ggml-webgpu: fix CI errors from #25025 and #25262 (#26566 )

* test new flash_attn test * rebase and fix to disable subgrou matrices when max_kv_tile == 0 * delete log output * Add i32 support to cpy and enables the all ops test * restore the non target ci tests * comment out of TODO of build-cpu.yml * fix format
Address review comment of PR 25532 (#26852 )
2026-08-11 22:26:21 +02:00 · 2026-08-11 07:10:00 +03:00 · 2026-08-11 00:02:25 +05:30 · 2026-08-10 11:09:19 -07:00 · 2026-08-10 19:53:12 +02:00 · 2026-08-10 16:58:56 +03:00
730 changed files with 47936 additions and 10250 deletions
@@ -57,7 +57,6 @@ COPY --from=web /app/tools/ui/dist tools/ui/dist
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build \
        -DGGML_HIP=ON \
-        -DGGML_HIP_ROCWMMA_FATTN=ON \
        -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
        -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
        -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
@@ -4,6 +4,10 @@ inputs:
  cuda_version:
    description: "CUDA toolkit version"
    required: true
+  cuda_arch:
+    description: "CUDA target architecture"
+    required: false
+    default: "x64"

 runs:
  using: "composite"
@@ -127,3 +131,26 @@ runs:
          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
          echo "CUDA_PATH_V13_3=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+    - name: Install Cuda Toolkit 13.4 for ARM64
+      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'arm64' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          choco install unzip -y
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cccl-windows-x86_64-13.3.4.1.2-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_crt-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_nvcc-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/libnvvm-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_cudart-windows-arm64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/libcublas-windows-arm64-13.7.0.10-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.1.2-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-arm64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-arm64-13.7.0.10-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
@@ -8,8 +8,26 @@ inputs:
 runs:
  using: "composite"
  steps:
-    - name: Setup ROCm
-      uses: ./.github/actions/install-exe
-      with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
-        args: -install
+    - name: Install ROCm with Wheels
+      shell: pwsh
+      run: |
+        $ErrorActionPreference = "Stop"
+        write-host "Setting up Python virtual environment"
+
+        # Create the venv directly at the cache location to avoid relocation issues
+        New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
+        python -m venv C:\TheRock\build\.venv
+        & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+        write-host "Upgrading pip"
+        python -m pip install --upgrade pip
+
+        write-host "Installing ROCm wheels for multi-arch support"
+        # Install ROCm wheels for multi-arch support (this may take several minutes)
+        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
+
+        # Pre-expand the devel tree so it is included in the cache
+        write-host "Initializing ROCm devel tree"
+        rocm-sdk init
+        if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+        write-host "Completed ROCm wheel installation to C:\TheRock\build"
@@ -60,7 +60,6 @@ jobs:
            -DCMAKE_BUILD_RPATH="@loader_path" \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=OFF \
            -DGGML_METAL_SHADER_DEBUG=ON \
            -DGGML_RPC=ON \
@@ -127,7 +126,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -178,7 +176,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -212,7 +209,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -257,7 +253,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -123,8 +123,8 @@ jobs:
    runs-on: windows-2022

    env:
-      # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
+      # Make sure this is in sync with release.yml and build-cuda-windows.yml
+      ROCM_VERSION: "7.14.0"

    steps:
      - name: Clone
@@ -135,11 +135,11 @@ jobs:
        uses: actions/cache@v5
        id: cache-rocm
        with:
-          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+          path: C:\TheRock\build
+          key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}

      - name: Setup ROCm
        if: steps.cache-rocm.outputs.cache-hit != 'true'
        uses: ./.github/actions/windows-setup-rocm
        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
+          version: ${{ env.ROCM_VERSION }}
@@ -99,7 +99,6 @@ jobs:
        run: |
          cmake -B build -S . \
            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
            -DGPU_TARGETS="gfx1030" \
            -DGGML_HIP=ON
          cmake --build build --config Release -j $(nproc)
@@ -83,7 +83,7 @@ jobs:

    env:
      # Make sure this is in sync with build-cache.yml
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
+      ROCM_VERSION: "7.14.0"

    strategy:
      matrix:
@@ -97,36 +97,53 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      - name: Grab rocWMMA package
-        id: grab_rocwmma
-        run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
-
-      - name: Use ROCm Installation Cache
+      - name: Cache ROCm Installation
        uses: actions/cache@v5
        id: cache-rocm
        with:
-          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+          path: C:\TheRock\build
+          key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}

      - name: Setup ROCm
        if: steps.cache-rocm.outputs.cache-hit != 'true'
        uses: ./.github/actions/windows-setup-rocm
        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
+          version: ${{ env.ROCM_VERSION }}
+
+      - name: Setup ROCm Environment
+        run: |
+          $ErrorActionPreference = "Stop"
+
+          # Activate venv from cache or fresh install
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          # Expand the devel tree (idempotent; no-op if already done during install)
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          $rocmPath = (rocm-sdk path --root)
+          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
+          $rocmPath = $rocmPath.Trim()
+          $cmakePath = (rocm-sdk path --cmake).Trim()
+          $binPath = (rocm-sdk path --bin).Trim()
+          write-host "ROCm root: $rocmPath"
+
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$binPath" >> $env:GITHUB_PATH
+
+          # Keep venv in PATH for subsequent steps
+          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH

      - name: Verify ROCm
        id: verify
        run: |
-          # Find and test ROCm installation
-          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
-          if (-not $clangPath) {
-            Write-Error "ROCm installation not found"
-            exit 1
-          }
-          & $clangPath.FullName --version
+          # Test the ROCm clang shipped in the installed wheel
+          & "${env:HIP_PATH}\lib\llvm\bin\clang.exe" --version

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -134,29 +151,27 @@ jobs:
          # TODO: this build does not match the build in release.yml, so we use a different cache key
          #       ideally, the builds should match, similar to the CUDA build above so that we would be able
          #       to populate the ccache for the release with manual runs of this workflow
-          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}

      - name: Build
        id: cmake_build
        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
            -DCMAKE_BUILD_TYPE=Release `
            -DLLAMA_BUILD_BORINGSSL=ON `
-            -DROCM_DIR="${env:HIP_PATH}" `
+            -DHIP_PATH="${env:HIP_PATH}" `
            -DGGML_HIP=ON `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGPU_TARGETS="gfx1100"  `
+            -DGPU_TARGETS="gfx1100" `
            -DGGML_RPC=ON
          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
@@ -15,6 +15,12 @@ on:
      '**/*.cpp'
    ]

+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-sanitize.yml'
+    ]
+
 concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
  cancel-in-progress: true
@@ -28,19 +34,35 @@ env:

 jobs:
  ctest:
-    runs-on: [self-hosted, X64, CPU, Linux]
-
    continue-on-error: true

    strategy:
      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        include:
+        - sanitizer: ADDRESS
+          machine: [self-hosted, X64, Linux]
+        # thread doesn't run properly on some self hosted machines, so run it on Github instead
+        - sanitizer: THREAD
+          machine: ubuntu-24.04
+        - sanitizer: UNDEFINED
+          machine: [self-hosted, X64, Linux]
+
+    runs-on: ${{ matrix.machine }}

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

+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        with:
+          key: ctest-thread-ubuntu-24.04
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
      # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
      - name: Build (undefined)
        id: cmake_build_undefined
@@ -71,6 +71,26 @@ jobs:
          nvidia-smi
          GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

+  gpu-rocm:
+    runs-on: [self-hosted, Linux, AMD]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        # HIP_LAUNCH_BLOCKING=1: workaround for an async-execution correctness
+        # issue on integrated RDNA3.5 (gfx1151) where batched inference returns
+        # incorrect output (perplexity ~88 vs ~9.4). Serializing kernel launches
+        # restores correctness. Remove once the underlying ROCm/HIP issue is fixed.
+        env:
+          HIP_LAUNCH_BLOCKING: "1"
+        run: |
+          rocminfo
+          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
  gpu-vulkan-nvidia-cm:
    runs-on: [self-hosted, Linux, NVIDIA]

@@ -0,0 +1,23 @@
+name: Convert PR to draft
+
+on:
+  pull_request_target:
+    types: [labeled]
+
+permissions:
+  pull-requests: write
+  issues: write
+  contents: write # required for "gh pr ready" command, see https://github.com/cli/cli/issues/8910
+
+jobs:
+  convert-to-draft:
+    if: github.event.label.name == 'draft' && github.event.pull_request.draft == false
+    runs-on: ubuntu-slim
+    steps:
+      - name: Convert PR to draft
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_URL: ${{ github.event.pull_request.html_url }}
+        run: |
+          gh pr ready --undo "$PR_URL"
+          gh pr edit "$PR_URL" --remove-label draft
@@ -93,13 +93,13 @@ jobs:
          - build: 'arm64'
            arch: 'arm64'
            os: macos-26
-            defines: "-DGGML_METAL_USE_BF16=ON -DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3"
+            defines: "-DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3"
          # TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23780)
          #       in order to enable it again, we have to provision dedicated runners  to run it
          #- build: 'arm64-kleidiai'
          #  arch: 'arm64'
          #  os: macos-14
-          #  defines: "-DGGML_METAL_USE_BF16=ON -DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3 -DGGML_CPU_KLEIDIAI=ON"
+          #  defines: "-DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3 -DGGML_CPU_KLEIDIAI=ON"
          - build: 'x64'
            arch: 'x64'
            os: macos-15-intel
@@ -748,6 +748,132 @@ jobs:
          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

