ggml/1434) * examples/yolo: fix load_model memory leak * fix/issue-1433 ggml_compute_forward_im2col_f16 assert error * fix/issue-1433	2026-03-16 17:22:06 +02:00
Pascal	dddca026bf	webui: add model information dialog to router mode (#20600 ) * webui: add model information dialog to router mode * webui: add "Available models" section header in model list * webui: remove nested scrollbar from chat template in model info dialog * chore: update webui build output * feat: UI improvements * refactor: Cleaner rendering + UI docs * chore: update webui build output --------- Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>	2026-03-16 15:38:11 +01:00
Aman Gupta	3c8521c4f5	llama-graph: replace cont with reshape for alpha in qwen35 (#20640 )	2026-03-16 22:07:13 +08:00
Aleksander Grygier	67a2209fab	webui: Add MCP CORS Proxy detection logic & UI (#20167 ) * refactor: MCP store cleanup * feat: Add MCP proxy availability detection * fix: Sidebar icon * chore: update webui build output * chore: Formatting * chore: update webui build output * chore: Update package lock * chore: update webui build output * chore: update webui build output * chore: update webui build output	2026-03-16 13:05:36 +01:00
Pascal	d65c4f2dc9	Fix model selector locked to first loaded model with multiple models (#20580 ) * webui: fix model selector being locked to first loaded model When multiple models are loaded, the auto-select effect would re-fire on every loadedModelIds change, overriding the user's manual model selection. Guard with selectedModelId so auto-select only kicks in when no model is chosen yet. * chore: update webui build output	2026-03-16 12:04:06 +01:00
Woof Dog	d8c331c0af	webui: use date in more human readable exported filename (#19939 ) * webui: use date in exported filename Move conversation naming and export to utils update index.html.gz * webui: move literals to message export constants file * webui: move export naming and download back to the conversation store * chore: update webui build output * webui: add comments to some constants * chore: update webui build output	2026-03-16 11:18:13 +01:00
Ruben Ortlam	46dba9fce8	vulkan: fix flash attention dot product precision (#20589 )	2026-03-16 10:45:49 +01:00
Sigbjørn Skjæret	de8f01c2d7	model : wire up Nemotron-H tensors for NVFP4 support (#20561 ) * wire up Nemotron-H tensors for NVFP4 support * add ssm tensors * alignment	2026-03-16 09:19:16 +01:00
Richard Davison	079e5a45f0	convert : support mixed-precision ModelOpt models with per-tensor NVFP4/FP8 quantization (#20539 ) * support mixed-precision ModelOpt models with per-tensor NVFP4/FP8 quantization * cleanup * fallback --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2026-03-16 09:18:47 +01:00
Masato Nakasaka	d3936498a3	common : fix iterator::end() dereference (#20445 )	2026-03-16 08:50:38 +02:00
Aman Gupta	34818ea6c0	CUDA: GDN hide memory latency (#20537 )	2026-03-16 11:41:45 +08:00
Piotr Wilkin (ilintar)	9e2e2198b0	tools/cli: fix disable reasoning (#20606 )	2026-03-15 22:40:53 +01:00
Georgi Gerganov	88915cb55c	server : fix wait in test_cancel_requests() test (#20601 ) * server : fix wait in test_cancel_requests() test * codeowners : add team for server tests	2026-03-15 20:54:37 +02:00
Sigbjørn Skjæret	ebbf544ed1	sycl : fix for untransposed GDA recurrent state (#20583 )	2026-03-15 19:10:15 +01:00
Sigbjørn Skjæret	b91d7dfe5b	ci : only save openvino caches on github-hosted master (#20593 ) * only save openvino ccache on master * disable toolkit cache if self-hosted * only cache on github-hosted runners * remove toolkit cache [no ci]	2026-03-15 18:58:13 +01:00
Johannes Gäßler	ae40cd27c8	CUDA: limit number of FA stream-k CUDA blocks (#20586 )	2026-03-15 18:30:47 +01:00