+  windows-rocm:
+    runs-on: windows-2022
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+            build: x64
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v5
+        with:
+          path: C:\TheRock\build
+          key: rocm-wheels-${{ matrix.ROCM_VERSION }}-multi-arch-${{ runner.os }}
+
+      - name: Setup ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        uses: ./.github/actions/windows-setup-rocm
+        with:
+          version: ${{ matrix.ROCM_VERSION }}
+
+      - name: Setup ROCm Environment
+        run: |
+          $ErrorActionPreference = "Stop"
+
+          # Activate venv from cache or fresh install
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          # Expand the devel tree (idempotent; no-op if already done during install)
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          $rocmPath = (rocm-sdk path --root)
+          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
+          $rocmPath = $rocmPath.Trim()
+          $cmakePath = (rocm-sdk path --cmake).Trim()
+          $binPath = (rocm-sdk path --bin).Trim()
+          write-host "ROCm root: $rocmPath"
+          write-host "CMake path: $cmakePath"
+          write-host "Bin path: $binPath"
+
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$binPath" >> $env:GITHUB_PATH
+
+          # Keep venv in PATH for subsequent steps
+          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
+
+      - name: Build
+        run: |
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_HIP=ON `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_C_FLAGS="-Wno-error=incompatible-pointer-types" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DHIP_PATH="${env:HIP_PATH}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+
+      - name: Verify HIP backend was built
+        run: |
+          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
+          if (-not $hipDll) {
+            Write-Host "##[error]ggml-hip*.dll was NOT produced. The HIP backend silently failed to build."
+            Write-Host "Contents of build\bin:"
+            Get-ChildItem build\bin | Format-Table -AutoSize
+            exit 1
+          }
+          Write-Host "HIP backend artifact found:"
+          $hipDll | Format-Table FullName, Length -AutoSize
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Get ROCm short version
+        run: |
+          $rocmVersionShort = ('${{ matrix.ROCM_VERSION }}'.Split('.')[0..1] -join '.')
+          echo "ROCM_VERSION_SHORT=$rocmVersionShort" >> $env:GITHUB_ENV
+
+      - name: Pack artifacts
+        run: |
+          cp "LICENSE" "build\bin\"
+          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
+          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
+
  windows:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -848,6 +974,7 @@ jobs:
          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip

  windows-cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

@@ -858,7 +985,16 @@ jobs:

    strategy:
      matrix:
-        cuda: ['12.4', '13.3']
+        include:
+          - cuda: '12.4'
+            arch: x64
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
+          - cuda: '13.3'
+            arch: x64
+            defines: ''
+          - cuda: '13.4'
+            arch: arm64
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -876,6 +1012,7 @@ jobs:
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -885,54 +1022,62 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Build
        id: cmake_build
        shell: cmd
        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
          cmake -S . -B build -G "Ninja Multi-Config" ^
            -DGGML_BACKEND_DL=ON ^
            -DGGML_NATIVE=OFF ^
            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Pack artifacts
        id: pack_artifacts
        run: |
-          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
+          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip .\build\bin\Release\ggml-cuda.dll

      - name: Upload artifacts
        uses: actions/upload-artifact@v6
        with:
-          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

-      - name: Copy and pack Cuda runtime
+      - name: Copy and pack Cuda runtime (x64)
+        if: ${{ matrix.arch == 'x64' }}
        run: |
          echo "Cuda install location: ${{ env.CUDA_PATH }}"
          $dst='.\build\bin\cudart\'
          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
          robocopy "${{env.CUDA_PATH}}\bin\x64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*
+
+      - name: Copy and pack Cuda runtime (ARM64)
+        if: ${{ matrix.arch == 'arm64' }}
+        run: |
+          echo "Cuda install location: ${{ env.CUDA_PATH }}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{env.CUDA_PATH}}\bin\arm64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*

      - name: Upload Cuda runtime
        uses: actions/upload-artifact@v6
        with:
-          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

  windows-sycl:
    needs: [check-release]
@@ -1149,8 +1294,8 @@ jobs:
    strategy:
      matrix:
        include:
-          - ROCM_VERSION: "7.2.1"
-            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
            build: 'x64'

    steps:
@@ -1182,38 +1327,36 @@ jobs:
        run: |
          sudo apt install -y build-essential git cmake wget

-      - name: Setup Legacy ROCm
-        if: matrix.ROCM_VERSION == '7.2.1'
-        id: legacy_env
-        run: |
-          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
-          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
-            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-
-          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
-          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
-          EOF
-
-          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
-          Package: *
-          Pin: release o=repo.radeon.com
-          Pin-Priority: 600
-          EOF
-
-          sudo apt update
-          sudo apt-get install -y libssl-dev rocm-hip-sdk
-
-      - name: Setup TheRock
-        if: matrix.ROCM_VERSION != '7.2.1'
+      - name: Setup TheRock with Wheels
        id: therock_env
        run: |
-          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
-          mkdir install
-          tar -xf *.tar.gz -C install
-          export ROCM_PATH=$(pwd)/install
-          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
-          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
-          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
+          # Create Python virtual environment
+          python3 -m venv .venv
+          source .venv/bin/activate
+
+          # Install ROCm wheels for build
+          # libraries = HIP runtime and CMake configs needed for linking
+          # devel = compilers, headers, static libs
+          python -m pip install --upgrade pip
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          ROCM_PATH=$(rocm-sdk path --root)
+          CMAKE_PATH=$(rocm-sdk path --cmake)
+          BIN_PATH=$(rocm-sdk path --bin)
+          echo "ROCM_PATH=$ROCM_PATH"
+          echo "CMAKE_PATH=$CMAKE_PATH"
+          echo "BIN_PATH=$BIN_PATH"
+
+          # Set environment variables
+          echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
+          echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
+          echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
+
+          # Keep venv activated for subsequent steps
+          echo "$(pwd)/.venv/bin" >> $GITHUB_PATH

      - name: Build with native CMake HIP support
        id: cmake_build
@@ -1229,7 +1372,6 @@ jobs:
            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
            -DGGML_HIP=ON \
            -DHIP_PLATFORM=amd \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)
@@ -1258,130 +1400,6 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz

-  windows-hip:
-    needs: [check-release, get-version]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
-    runs-on: windows-2022
-
-    permissions:
-      actions: write
-
-    env:
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
-
-    strategy:
-      matrix:
-        include:
-          - name: "radeon"
-            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
-      - name: Grab rocWMMA package
-        id: grab_rocwmma
-        run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
-
-      - name: Cache ROCm Installation
-        id: cache-rocm
-        uses: actions/cache@v5
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-
-      - name: Install ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
-        id: depends
-        run: |
-          $ErrorActionPreference = "Stop"
-          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP SDK"
-          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $completed = $proc.WaitForExit(600000)
-          if (-not $completed) {
-              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
-              $proc.Kill()
-              exit 1
-          }
-          if ($proc.ExitCode -ne 0) {
-              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
-              exit 1
-          }
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        id: verify
-        run: |
-          # Find and test ROCm installation
-          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
-          if (-not $clangPath) {
-            Write-Error "ROCm installation not found"
-            exit 1
-          }
-          & $clangPath.FullName --version
-
-      - name: Build
-        id: cmake_build
-        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DGGML_BACKEND_DL=ON `
-            -DGGML_NATIVE=OFF `
-            -DGGML_CPU=OFF `
-            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_HIP=ON `
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} `
-            -DLLAMA_BUILD_BORINGSSL=ON
-          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
-          md "build\bin\rocblas\library\"
-          md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
-          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
-          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
-
  ios-xcode:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1402,7 +1420,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -1556,7 +1573,7 @@ jobs:
      - windows-cpu
      - windows-cuda
      #- windows-sycl
-      - windows-hip
+      - windows-rocm
      - windows-openvino
      - ubuntu-22-rocm
      - ubuntu-cpu
@@ -1668,7 +1685,7 @@ jobs:
            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
            - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
            - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
            - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1682,10 +1699,11 @@ jobs:
            - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-opencl-adreno-arm64.zip)
            - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
            - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.3-x64.zip) - [CUDA 13.3 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.3-x64.zip)
+            - [Windows arm64 (CUDA 13) (preview)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.4-arm64.zip) - [CUDA 13.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.4-arm64.zip)
            - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
            - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
            - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
+            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)

            **openEuler:**
            - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)
@@ -25,6 +25,12 @@ on:
      'tools/server/**.*'
    ]

+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/server-sanitize.yml'
+    ]
+
 env:
  LLAMA_ARG_LOG_COLORS: 1
  LLAMA_ARG_LOG_PREFIX: 1
@@ -90,15 +96,18 @@ jobs:

      - name: Python setup
        id: setup_python
-        uses: actions/setup-python@v6
-        with:
-          python-version: '3.11'
-          pip-install: -r tools/server/tests/requirements.txt
+        uses: actions/setup-python@v7
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m venv .venv
+          .venv/bin/pip install -r tools/server/tests/requirements.txt

      - name: Tests
        id: server_integration_tests
        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
        run: |
+          source .venv/bin/activate
          cd tools/server/tests
          export ${{ matrix.extra_args }}
          pytest -v -x -m "not slow"
@@ -107,6 +116,7 @@ jobs:
        id: server_integration_tests_slow
        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
+          source .venv/bin/activate
          cd tools/server/tests
          export ${{ matrix.extra_args }}
          SLOW_TESTS=1 pytest -v -x
@@ -12,7 +12,7 @@
 [![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)

-[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev branches](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-features.md) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)

 </div>

@@ -21,11 +21,18 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/

 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.

+### AI-powered code scan
+
+llama.cpp has an AI security scanner that scans the code periodically. The full prompts and tool set can be found in [ggml-org/security-scan-prompt](https://github.com/ggml-org/security-scan-prompt).
+
+We greatly appreciate reports that reflect genuine research effort, and we are happy to spend our time reviewing them. Findings that an autonomous AI agent can surface on its own add little on top of the scans we already run.
+
 ### Requirements

 Before submitting your report, ensure you meet the following requirements:

 - You have read this policy and fully understand it.
+- You have searched for existing discussions of the issue. If it has already been reported, your report will likely be rejected as a duplicate.
 - AI is only permitted in an assistive capacity as stated in [AGENTS.md](AGENTS.md). We do not accept reports that are written exclusively by AI.
 - Your report must include a working Proof-of-Concept in the form of a script and/or attached files.

@@ -46,6 +53,8 @@ Only vulnerabilities that fall within these parts of the project are considered

 Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.

+Denial-of-Service (DoS) bugs are generally not treated as vulnerabilities. We don't reject them outright, but we look at them case-by-case and only accept those that are genuinely worth fixing.
+
 For vulnerabilities that fall within the `vendor` directory, please report them directly to the third-party project.

 ## Using llama.cpp securely
@@ -17,7 +17,6 @@ LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
 GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
-GGML_METAL_USE_BF16=ON
 GGML_OPENMP=OFF

 COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
@@ -44,7 +43,6 @@ COMMON_CMAKE_ARGS=(
    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
    -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
    -DGGML_METAL=${GGML_METAL}
-    -DGGML_METAL_USE_BF16=${GGML_METAL_USE_BF16}
    -DGGML_NATIVE=OFF
    -DGGML_OPENMP=${GGML_OPENMP}
 )
@@ -10,6 +10,9 @@
 # # with CUDA support
 # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with ROCm support
+# GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 # # with SYCL support
 # GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
@@ -89,7 +92,7 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
 fi

 if [ ! -z ${GG_BUILD_ROCM} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_HIP=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON"
    if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
        echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
        exit 1
@@ -640,39 +643,52 @@ function gg_sum_rerank_tiny {

 function gg_check_build_requirements {
    if ! command -v git &> /dev/null; then
-        gg_printf 'git not found, please install'
+        gg_printf 'git not found, please install\n'
+        exit 1
    fi

    if ! command -v git-lfs &> /dev/null; then
-        gg_printf 'git-lfs not found, please install'
+        gg_printf 'git-lfs not found, please install\n'
+        exit 1
+    fi
+
+    if ! git config --get filter.lfs.clean &> /dev/null; then
+        gg_printf 'git-lfs not initialized, please run `git lfs install`\n'
+        exit 1
    fi

    if ! command -v wget &> /dev/null; then
-        gg_printf 'wget not found, please install'
+        gg_printf 'wget not found, please install\n'
+        exit 1
    fi

    if ! command -v python3 &> /dev/null; then
-        gg_printf 'python3 not found, please install'
+        gg_printf 'python3 not found, please install\n'
+        exit 1
    fi

    if ! command -v pip3 &> /dev/null; then
-        gg_printf 'pip3 not found, please install'
+        gg_printf 'pip3 not found, please install\n'
+        exit 1
    fi

    if ! python3 -m ensurepip --help &> /dev/null; then
-        gg_printf 'ensurepip not found, please install python3-venv package'
+        gg_printf 'ensurepip not found, please install python3-venv package\n'
+        exit 1
    fi

    if ! command -v cmake &> /dev/null; then
-        gg_printf 'cmake not found, please install'
+        gg_printf 'cmake not found, please install\n'
+        exit 1
    fi

    if ! command -v ccache &> /dev/null; then
-        gg_printf 'ccache not found, please consider installing for faster builds'
+        gg_printf 'ccache not found, please consider installing for faster builds\n'
    fi

    if ! command -v ctest &> /dev/null; then
-        gg_printf 'ctest not found, please install'
+        gg_printf 'ctest not found, please install\n'
+        exit 1
    fi
 }

@@ -0,0 +1,26 @@
+# Used to cross-compile ggml-cuda for Windows ARM64 on an x64 Windows host.
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+if ( DEFINED CUDAToolkit_ROOT )
+    file( TO_CMAKE_PATH "${CUDAToolkit_ROOT}" CUDA_ROOT )
+elseif ( DEFINED ENV{CUDA_PATH} )
+    file( TO_CMAKE_PATH "$ENV{CUDA_PATH}" CUDA_ROOT )
+else()
+    message( FATAL_ERROR "Set CUDAToolkit_ROOT or CUDA_PATH to a Windows CUDA Toolkit with ARM64 target libraries" )
+endif()
+
+if ( DEFINED ENV{VCToolsInstallDir} )
+    file( TO_CMAKE_PATH "$ENV{VCToolsInstallDir}" MSVC_TOOLS_ROOT )
+    set( CMAKE_CUDA_HOST_COMPILER "${MSVC_TOOLS_ROOT}/bin/Hostx64/arm64/cl.exe" CACHE FILEPATH "" )
+endif()
+
+set( CMAKE_CUDA_COMPILER "${CUDA_ROOT}/bin/nvcc.exe" CACHE FILEPATH "" )
+set( CMAKE_CUDA_FLAGS_INIT "-target-dir=arm64" )
+
+# FindCUDAToolkit selects lib/x64 from the host architecture on Windows.
+set( CUDA_CUDART              "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
+set( CUDA_cudart_LIBRARY      "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
+set( CUDA_cublas_LIBRARY      "${CUDA_ROOT}/lib/arm64/cublas.lib"   CACHE FILEPATH "" )
+set( CUDA_cublasLt_LIBRARY    "${CUDA_ROOT}/lib/arm64/cublasLt.lib" CACHE FILEPATH "" )
+set( CUDA_cuda_driver_LIBRARY "${CUDA_ROOT}/lib/arm64/cuda.lib"     CACHE FILEPATH "" )
@@ -3308,6 +3308,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.server_tools = parse_csv_row(value);
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS"));
+    add_opt(common_arg(
+        {"--tools-runtime"}, "OPTION",
+        "experimental: run tools in a separate runtime environment (default: none, use host environment)\n"
+        "available options:\n"
+        "  'docker:<image>', 'podman:<image>': spin up a new container and reuse it for all invocations, clean up on server exit\n"
+        "  'docker-container:<id>', 'podman-container:<id>': use an existing container by ID, won't stop on server exit\n"
+        "  'ssh:<target>': run tools on a remote POSIX host over SSH, key-based auth and a trusted host key are required\n",
+        [](common_params & params, const std::string & value) {
+            params.server_tools_runtime = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS_RUNTIME"));
    add_opt(common_arg(
        {"--mcp-servers-config"}, "PATH",
        "experimental: path to JSON file with MCP server definitions (Cursor-compatible format) - do not enable in untrusted environments (default: none)\n"
@@ -3086,6 +3086,151 @@ static common_chat_params common_chat_params_init_minicpm5(const common_chat_tem
    return data;
 }

+// An assistant turn is rendered as one or more messages, each
+// "<|start|>assistant to=<recipient><|message|>{content}{END}" where END is
+// <|eom|> (more messages follow) or <|eot|> (end of turn):
+//   - chain-of-thought: to=self, terminated by <|eom|>
+//   - final answer:     to=user, terminated by <|eot|>
+// The generation prompt is just "<|start|>assistant"; the model emits its own
+// " to=...<|message|>".
+static common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template &          tmpl,
+                                                               const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = "<|start|>assistant";
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+
+    data.preserved_tokens = {
+        "<|start|>", "<|message|>", "<|eom|>", "<|eot|>",
+        // ATEM tool-call markup emitted on " to=<tool>" turns.
+        "<atem:function_calls>", "<atem:invoke", "<atem:parameter", "</atem:parameter>",
+        "</atem:invoke>", "</atem:function_calls>",
+    };
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
+        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
+        { COMMON_CHAT_ROLE_TOOL,      "<|start|>tool"      },
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "<|start|>assistant to=self<|message|>" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<|eom|><|start|>assistant to=user<|message|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+
+    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    // Constrained grammar whenever tools are offered.
+    auto include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto start = p.rule("start", p.literal("<|start|>assistant"));
+
+        if (!extract_reasoning && !include_grammar) {
+            return start + p.content(p.rest());
+        }
+
+        if (extract_reasoning) {
+            p.rule("analysis", p.literal(" to=self<|message|>") + p.reasoning(p.until("<|eom|>")) + p.literal("<|eom|>"));
+        } else {
+            p.rule("analysis", p.literal(" to=self<|message|>") + p.content(p.until("<|eom|>")) + p.literal("<|eom|>"));
+        }
+        auto analysis = p.ref("analysis");
+
+        auto recipient  = p.optional(p.literal(" to=user"));
+        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") + p.content(p.until("<|eot|>")));
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto string_value = p.ac(
+                p.tool_arg_string_value(p.until("</atem:parameter>")) + p.tool_arg_close(p.literal("</atem:parameter>")),
+                "</atem:parameter>");
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto &      function = tool.at("function");
+                const std::string name     = function.at("name");
+                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto args = p.eps();
+                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
+                    auto schema_info = common_schema_info();
+                    schema_info.resolve_refs(params);
+
+                    auto arg_choice = p.choice();
+                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
+                        auto value_parser = p.eps();
+                        if (schema_info.resolves_to_string(prop_schema)) {
+                            value_parser = string_value;
+                        } else {
+                            value_parser = p.tool_arg_json_value(
+                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false))
+                                + p.tool_arg_close(p.literal("</atem:parameter>"));
+                        }
+
+                        auto arg_rule = p.tool_arg(
+                            p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
+                            value_parser);
+
+                        arg_choice |= arg_rule;
+                    }
+                    args = p.zero_or_more(arg_choice + p.space());
+                }
+
+                auto tool_parser = p.tool(
+                    p.tool_open(p.literal(" to=") + p.until("<|message|>") +
+                                p.literal("<|message|><atem:function_calls>") + p.space() +
+                                p.literal("<atem:invoke name=\"") + p.tool_name(p.literal(name)) + p.literal("\">") + p.space())
+                    << p.tool_args(args)
+                    << p.tool_close(p.literal("</atem:invoke>") + p.space() + p.literal("</atem:function_calls>")));
+
+                tool_choice |= p.rule("tool-" + name, tool_parser);
+            });
+
+            auto tool_calls = inputs.parallel_tool_calls
+                ? p.trigger_rule("tool-call", tool_choice + p.zero_or_more(p.literal("<|eom|>") + start + tool_choice))
+                : p.trigger_rule("tool-call", tool_choice);
+
+
+            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+                return p.zero_or_more(start + analysis) + start + tool_calls;
+            }
+            return p.zero_or_more(start + analysis) + start + (tool_calls | final_msg);
+        }
+
+        return p.zero_or_more(start + analysis) + start + final_msg;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+              "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
+        };
+    }
+
+    return data;
+}
+
 static json common_chat_extra_context() {
    json ctx = json::object();
    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
@@ -3114,6 +3259,12 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
        return common_chat_params_init_gpt_oss(tmpl, params);
    }

+    // Muse Glimmer format using " to=<recipient>" recipients and <|eom|>/<|eot|> message terminators.
+    if (src.find("<atem:function_calls>") != std::string::npos && src.find("<|eom|>") != std::string::npos) {
+        LOG_DBG("Using specialized template: Muse Glimmer\n");
+        return common_chat_params_init_muse_glimmer(tmpl, params);
+    }
+
    // Functionary v3.2 - uses recipient-based format with >>>recipient\n{content}
    // Detection: template has ">>>all" for content and ">>>" prefix for tool calls
    if (src.find(">>>all") != std::string::npos && src.find(">>>${recipient}") != std::string::npos) {
@@ -1639,6 +1639,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.n_seq_max         = params.n_parallel;
    cparams.n_rs_seq          = params.speculative.need_n_rs_seq();
    cparams.n_outputs_max     = std::max(params.n_outputs_max, 0);
+    cparams.n_outputs_max_per_seq = std::max(params.n_outputs_max_per_seq, 0);
    cparams.n_batch           = params.n_batch;
    cparams.n_ubatch          = params.n_ubatch;
    cparams.n_threads         = params.cpuparams.n_threads;
@@ -447,6 +447,7 @@ struct common_params {
    int32_t n_parallel            =     1; // number of parallel sequences to decode
    int32_t n_sequences           =     1; // number of sequences to decode
    int32_t n_outputs_max         =     0; // max outputs in a batch (0 = n_batch)
+    int32_t n_outputs_max_per_seq =     1; // max outputs per sequence
    int32_t grp_attn_n            =     1; // group-attention factor
    int32_t grp_attn_w            =   512; // group-attention width
    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -655,6 +656,7 @@ struct common_params {

    // enable built-in tools
    std::vector<std::string> server_tools;
+    std::string server_tools_runtime;

    // MCP server configs (Cursor-compatible JSON)
    std::string mcp_servers_config;   // path to JSON file with MCP server definitions
@@ -136,7 +136,10 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
        devs.push_back(llama_model_get_device(model, i));
    }

-    hp_ngl         = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);
+    hp_ngl         = llama_model_n_layer(model);
+    if (mparams->load_mtp) {
+        hp_ngl    += llama_model_n_layer_nextn(model);
+    }
    hp_n_ctx_train = llama_model_n_ctx_train(model);
    hp_n_expert    = llama_model_n_expert(model);