Pascal	ceef6b5233	ggml: avoid creating CUDA context during device init (#20595 )	2026-03-16 00:42:56 +08:00
Adrien Gallouët	07c6a59b4f	vendor : update cpp-httplib to 0.38.0 (#20578 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2026-03-15 17:30:06 +01:00
MoonShadow	8b7d340b6f	ggml/hip: fix APU compatibility - soft error handling for hipMemAdviseSetCoarseGrain (#20536 ) * ggml/hip: fix APU compatibility - soft error handling for hipMemAdviseSetCoarseGrain On AMD APU/iGPU devices (unified memory architecture), hipMemAdviseSetCoarseGrain returns hipErrorInvalidValue because the hint is not applicable to UMA systems. The previous CUDA_CHECK() call treated this as a fatal error, causing crashes on APU systems such as AMD Strix Halo (gfx1151). Fix: treat hipMemAdviseSetCoarseGrain as an optional performance hint - call it without error checking and clear any resulting error with hipGetLastError(). Also add pre-allocation debug logging (GGML_LOG_DEBUG) to help diagnose memory issues on APU systems, and store totalGlobalMem in device info. Context: AMD APUs on Windows are affected by a ROCm runtime bug that limits hipMallocManaged to ~64GB regardless of available system RAM. A fix has been submitted upstream: https://github.com/ROCm/rocm-systems/pull/4077 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * ggml/hip: remove unrelated changes, keep only hipMemAdviseSetCoarseGrain fix --------- Co-authored-by: moonshadow-25 <moonshadow-25@users.noreply.github.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-03-15 17:23:58 +01:00
Eric Hsieh	559646472d	fix: prevent nullptr dereference (#20552 )	2026-03-15 16:51:49 +01:00
Sigbjørn Skjæret	cf45437d35	codeowners : use teams (#20526 ) * use teams * update * update * update * update * update	2026-03-15 14:26:10 +01:00
Georgi Gerganov	9cd4ebcfb1	ci : split build.yml + server.yml (#20546 ) * ci : split build.yml * cont : split server.yml * cont : reduce paths * cont : split build-android.yml + update paths * ci : make msys workflows manual (#20588) * ci : make cross-build workflows manual (#20585) * cont : fix release paths Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2026-03-15 15:11:17 +02:00
Sigbjørn Skjæret	89d0aec042	convert : support contiguous method on lora tensors (#20489 )	2026-03-15 12:15:12 +01:00
Bartowski	b9da4444df	ggml : guard against sumq2 being 0 in IQ4_NL (#20460 )	2026-03-15 10:47:28 +02:00
PikaPikachu	617db241aa	cuda : add RDNA4-specific MMVQ parameter table for bs=1 decode (#19478 ) * mmvq: add RDNA3/RDNA4-specific parameter table (nwarps=8, rows=1) * mmvq: add dedicated RDNA3 parameter table * mmvq: exclude RDNA3.5 (gfx1150/1151) from RDNA3 table	2026-03-15 08:33:39 +01:00
Ruben Ortlam	1a3d8edbba	vulkan: use graphics queue on AMD (#20551 ) * vulkan: use graphics queue on AMD for slightly better performance * disable async transfer queue on AMD	2026-03-15 08:18:54 +01:00
sprayandwipe	6b10a82c00	kv-cache : fix reading llama_kv_cell_ext during state read (#20273 ) Co-authored-by: sid <sid@ragingfist.net>	2026-03-15 09:11:19 +02:00
Michael Wand	d23355afc3	model : wire up Qwen3.5/Qwen3.5MoE tensors for NVFP4 support (#20506 )	2026-03-14 22:44:42 +01:00
Georgi Gerganov	b30a5fdf37	metal : add FA specialization for HSK = 320, HSV = 256 (#20549 )	2026-03-14 23:15:47 +02:00
Georgi Gerganov	b4768955c4	ci : move self-hosted workflows to separate files (#20540 )	2026-03-14 23:15:35 +02:00
Gerard Guillemas Martos	fc350fdf96	docker : force Python 3.13 in Vulkan container (#20530 ) * ci: force Python 3.13 in Vulkan container * remove unnecessary `update-alternatives` line	2026-03-14 21:37:09 +01:00