@@ -116,6 +116,8 @@ static llama_sampler_i llama_sampler_llg_i = {
    /* .backend_accept    = */ NULL,
    /* .backend_apply     = */ NULL,
    /* .backend_set_input = */ NULL,
+    /* .backend_reset     = */ NULL,
+    /* .copy_state        = */ NULL,
 };

 static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,
@@ -217,6 +217,8 @@ static struct llama_sampler_i common_reasoning_budget_i = {
    /* .backend_accept    = */ nullptr,
    /* .backend_apply     = */ nullptr,
    /* .backend_set_input = */ nullptr,
+    /* .backend_reset     = */ nullptr,
+    /* .copy_state        = */ nullptr,
 };

 static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
@@ -518,6 +518,26 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
    };
 }

+void common_sampler_copy(const common_sampler * src, common_sampler * dst) {
+    if (!src || !dst || src == dst) {
+        return;
+    }
+
+    GGML_ASSERT((src->grmr == nullptr) == (dst->grmr == nullptr));
+    GGML_ASSERT((src->rbudget == nullptr) == (dst->rbudget == nullptr));
+
+    llama_sampler_copy(src->grmr,    dst->grmr);
+    llama_sampler_copy(src->rbudget, dst->rbudget);
+    llama_sampler_copy(src->chain,   dst->chain);
+
+    dst->params     = src->params;
+    dst->prev       = src->prev;
+    dst->cur        = src->cur;
+    dst->cur_p      = src->cur_p;
+    dst->cur_p.data = src->cur_p.data ? dst->cur.data() : nullptr; // re-point to dst's buffer
+    dst->t_total_us = src->t_total_us;
+}
+
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
    // TODO: measure grammar performance

@@ -47,6 +47,7 @@ void common_sampler_free(struct common_sampler * gsmpl);
 void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool is_generated);
 void                    common_sampler_reset (struct common_sampler * gsmpl);
 struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
+void                    common_sampler_copy  (const struct common_sampler * src, struct common_sampler * dst);

 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
@@ -1032,7 +1032,14 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            return true;
        }

-        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
+        // Target prefill may contain token IDs or multimodal embeddings. Both
+        // produce the target-layer features used to seed the draft KV cache, so
+        // skipping the embedding batches leaves a hole in the draft's cache and
+        // the next injection fails to initialize.
+        // TODO: revisit after https://github.com/ggml-org/llama.cpp/pull/24669 is merged
+        const bool has_tokens     = batch_in.token != nullptr;
+        const bool has_embeddings = batch_in.embd  != nullptr;
+        if (has_tokens == has_embeddings) {
            return true;
        }

@@ -2292,6 +2299,7 @@ common_params common_base_params_to_speculative(const common_params & params) {
    result.cache_type_k  = params_spec.cache_type_k;
    result.cache_type_v  = params_spec.cache_type_v;
    result.n_outputs_max = params.n_parallel;
+    result.n_outputs_max_per_seq = 1;

    return result;
 }
@@ -2377,6 +2385,17 @@ common_speculative_init_result_ptr common_speculative_init_from_params(common_pa
    return std::make_unique<common_speculative_init_result>(params, model_tgt, ctx_tgt);
 }

+common_speculative_output_limits common_speculative_get_output_limits(
+        int32_t n_batch, int32_t n_parallel, int32_t n_draft) {
+    const int64_t per_seq = 1 + (int64_t) std::max(0, n_draft);
+    const int64_t total   = (int64_t) n_parallel * per_seq;
+
+    return {
+        /* .total   = */ (int32_t) std::min<int64_t>(n_batch, total),
+        /* .per_seq = */ (int32_t) std::min<int64_t>(n_batch, per_seq),
+    };
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -25,6 +25,15 @@ int32_t common_speculative_n_max(const common_params_speculative * spec);

 common_params common_base_params_to_speculative(const common_params & params);

+struct common_speculative_output_limits {
+    int32_t total;
+    int32_t per_seq;
+};
+
+// return the output limits needed for speculative decoding
+common_speculative_output_limits common_speculative_get_output_limits(
+        int32_t n_batch, int32_t n_parallel, int32_t n_draft);
+
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq);

 void common_speculative_free(common_speculative * spec);
@@ -70,6 +70,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Exaone4ForCausalLM": "exaone",
    "ExaoneForCausalLM": "exaone",
    "ExaoneMoEForCausalLM": "exaone",
+    "ExaoneMoeForCausalLM": "exaone",
    "FalconForCausalLM": "falcon",
    "FalconH1ForCausalLM": "falcon_h1",
    "FalconMambaForCausalLM": "mamba",
@@ -102,6 +103,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteMoeForCausalLM": "granite",
    "GraniteMoeHybridForCausalLM": "granite",
    "GraniteMoeSharedForCausalLM": "granite",
+    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
    "Grok1ForCausalLM": "grok",
@@ -181,6 +183,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Olmo3ForCausalLM": "olmo",
    "OlmoForCausalLM": "olmo",
    "OlmoeForCausalLM": "olmo",
+    "MuseGlimmerAssistantModel": "muse_glimmer",
+    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
    "OpenELMForCausalLM": "openelm",
    "OrionForCausalLM": "orion",
    "PLMForCausalLM": "plm",
@@ -296,6 +300,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "MiniCPMV4_6ForConditionalGeneration": "minicpm",
    "Mistral3ForConditionalGeneration": "llava",
    "NemotronH_Nano_VL_V2": "nemotron",
+    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
    "PaddleOCRVisionModel": "ernie",
    "Phi4ForCausalLMV": "phi",
    "Qwen2AudioForConditionalGeneration": "ultravox",
@@ -17,8 +17,11 @@ from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logg
 from .qwen import QwenModel


-@ModelBase.register("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
+@ModelBase.register("DeepseekOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
+    # HF dynamic_preprocess() max_num, which differs per model
+    preproc_max_tiles = 9
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
@@ -43,6 +46,9 @@ class DeepseekOCRVisionModel(MmprojModel):
        # @bluebread: there's no window_size in config but just add it here anyway
        self.gguf_writer.add_vision_window_size(self.hparams.get("window_size", 14))

+        self.gguf_writer.add_vision_preproc_min_tiles(2)
+        self.gguf_writer.add_vision_preproc_max_tiles(self.preproc_max_tiles)
+
        # SAM configuration
        sam_hparams = hparams['sam']
        self.gguf_writer.add_vision_sam_layers_count(sam_hparams['layers'])
@@ -93,8 +99,15 @@ class DeepseekOCRVisionModel(MmprojModel):
        return super().filter_tensors((name, gen))


+@ModelBase.register("UnlimitedOCRForCausalLM")
+class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 32
+
+
@ModelBase.register("DeepseekOCR2ForCausalLM")
 class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 6
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2
@@ -520,6 +533,13 @@ class DeepseekV4Model(TextModel):
        for key, value in raw_hparams.items():
            self.hparams.setdefault(key, value)

+        # workaround for special rope_parameters (main/compress) in transformers 5.x
+        if self.rope_parameters.get("full_attention", self.rope_parameters).get("rope_type") is None:
+            if (rope_scaling := raw_hparams.get("rope_scaling")) is not None:
+                if "rope_type" not in rope_scaling and (rope_type := rope_scaling.get("type")) is not None:
+                    rope_scaling["rope_type"] = rope_type
+                self.rope_parameters.update(**rope_scaling)
+
        self.block_count = self.hparams["num_hidden_layers"]
        if self.mtp_only:
            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
@@ -123,7 +123,9 @@ class Exaone4Model(TextModel):
                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))


-@ModelBase.register("ExaoneMoEForCausalLM")
+# note: transformers >= 5.1 renamed the class to "ExaoneMoeForCausalLM" (lowercase 'e'),
+#       so accept both spellings - LG AI have updated the configs of already-released models
+@ModelBase.register("ExaoneMoEForCausalLM", "ExaoneMoeForCausalLM")
 class ExaoneMoEModel(Exaone4Model):
    model_arch = gguf.MODEL_ARCH.EXAONE_MOE

@@ -123,6 +123,166 @@ class GraniteMoeModel(GraniteModel):
        yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("GraniteSwitchForCausalLM")
+class GraniteSwitchModel(GraniteMoeModel):
+    """Dense, all-attention Granite with N per-token embedded LoRA adapters, stacked
+    over the adapter dim with a zero adapter at slot 0 (N = num_adapters + 1)."""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWITCH
+
+    # permute q/k per-slice below (NORM-rope layout), not via the parent's auto-permute
+    undo_permute = False
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # the weightless switch reserves one cache slot: one fewer block than num_hidden_layers
+        self.block_count = self.block_count - 1
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self._n_adapters = int(self.hparams["num_adapters"])
+        self._max_lora_rank = int(self.hparams["max_lora_rank"])
+        self._n_slots = self._n_adapters + 1  # +1 for the zero slot at index 0
+
+        n_head = int(self.hparams["num_attention_heads"])
+        n_kv_head = int(self.hparams["num_key_value_heads"])
+        head_dim = (
+            self.hparams.get("projection_head_dim")
+            or self.hparams.get("head_dim")
+            or (self.hparams["hidden_size"] // n_head)
+        )
+        self._n_head = n_head
+        self._n_kv_head = n_kv_head
+        self._head_dim = int(head_dim)
+        self._q_size = n_head * self._head_dim
+        self._kv_size = n_kv_head * self._head_dim
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # dense: pin expert_used_count to 0 (config carries a leftover num_experts_per_tok)
+        if not self.hparams.get("num_local_experts"):
+            self.gguf_writer.add_expert_used_count(0)
+
+        self.gguf_writer.add_adapter_count(self._n_adapters)
+        self.gguf_writer.add_adapter_lora_rank(self._max_lora_rank)
+        self.gguf_writer.add_adapter_token_ids_activate(self.hparams["adapter_token_ids"])
+        self.gguf_writer.add_adapter_token_ids_substitute(self.hparams["adapter_substitute_token_ids"])
+        router_gain = float(self.hparams.get("control_token_gain", 15.0))
+        self.gguf_writer.add_adapter_router_gain(router_gain)
+        logger.info("gguf: (graniteswitch) num_adapters=%s max_lora_rank=%s n_slots=%s router_gain=%s", self._n_adapters, self._max_lora_rank, self._n_slots, router_gain)
+
+    def _lora_a(self, data: Tensor) -> Tensor:
+        # on-disk A: [n_adapters, 1, max_rank, in] -> [n_adapters+1, max_rank, in]
+        a = data.squeeze(1)
+        zero = torch.zeros_like(a[:1])
+        return torch.cat([zero, a], dim=0).contiguous()
+
+    def _lora_b(self, data: Tensor, permute_n_head: int | None = None) -> Tensor:
+        # on-disk B: [n_adapters, 1, out, max_rank] -> [n_adapters+1, out, max_rank]
+        b = data.squeeze(1)
+        if permute_n_head is not None:
+            # permute each adapter's B output rows to match the permuted q/k base
+            b = torch.stack([self.permute(b[i], permute_n_head, permute_n_head) for i in range(b.shape[0])], dim=0)
+        zero = torch.zeros_like(b[:1])
+        return torch.cat([zero, b], dim=0).contiguous()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+
+        # skip the weightless switch + control-token buffers (rebuilt at load time)
+        bare = name.split(".")[-1]
+        if (
+            name.startswith("model.switch.") or name.startswith("switch.")
+            or bare in ("adapter_token_ids", "control_to_substitute_lut")
+        ):
+            return
+
+        if "self_attn.qkv_proj" in name:
+            if name.endswith("base_layer.weight"):
+                # fused [q|k|v] rows: permute q/k row-blocks for ggml's NORM-rope layout
+                q, k, v = data_torch.split([self._q_size, self._kv_size, self._kv_size], dim=0)
+                q = self.permute(q, self._n_head, self._n_head)
+                k = self.permute(k, self._n_kv_head, self._n_kv_head)
+                fused = torch.cat([q, k, v], dim=0)
+                yield (self.format_tensor_name(T.ATTN_QKV, bid), fused)
+                return
+            if "lora_A_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key = {0: T.ATTN_Q, 1: T.ATTN_K, 2: T.ATTN_V}[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if "lora_B_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key, ph = {
+                    0: (T.ATTN_Q, self._n_head),
+                    1: (T.ATTN_K, self._n_kv_head),
+                    2: (T.ATTN_V, None),
+                }[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_b"), self._lora_b(data_torch, ph))
+                return
+            raise ValueError(f"Unexpected qkv_proj tensor: {name}")
+
+        if "self_attn.o_proj" in name:
+            if name.endswith("base_layer.weight"):
+                yield (self.format_tensor_name(T.ATTN_OUT, bid), data_torch)
+                return
+            if name.endswith("lora_A"):
+                yield (self.format_tensor_name(T.ATTN_OUT, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if name.endswith("lora_B"):
+                yield (self.format_tensor_name(T.ATTN_OUT, bid, suffix=".lora_b"), self._lora_b(data_torch))
+                return
+            raise ValueError(f"Unexpected o_proj tensor: {name}")
+
+        if "shared_mlp.input_linear" in name:
+            ffn = self.hparams["shared_intermediate_size"]
+            if name.endswith("base_layer.weight"):
+                gate, up = data_torch.split([ffn, ffn], dim=0)
+                yield (self.format_tensor_name(T.FFN_GATE, bid), gate)
+                yield (self.format_tensor_name(T.FFN_UP, bid), up)
+                return
+            if "lora_A_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key = {0: T.FFN_GATE, 1: T.FFN_UP}[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if "lora_B_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key = {0: T.FFN_GATE, 1: T.FFN_UP}[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_b"), self._lora_b(data_torch))
+                return
+            raise ValueError(f"Unexpected shared_mlp.input_linear tensor: {name}")
+
+        if "shared_mlp.output_linear" in name:
+            if name.endswith("base_layer.weight"):
+                yield (self.format_tensor_name(T.FFN_DOWN, bid), data_torch)
+                return
+            if name.endswith("lora_A"):
+                yield (self.format_tensor_name(T.FFN_DOWN, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if name.endswith("lora_B"):
+                yield (self.format_tensor_name(T.FFN_DOWN, bid, suffix=".lora_b"), self._lora_b(data_torch))
+                return
+            raise ValueError(f"Unexpected shared_mlp.output_linear tensor: {name}")
+
+        if bid is not None and ".layers." in name and (
+            "input_layernorm" in name or "post_attention_layernorm" in name
+        ):
+            key = T.ATTN_NORM if "input_layernorm" in name else T.FFN_NORM
+            yield (self.format_tensor_name(key, bid), data_torch)
+            return
+
+        if name in ("model.embed_tokens.weight", "embed_tokens.weight"):
+            yield (self.format_tensor_name(T.TOKEN_EMBD), data_torch)
+            return
+        if name in ("model.norm.weight", "norm.weight"):
+            yield (self.format_tensor_name(T.OUTPUT_NORM), data_torch)
+            return
+        if name == "lm_head.weight":
+            return  # tied to token_embd
+
+        raise ValueError(f"graniteswitch: unhandled tensor {name!r} (bid={bid})")
+
+
@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
 class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
    """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
@@ -0,0 +1,179 @@
+from __future__ import annotations
+
+import json
+from typing import Any, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import MmprojModel, ModelBase, TextModel, gguf
+
+
+def _unpermute_for_rope(tensor: "Tensor", n_heads: int) -> "Tensor":
+    """Invert transformers' `_permute_for_rope`: HF stores Q/K in rotate_half layout,
+    llama.cpp consumes the interleaved (NORM) layout."""
+    if tensor.ndim == 2:
+        dim1, dim2 = tensor.shape
+        return tensor.view(n_heads, 2, dim1 // n_heads // 2, dim2).transpose(1, 2).reshape(dim1, dim2)
+    if tensor.ndim == 1:
+        (dim1,) = tensor.shape
+        return tensor.view(n_heads, 2, dim1 // n_heads // 2).transpose(1, 2).reshape(dim1)
+    raise ValueError(f"_unpermute_for_rope: unexpected shape {tuple(tensor.shape)}")
+
+
+@ModelBase.register("MuseGlimmerForConditionalGeneration")
+class MuseGlimmerModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.MUSE_GLIMMER
+
+    def norm_shift(self, name: str) -> float:
+        # All four layer norms use 1, the final norm uses 0.
+        return 1.0 if name.endswith("layernorm.weight") else 0.0
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+        from transformers import AutoTokenizer
+        tok = AutoTokenizer.from_pretrained(self.dir_model)
+        eot_id = tok.convert_tokens_to_ids("<|eot|>")
+        if isinstance(eot_id, int) and eot_id >= 0:
+            self.gguf_writer.add_eot_token_id(eot_id)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        self.gguf_writer.add_final_logit_softcapping(hparams["final_logit_softcapping"])
+        self.gguf_writer.add_logit_scale(hparams["output_multiplier"])
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in hparams["layer_types"]])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        shift = self.norm_shift(name)
+        if shift != 0.0:
+            data_torch = data_torch + shift
+
+        # Invert transformers' `_permute_for_rope` on Q/K, we keep ggml's NORM (interleaved) rope
+        if ".self_attn.q_proj." in name:
+            data_torch = _unpermute_for_rope(data_torch, int(self.hparams["num_attention_heads"]))
+        elif ".self_attn.k_proj." in name:
+            data_torch = _unpermute_for_rope(data_torch, int(self.hparams["num_key_value_heads"]))
+
+        # Synthesize QK-norm weights to absorb qk_scale_factor.
+        # MuseGlimmer implementation: scaleless RMSNorm followed by qk_scale_factor..
+        if bid is not None and name.endswith(f"model.layers.{bid}.self_attn.q_proj.weight"):
+            head_dim = self.hparams["head_dim"]
+            q_scale = float(self.hparams["qk_scale_factor"])
+            yield (
+                self.map_tensor_name(f"model.layers.{bid}.self_attn.q_norm.weight"),
+                torch.full((head_dim,), q_scale, dtype=torch.float32),
+            )
+            yield (
+                self.map_tensor_name(f"model.layers.{bid}.self_attn.k_norm.weight"),
+                torch.ones((head_dim,), dtype=torch.float32),
+            )
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("MuseGlimmerForConditionalGeneration")
+class MuseGlimmerVisionModel(MmprojModel):
+    def get_vision_config(self) -> dict[str, Any] | None:
+        c = self.global_config.get("vision_config")
+        if not c:
+            return None
+        # MuseGlimmer actually uses dynamic size, initialize with nominal size
+        image_size = c["pos_emb_height"] * c["patch_size"] * c["merge_size"]
+        return {**c, "image_size": image_size}
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        c = self.hparams_vision  # enriched vision_config from get_vision_config()
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MUSE_GLIMMER)
+        self.gguf_writer.add_vision_attention_layernorm_eps(float(c["layer_norm_eps"]))
+        self.gguf_writer.add_vision_spatial_merge_size(int(c["merge_size"]))
+
+    @classmethod
+    def filter_tensors(cls, item):
+        name, gen = item
+        keep = ("model.vision_tower.", "model.vision_adapter.", "model.vision_projection.")
+        if not any(name.startswith(k) for k in keep):
+            return None
+        return super().filter_tensors((name, gen))
+
+    # 3-layer projector MLP
+    _MM_MLP_MAP = {
+        "model.vision_adapter.fc1": (gguf.MODEL_TENSOR.V_MMPROJ, 0),
+        "model.vision_adapter.fc2": (gguf.MODEL_TENSOR.V_MMPROJ, 1),
+        "model.vision_projection":  (gguf.MODEL_TENSOR.V_MMPROJ, 2),
+    }
+
+    def modify_tensors(self, data_torch, name, bid):
+        assert self.hparams_vision is not None
+        if ".attn.q_proj." in name or ".attn.k_proj." in name:
+            n_heads = int(self.hparams_vision["num_attention_heads"])
+            data_torch = _unpermute_for_rope(data_torch, n_heads)
+        # Lay out the pt=2 temporal slabs of the patch embedding as a conv2d for build_inp()
+        if name.endswith("patch_embedder.patch_embedding.weight"):
+            n_embd = data_torch.shape[0]
+            pt = int(self.hparams_vision["patch_temporal"])
+            ps = int(self.hparams_vision["patch_size"])
+            data_torch = data_torch.view(n_embd, pt, 3, ps, ps).sum(dim=1)  # (n_embd, 3, ps, ps)
+        stem, _, suffix = name.rpartition(".")
+        if stem in self._MM_MLP_MAP:
+            tensor_key, idx = self._MM_MLP_MAP[stem]
+            yield (self.format_tensor_name(tensor_key, bid=idx, suffix="." + suffix), data_torch)
+            return
+        yield (self.map_tensor_name(name), data_torch)
+
+
+@ModelBase.register("MuseGlimmerAssistantModel")
+class MuseGlimmerAssistantModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    def set_vocab(self):
+        if self.target_model_dir is None:
+            raise ValueError(
+                "MuseGlimmerAssistant (DFlash drafter) requires --target-model-dir pointing to the "
+                "target MuseGlimmer HF directory"
+            )
+
+        original_dir = self.dir_model
+        self.dir_model = self.target_model_dir
+
+        from . import get_model_class
+        with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
+            target_arch = json.load(f)["architectures"][0]
+        target_cls = get_model_class(target_arch)
+        if target_cls is not type(self):
+            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
+        else:
+            super().set_vocab()
+
+        self.dir_model = original_dir
+
+        mask_token_id = self.hparams.get("mask_token_id")
+        if mask_token_id is not None:
+            self.gguf_writer.add_mask_token_id(int(mask_token_id))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        h = self.hparams
+
+        self.gguf_writer.add_block_size(int(h["block_size"]))
+
+        # dflash.target_layers[k] refers to the inputs going into the ith layer, which come from the (i-1)th layer's output.
+        # The transformers configuration refers to the outputs being recorded.
+        self.gguf_writer.add_target_layers([int(x) + 1 for x in h["target_layer_ids"]])
+
+        if h.get("sliding_window") and h.get("layer_types"):
+            self.gguf_writer.add_sliding_window(int(h["sliding_window"]))
+            self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in h["layer_types"]])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # DFlash defaults to NEOX (rotate_half) rope, matching transformers HF layout for Q/K, QK-norms
+        # no permutation needed.
+        yield (self.map_tensor_name(name), data_torch)
@@ -197,6 +197,7 @@ class NemotronHModel(GraniteHybridModel):
    """Hybrid mamba2/attention model from NVIDIA"""
    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
    is_moe: bool = False
+    supports_mtp_export = True

    def __init__(self, *args, **kwargs):
        # We have to determine the correct model architecture (MoE vs non-MoE) before
@@ -236,6 +237,25 @@ class NemotronHModel(GraniteHybridModel):
            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "mamba"]
            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]

+        # `--no-mtp` drops it entirely; `--mtp` exports only the MTP head
+        self._mtp_bid: int | None = None
+        if self.is_moe and not self.no_mtp:
+            n_nextn = self.hparams.get("num_nextn_predict_layers", 0) or 0
+            if n_nextn > 0:
+                assert n_nextn == 1, (
+                    "NemotronH MTP conversion currently supports num_nextn_predict_layers == 1"
+                )
+                self._mtp_bid = self.block_count
+                self.block_count += 1
+                # The folded MTP block carries both an attention sub-layer and a
+                # MoE sub-layer, so register it as both so the per-layer metadata arrays cover it
+                self._attn_layers.append(self._mtp_bid)
+                self._mlp_layers.append(self._mtp_bid)
+                self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        if self.mtp_only and self._mtp_bid is None:
+            raise ValueError("--mtp was requested, but this model does not contain a supported MTP head")
+
    def get_attn_layers(self):
        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
        if pattern is None:
@@ -246,6 +266,36 @@ class NemotronHModel(GraniteHybridModel):

        return [i for i, val in enumerate(pattern) if val == "attention"]

+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if name.startswith("mtp."):
+            # --no-mtp: drop the MTP head entirely
+            if cls.no_mtp:
+                return None
+        elif cls.mtp_only:
+            # --mtp: export the MTP head plus the tensors it shares with the target model
+            keep = name in (
+                "backbone.embeddings.weight",
+                "backbone.norm_f.weight",
+                "lm_head.weight",
+            )
+            if not keep:
+                return None
+        return super().filter_tensors((name, gen))
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
    def set_gguf_parameters(self):
        super().set_gguf_parameters()

@@ -284,6 +334,10 @@ class NemotronHModel(GraniteHybridModel):
            if (latent_size := self.hparams.get("moe_latent_size")) is not None:
                self.gguf_writer.add_moe_latent_size(latent_size)

+        # MTP head: number of trailing NextN blocks
+        if self._mtp_bid is not None:
+            self.gguf_writer.add_nextn_predict_layers(self.hparams["num_nextn_predict_layers"])
+
    def set_vocab(self):
        # The NemotronH config uses pattern characters (e.g. '-') that may not
        # be supported by the installed transformers version. AutoTokenizer
@@ -350,15 +404,24 @@ class NemotronHModel(GraniteHybridModel):
        if not self.is_moe:
            self.gguf_writer.add_add_bos_token(True)

-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if self.is_moe and bid is not None:
-            # Skip Multi-Token Prediction (MTP) tensors. These are used for
-            # for speculative decoding but we don't include them in this model
-            # conversion. See https://github.com/ggml-org/llama.cpp/pull/18886
-            if name.startswith("mtp."):
-                logger.info(f"gguf: Skipping MTP (Speculative) layer: {name}")
-                return
+    _MTP_SPECIAL_RENAMES = {
+        "mtp.layers.0.enorm.weight":           "model.layers.{bid}.enorm.weight",
+        "mtp.layers.0.hnorm.weight":           "model.layers.{bid}.hnorm.weight",
+        "mtp.layers.0.eh_proj.weight":         "model.layers.{bid}.eh_proj.weight",
+        "mtp.layers.1.norm.weight":            "model.layers.{bid}.post_attention_layernorm.weight",
+        "mtp.layers.1.final_layernorm.weight": "model.layers.{bid}.shared_head.norm.weight",
+    }

+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        #   mtp.layers.0: NextN input fusion + attention
+        #   mtp.layers.1: MoE + final head norm
+        if self._mtp_bid is not None and name.startswith(("mtp.layers.0.", "mtp.layers.1.")):
+            suffix = name.split(".", 3)[3]
+            bid = self._mtp_bid
+            renamed = self._MTP_SPECIAL_RENAMES.get(name)
+            name = renamed.format(bid=bid) if renamed else f"backbone.layers.{bid}.{suffix}"
+
+        if self.is_moe and bid is not None:
            if name.endswith("mixer.gate.e_score_correction.bias"):
                yield from ModelBase.modify_tensors(self, data_torch, name, bid)
                return
@@ -449,6 +449,8 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```

+User can use the device management in [docs/multi-gpu.md](https://github.com/ggml-org/llama.cpp/blob/master/docs/multi-gpu.md), like parameter `--device SYCL0,SYCL1` to assign one or more devices.
+
 ## Windows

 ### Install GPU driver
@@ -763,6 +765,7 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```

+User can use the device management in [docs/multi-gpu.md](https://github.com/ggml-org/llama.cpp/blob/master/docs/multi-gpu.md), like parameter `--device SYCL0,SYCL1` to assign one or more devices.

 ## Environment Variable

@@ -895,6 +898,45 @@ Pass these via `CXXFLAGS` or add a one-off `#define` to enable a flag on the spo
    set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
  ```

+- When I set `SYCL_CACHE_PERSISTENT=1` in running time, I meet crash.
+
+  `SYCL_CACHE_PERSISTENT=1` is not recommended by llama.cpp SYCL backend.
+  When cache is enabled, SYCL runtime will try to cache and reuse JIT-compiled binaries.
+
+  We find some AI will tell user this cmd to speed up SYCL backend. It only speeds up the startup to skip the JIT process, instead of running speed.
+
+  It will bring negative impact when the SYCL binary file is changed frequently in your running environment. The new & old codes mix will lead to crash.
+
+  Compare to the benefit, it has brought more failed cases.
+  If you are not familiar with the SYCL compiler principle of JIT and AOT, please don't use it.
+
+  To restore, you need to remove the local cache: `~/.cache/libsycl_cache/` and execute `unset SYCL_CACHE_PERSISTENT` in running time.
+
+- How to use iGPU and dGPU in same time?
+
+  1. Detect the devices in your running time.
+  ```
+  source /opt/intel/oneapi/setvars.sh
+  ./build/bin/llama-server --list-devices
+
+  or
+  ./build/bin/llama-cli --list-devices
+  ./build/bin/llama-bench --list-devices
+  ./build/bin/llama-completion --list-devices
+
+  Available devices:
+    SYCL0: Intel(R) Arc(TM) A770 Graphics (15473 MiB, 15473 MiB free)
+    SYCL1: Intel(R) UHD Graphics 770 (59675 MiB, 44986 MiB free)
+  ```
+
+  The dGPU will be in the head of this list and iGPU will be the end.
+  If not all GPUs are listed, please check the env var: ONEAPI_DEVICE_SELECTOR and unset it.
+
+  2. Set the iGPU and dGPU
+
+  Set the iGPU and dGPU by `./build/bin/llama-server --device SYCL0,SYCL1,SYCLxxx`.
+
+
 ### **GitHub contribution**:
 Please add the `[SYCL]` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.

@@ -15,7 +15,7 @@ Legend:
 | Operation | BLAS | CANN | CPU | CUDA | ET | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|------|------|------|
 |                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -41,9 +41,9 @@ Legend:
 |                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -59,7 +59,7 @@ Legend:
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -68,7 +68,7 @@ Legend:
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -202,6 +202,12 @@ Example Video:

 If a draft model is combined with a draftless decoding the draftless decoding has higher precedence.

+### Backend Sampling
+
+Use `--backend-sampling` to run supported target-model samplers on the model backend. Draft-model sampling uses the backend by default and can be controlled with `--spec-draft-backend-sampling` and `--no-spec-draft-backend-sampling`.
+
+Unsupported samplers and device layouts fall back to CPU sampling. Tensor split mode does not support backend sampling. A fixed seed produces repeatable random draws, but stochastic CPU and backend sampling can still select different tokens because floating-point operations can differ between implementations and devices. Use greedy sampling when exact output matching is required.
+
 ### General Speculative Parameters

 ```
@@ -3,9 +3,11 @@
 #include "common.h"
 #include "ngram-cache.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "log.h"
 #include "llama.h"

+#include <algorithm>
 #include <clocale>
 #include <cstdint>
 #include <cstdio>
@@ -27,6 +29,10 @@ int main(int argc, char ** argv){
    // max. number of additional tokens to draft if match is found
    const int n_draft = params.speculative.draft.n_max;

+    const auto output_limits = common_speculative_get_output_limits(params.n_batch, params.n_parallel, n_draft);
+    params.n_outputs_max = output_limits.total;
+    params.n_outputs_max_per_seq = output_limits.per_seq;
+
    // init llama.cpp
    llama_backend_init();
    llama_numa_init(params.numa);
@@ -47,6 +47,7 @@ CMD_ARGS+=("../../convert_hf_to_gguf.py" "--verbose")
 CMD_ARGS+=("${MODEL_PATH}")
 CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
 CMD_ARGS+=("--outtype" "${TYPE}")
+CMD_ARGS+=("--model-name" "${MODEL_NAME}")
 [[ -n "$METADATA_OVERRIDE" ]] && CMD_ARGS+=("--metadata" "${METADATA_OVERRIDE}")
 [[ -n "$MMPROJ" ]] && CMD_ARGS+=("${MMPROJ}")

@@ -31,6 +31,7 @@ python ../../convert_hf_to_gguf.py --verbose \
    ${EMBEDDING_MODEL_PATH} \
    --outfile ${CONVERTED_MODEL} \
    --outtype ${TYPE} \
+    --model-name ${MODEL_NAME} \
    ${SENTENCE_TRANSFORMERS}

 echo ""
@@ -5,6 +5,7 @@
 #include "log.h"
 #include "llama.h"

+#include <algorithm>
 #include <clocale>
 #include <cstdio>
 #include <cstring>
@@ -29,6 +30,11 @@ int main(int argc, char ** argv) {
        return 1;
    }

+    const auto output_limits = common_speculative_get_output_limits(
+            params.n_batch, params.n_parallel, common_speculative_n_max(&params.speculative));
+    params.n_outputs_max = output_limits.total;
+    params.n_outputs_max_per_seq = output_limits.per_seq;
+
    // init llama.cpp
    llama_backend_init();
    llama_numa_init(params.numa);
@@ -55,6 +61,9 @@ int main(int argc, char ** argv) {

        auto params_dft = params;

+        params_dft.n_outputs_max = params.n_parallel;
+        params_dft.n_outputs_max_per_seq = 1;
+
        params_dft.devices      = params_spec.devices;
        params_dft.model        = params_spec.mparams;
        params_dft.n_gpu_layers = params_spec.n_gpu_layers;
@@ -1,6 +1,7 @@
 #include "arg.h"
 #include "common.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "log.h"
 #include "llama.h"

@@ -57,6 +58,11 @@ int main(int argc, char ** argv) {
    // max number of parallel drafting sequences (i.e. tree branches)
    const int n_seq_dft = params.n_parallel;

+    const auto output_limits = common_speculative_get_output_limits(
+            params.n_batch, params.n_parallel, params.speculative.draft.n_max);
+    params.n_outputs_max = output_limits.total;
+    params.n_outputs_max_per_seq = output_limits.per_seq;
+
    // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
    const float p_draft_split = params.speculative.draft.p_split;

@@ -83,6 +89,8 @@ int main(int argc, char ** argv) {
    params.devices = params.speculative.draft.devices;
    params.model = params.speculative.draft.mparams;
    params.n_gpu_layers = params.speculative.draft.n_gpu_layers;
+    params.n_outputs_max = params.n_parallel;
+    params.n_outputs_max_per_seq = 1;
    if (params.speculative.draft.cpuparams.n_threads > 0) {
        params.cpuparams.n_threads = params.speculative.draft.cpuparams.n_threads;
    }
@@ -12,6 +12,7 @@ This script processes files with specified options.

 Options:
  -h, --help    Display this help message and exit.
+  -d, --device  <value>    Set SYCL devices (default: SYCL0).
  -c, --context <value>    Set context length. Bigger need more memory.
  -p, --promote <value>    Prompt to start generation with.
  -m, --model   <value>    Full model file path.
@@ -41,10 +42,16 @@ MODEL_FILE=../models/Qwen3.5-4B-Q4_0.gguf
 NGL=99
 CONTEXT=4096
 GGML_SYCL_DEVICE=-1
+SYCL_DEVICES="SYCL0"
 SPLIT_MODE=layer
 LOG_VERBOSE=3
 while [[ $# -gt 0 ]]; do
    case "$1" in
+        -d|--device)
+            SYCL_DEVICES="$2"
+            shift
+            shift
+            ;;
        -c|--context)
            CONTEXT=$2
            # Shift twice to consume both the option flag and its value
@@ -95,8 +102,6 @@ while [[ $# -gt 0 ]]; do
    esac
 done

-
-
 source /opt/intel/oneapi/setvars.sh

 #export GGML_SYCL_DEBUG=1
@@ -107,17 +112,19 @@ source /opt/intel/oneapi/setvars.sh
 export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 echo "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=${UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS}"

+echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
+
+
 if [ $GGML_SYCL_DEVICE -ne -1 ]; then
    echo "Use $GGML_SYCL_DEVICE as main GPU"
    #use signle GPU only
    GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
-    echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
 else
-    echo "Use all Intel GPUs, including iGPU & dGPU"
+    echo "Use Intel GPUs: ${SYCL_DEVICES}"
    GPUS_SETTING="-sm ${SPLIT_MODE}"
- fi
+fi

-echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap --host 0.0.0.0 --port 8000"
-ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap --host 0.0.0.0 --port 8000
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --device ${SYCL_DEVICES} --mmap --host 0.0.0.0 --port 8000"
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --device ${SYCL_DEVICES} --mmap --host 0.0.0.0 --port 8000


@@ -12,6 +12,7 @@ This script processes files with specified options.

 Options:
  -h, --help    Display this help message and exit.
+  -d, --device  <value>    Set SYCL devices (default: SYCL0).
  -c, --context <value>    Set context length. Bigger need more memory.
  -p, --promote <value>    Prompt to start generation with.
  -m, --model   <value>    Full model file path.
@@ -42,10 +43,16 @@ MODEL_FILE=../models/llama-2-7b.Q4_0.gguf
 NGL=99
 CONTEXT=4096
 GGML_SYCL_DEVICE=-1
+SYCL_DEVICES="SYCL0"
 SPLIT_MODE=layer
 LOG_VERBOSE=3
 while [[ $# -gt 0 ]]; do
    case "$1" in
+        -d|--device)
+            SYCL_DEVICES="$2"
+            shift
+            shift
+            ;;
        -c|--context)
            CONTEXT=$2
            # Shift twice to consume both the option flag and its value
@@ -115,16 +122,17 @@ source /opt/intel/oneapi/setvars.sh
 export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 echo "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=${UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS}"

+echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
+
 if [ $GGML_SYCL_DEVICE -ne -1 ]; then
    echo "Use $GGML_SYCL_DEVICE as main GPU"
    #use signle GPU only
    GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
-    echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
 else
-    echo "Use all Intel GPUs, including iGPU & dGPU"
+    echo "Use Intel GPUs: ${SYCL_DEVICES}"
    GPUS_SETTING="-sm ${SPLIT_MODE}"
 fi

-echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
-ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --device ${SYCL_DEVICES} --mmap "
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --device ${SYCL_DEVICES} --mmap

@@ -13,6 +13,7 @@ set "MODEL_FILE=..\models\Qwen3.5-4B-Q4_0.gguf"
 set "NGL=99"
 set "CONTEXT=4096"
 set "GGML_SYCL_DEVICE=-1"
+set "SYCL_DEVICES=SYCL0"
 set "SPLIT_MODE=layer"
 set "LOG_VERBOSE=3"

@@ -36,6 +37,21 @@ if /I "%~1"=="--context" (
  goto parse_args
 )

+if /I "%~1"=="-d" (
+  if "%~2"=="" goto missing_value
+  set "SYCL_DEVICES=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--device" (
+  if "%~2"=="" goto missing_value
+  set "SYCL_DEVICES=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
 if /I "%~1"=="-m" (
  if "%~2"=="" goto missing_value
  set "MODEL_FILE=%~2"
@@ -130,6 +146,7 @@ echo This script processes files with specified options.
 echo.
 echo Options:
 echo   -h, --help    Display this help message and exit.
+echo   -d, --device ^<value^>    Set SYCL devices (default: SYCL0).
 echo   -c, --context ^<value^>    Set context length. Bigger need more memory.
 echo   -m, --model   ^<value^>    Full model file path.
 echo   -mg,--main-gpu ^<value^>   Set main GPU ID (0 - n) for single GPU mode.
@@ -160,19 +177,20 @@ REM Support malloc device memory more than 4GB.
 set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"
 echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%

+echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
+
 if not "%GGML_SYCL_DEVICE%"=="-1" (
  echo Use %GGML_SYCL_DEVICE% as main GPU
  REM Use single GPU only.
  set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
-  echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
-) else (
-  echo Use all Intel GPUs, including iGPU ^& dGPU
+  ) else (
+  echo Use Intel GPUs: %SYCL_DEVICES%
  set "GPUS_SETTING=-sm %SPLIT_MODE%"
 )

-echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap --host 0.0.0.0 --port 8000
+echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --device %SYCL_DEVICES% --mmap --host 0.0.0.0 --port 8000
 set "ZES_ENABLE_SYSMAN=1"
-%BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap --host 0.0.0.0 --port 8000
+%BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --device "%SYCL_DEVICES%" --mmap --host 0.0.0.0 --port 8000

 endlocal

@@ -19,6 +19,7 @@ set "MODEL_FILE=..\models\llama-2-7b.Q4_0.gguf"
 set "NGL=99"
 set "CONTEXT=4096"
 set "GGML_SYCL_DEVICE=-1"
+set "SYCL_DEVICES=SYCL0"
 set "SPLIT_MODE=layer"
 set "LOG_VERBOSE=3"

@@ -42,6 +43,21 @@ if /I "%~1"=="--context" (
  goto parse_args
 )

+if /I "%~1"=="-d" (
+  if "%~2"=="" goto missing_value
+  set "SYCL_DEVICES=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--device" (
+  if "%~2"=="" goto missing_value
+  set "SYCL_DEVICES=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
 if /I "%~1"=="-p" (
  if "%~2"=="" goto missing_value
  set "INPUT_PROMPT=%~2"
@@ -151,6 +167,7 @@ echo This script processes files with specified options.
 echo.
 echo Options:
 echo   -h, --help    Display this help message and exit.
+echo   -d, --device ^<value^>    Set SYCL devices (default: SYCL0).
 echo   -c, --context ^<value^>    Set context length. Bigger need more memory.
 echo   -p, --promote ^<value^>    Prompt to start generation with.
 echo   -m, --model   ^<value^>    Full model file path.
@@ -182,19 +199,21 @@ REM Support malloc device memory more than 4GB.
 set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"
 echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%

+echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
+
 if not "%GGML_SYCL_DEVICE%"=="-1" (
  echo Use %GGML_SYCL_DEVICE% as main GPU
  REM Use single GPU only.
  set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
-  echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
-) else (
-  echo Use all Intel GPUs, including iGPU ^& dGPU
+  )
+else (
+  echo Use Intel GPUs: %SYCL_DEVICES%
  set "GPUS_SETTING=-sm %SPLIT_MODE%"
 )

-echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m %MODEL_FILE% -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap
+echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m %MODEL_FILE% -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --device %SYCL_DEVICES% --mmap
 set "ZES_ENABLE_SYSMAN=1"
-%BIN_FILE% -m "%MODEL_FILE%" -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap
+%BIN_FILE% -m "%MODEL_FILE%" -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --device "%SYCL_DEVICES%" --mmap

 endlocal

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 18)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_MINOR 19)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -2788,6 +2788,12 @@ extern "C" {
            struct ggml_cgraph * cgraph,
            struct ggml_tensor * tensor);

+    // add the tensor and its parents to the graph without marking them for compute
+    // the flag is set later, when the tensor is reached from a node that computes
+    GGML_API void ggml_build_forward_order(
+            struct ggml_cgraph * cgraph,
+            struct ggml_tensor * tensor);
+
    GGML_API void ggml_build_backward_expand(
        struct ggml_context *  ctx,        // context for gradient computation
        struct ggml_cgraph  *  cgraph,
@@ -8,6 +8,22 @@
 #include <sys/sysctl.h>
 #endif

+#if !defined(HWCAP_FPHP)
+#define HWCAP_FPHP (1 << 9)
+#endif
+
+#if !defined(HWCAP_ASIMDHP)
+#define HWCAP_ASIMDHP (1 << 10)
+#endif
+
+#if !defined(HWCAP_ASIMDDP)
+#define HWCAP_ASIMDDP (1 << 20)
+#endif
+
+#if !defined(HWCAP_SVE)
+#define HWCAP_SVE (1 << 22)
+#endif
+
 #if !defined(HWCAP2_SVE2)
 #define HWCAP2_SVE2 (1 << 1)
 #endif
@@ -23,7 +39,7 @@
 struct aarch64_features {
    // has_neon not needed, aarch64 has NEON guaranteed
    bool has_dotprod     = false;
-    bool has_fp16_va     = false;
+    bool has_fp16        = false;
    bool has_sve         = false;
    bool has_sve2        = false;
    bool has_i8mm        = false;
@@ -36,7 +52,7 @@ struct aarch64_features {
        uint32_t hwcap2 = getauxval(AT_HWCAP2);

        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-        has_fp16_va = !!(hwcap & HWCAP_FPHP);
+        has_fp16    = !!(hwcap & HWCAP_FPHP) && !!(hwcap & HWCAP_ASIMDHP);
        has_sve     = !!(hwcap & HWCAP_SVE);
        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
@@ -75,7 +91,7 @@ static int ggml_backend_cpu_aarch64_score() {
    score += 1<<1;
 #endif
 #ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
-    if (!af.has_fp16_va) { return 0; }
+    if (!af.has_fp16) { return 0; }
    score += 1<<2;
 #endif
 #ifdef GGML_USE_SVE
@@ -2608,7 +2608,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
    return true;
 }

-#elif defined(__gnu_linux__)
+#elif defined(__linux__)
 // TODO: this may not work on BSD, to be verified

 static bool ggml_thread_apply_affinity(const bool * mask) {
@@ -195,6 +195,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -214,6 +215,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -253,9 +253,9 @@ static void ggml_cpy_f32_q8_0_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    GGML_ASSERT(ne % QK8_0 == 0);
-    const int64_t num_blocks = ne / QK8_0;
+    const int64_t num_blocks = (ne/QK8_0 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
+    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -264,9 +264,9 @@ static void ggml_cpy_q8_0_f32_cuda(
    const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t nb00, const int64_t nb01, const int64_t nb02,
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

-    const int64_t num_blocks = ne;
+    const int64_t num_blocks = (ne/QK8_0 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
+    cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -276,9 +276,9 @@ static void ggml_cpy_f32_q4_0_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    GGML_ASSERT(ne % QK4_0 == 0);
-    const int64_t num_blocks = ne / QK4_0;
+    const int64_t num_blocks = (ne/QK4_0 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
+    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -289,9 +289,9 @@ static void ggml_cpy_q4_0_f32_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12,
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
-    const int64_t num_blocks = ne;
+    const int64_t num_blocks = (ne/QK4_0 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
+    cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -302,9 +302,9 @@ static void ggml_cpy_f32_q4_1_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    GGML_ASSERT(ne % QK4_1 == 0);
-    const int64_t num_blocks = ne / QK4_1;
+    const int64_t num_blocks = (ne/QK4_1 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
+    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -315,9 +315,9 @@ static void ggml_cpy_q4_1_f32_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12,
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
-    const int64_t num_blocks = ne;
+    const int64_t num_blocks = (ne/QK4_1 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
+    cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -328,9 +328,9 @@ static void ggml_cpy_f32_q5_0_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    GGML_ASSERT(ne % QK5_0 == 0);
-    const int64_t num_blocks = ne / QK5_0;
+    const int64_t num_blocks = (ne/QK5_0 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
+    cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -341,9 +341,9 @@ static void ggml_cpy_q5_0_f32_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12,
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
-    const int64_t num_blocks = ne;
+    const int64_t num_blocks = (ne/QK5_0 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
+    cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -354,9 +354,9 @@ static void ggml_cpy_f32_q5_1_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    GGML_ASSERT(ne % QK5_1 == 0);
-    const int64_t num_blocks = ne / QK5_1;
+    const int64_t num_blocks = (ne/QK5_1 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
+    cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -367,9 +367,9 @@ static void ggml_cpy_q5_1_f32_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12,
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
-    const int64_t num_blocks = ne;
+    const int64_t num_blocks = (ne/QK5_1 + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
+    cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -380,9 +380,9 @@ static void ggml_cpy_f32_iq4_nl_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    GGML_ASSERT(ne % QK4_NL == 0);
-    const int64_t num_blocks = ne / QK4_NL;
+    const int64_t num_blocks = (ne/QK4_NL + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    GGML_ASSERT(num_blocks <= INT_MAX);
-    cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
+    cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

@@ -2651,6 +2651,52 @@ static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
    return true;
 }

+static bool ggml_cuda_should_fuse_rms_norm_mul_rope(const ggml_tensor * rms_norm,
+                                                    const ggml_tensor * mul,
+                                                    const ggml_tensor * rope) {
+    if (rms_norm->op != GGML_OP_RMS_NORM || mul->op != GGML_OP_MUL || rope->op != GGML_OP_ROPE) {
+        return false;
+    }
+
+    if (rms_norm->src[0]->type != GGML_TYPE_F32 || rms_norm->type != GGML_TYPE_F32 ||
+        mul->src[0]->type != GGML_TYPE_F32 || mul->src[1]->type != GGML_TYPE_F32 ||
+        mul->type != GGML_TYPE_F32 || rope->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (rope->src[0] != mul) {
+        return false;
+    }
+
+    //if rms norm is the B operand, then we don't handle broadcast
+    if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
+        return false;
+    }
+
+    if (!ggml_are_same_shape(rms_norm, mul)) {
+        return false;
+    }
+
+    //rms_norm kernel assumes contiguous rows
+    if (!ggml_is_contiguous_rows(rms_norm->src[0]) ||
+        !ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+        return false;
+    }
+
+    // the fused kernel handles the norm/neox rope modes only
+    const int mode = ((const int32_t *) rope->op_params)[2];
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+        return false;
+    }
+
+    const int n_dims = ((const int32_t *) rope->op_params)[1];
+    if (n_dims % 2 != 0 || rope->src[0]->ne[0] % 2 != 0) {
+        return false;
+    }
+
+    return true;
+}
+
 // match gated_delta_net + the strided cpy that scatters its state snapshots into the cache
 // (slot i -> rollback group i, slot 0 newest), so the kernel can write them and skip the cpy.
 static int ggml_cuda_try_gdn_cache_fusion(
@@ -2980,6 +3026,36 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
        }
    }

+    std::initializer_list<enum ggml_op> rms_norm_mul_rope_ops          = { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE };
+    std::initializer_list<enum ggml_op> rms_norm_mul_rope_set_rows_ops = { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };
+
+    if (is_equal(rms_norm_mul_rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) {
+        const ggml_tensor * rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor * mul      = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * rope     = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * view     = cgraph->nodes[node_idx + 3];
+        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 4];
+
+        if (ggml_check_edges(cgraph, node_idx, {{1, 0, 0}, {2, 0, 1}, {3, 0, 2}, {4, 0, 3}}) &&
+            ggml_cuda_should_fuse_rms_norm_mul_rope(rms_norm, mul, rope) &&
+            ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
+            int out_nodes[] = { node_idx + 4 };
+            return ggml_cuda_check_fusion_memory_ranges(cgraph, node_idx, (int)ops.size(), out_nodes, 1);
+        }
+    }
+
+    if (is_equal(rms_norm_mul_rope_ops, ops) && ggml_can_fuse(cgraph, node_idx, ops)) {
+        const ggml_tensor * rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor * mul      = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * rope     = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_rms_norm_mul_rope(rms_norm, mul, rope)) {
+            int out_nodes[] = { node_idx + 2 };
+            return ggml_cuda_check_fusion_memory_ranges(cgraph, node_idx, (int)ops.size(), out_nodes, 1);
+        }
+        return false;
+    }
+
    std::initializer_list<enum ggml_op> rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };

    if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
@@ -2988,7 +3064,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];

        if (ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
-            return true;
+            int out_nodes[] = { node_idx + 2 };
+            return ggml_cuda_check_fusion_memory_ranges(cgraph, node_idx, (int)ops.size(), out_nodes, 1);
        }
    }

@@ -3840,6 +3917,16 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
        return fused_node_count - 1;
    }

+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
+        ggml_cuda_op_rms_norm_mul_rope_fused(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2], cgraph->nodes[i + 4]);
+        return 4;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE }, {})) {
+        ggml_cuda_op_rms_norm_mul_rope_fused(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2], nullptr);
+        return 2;
+    }
+
    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD }, {})) {
        ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2]);
        return 2;
@@ -4033,7 +4120,11 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                    continue;
                }
 #ifndef NDEBUG
-                assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
+                // On integrated GPUs (APUs, e.g. RDNA3.5) the scheduler may place a
+                // node's output on the host-visible buffer, which the compute path
+                // handles. Allow that here, mirroring the src-tensor check below.
+                assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) ||
+                       (integrated && ggml_backend_buft_is_cuda_host(node->buffer->buft)));
                for (int j = 0; j < GGML_MAX_SRC; j++) {
                    if (node->src[j] != nullptr) {
                        assert(node->src[j]->buffer);
@@ -5094,7 +5185,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            return max_bias == 0.0f;
        }
        case GGML_OP_ROLL:
-            if(op->src[0]->type == GGML_TYPE_F32) {
+            if(op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0])) {
                return true;
            }
            return false;
@@ -5205,6 +5296,7 @@ static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const gg

 static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
 #ifdef GGML_CUDA_NO_PEER_COPY
+    GGML_UNUSED(dev);
    return nullptr;
 #else
    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
@@ -8,7 +8,6 @@ struct __builtin_align__(32) float8 {
    float x; float y; float z; float w;
    float p; float q; float r; float s;
 };
-#endif

 #if CUDART_VERSION >= 12080
 static __device__ __forceinline__ float nvfp4_native_scale_error(
@@ -49,6 +48,7 @@ static __device__ __forceinline__ float nvfp4_native_scale_error(
    return err;
 }
 #endif // CUDART_VERSION >= 12080
+#endif // defined(BLACKWELL_MMA_AVAILABLE)

 __launch_bounds__(CUDA_QUANTIZE_BLOCK_SIZE, 1)
 static __global__ void quantize_q8_1(
@@ -670,3 +670,238 @@ void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rope, ggml_tensor * set_rows) {
    ggml_cuda_op_rope_impl<true>(ctx, rope, set_rows);
 }
+
+// fused RMS_NORM + MUL + ROPE (+ VIEW + SET_ROWS)
+// one block per row: block_reduce gives the norm scale, then each thread applies mul and rope to the elements it owns
+template <int block_size, bool has_ff, typename D>
+static __global__ void rms_norm_mul_rope_f32(
+        const float * x, D * dst, const int ncols,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s1, const int64_t s2, const int64_t s3,
+        const float eps,
+        const float * mul,
+        const int64_t mul_s01, const int64_t mul_s02, const int64_t mul_s03,
+        const uint3 mul_ncols_packed, const uint3 mul_nrows_packed,
+        const uint3 mul_nchannels_packed, const uint3 mul_nsamples_packed,
+        const int n_dims, const int32_t * pos,
+        const float freq_scale, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float theta_scale,
+        const float * freq_factors,
+        const int64_t * row_indices, const int set_rows_stride,
+        const bool is_neox) {
+    ggml_cuda_pdl_lc();
+    const int row     = blockIdx.x;
+    const int channel = blockIdx.y;
+    const int sample  = blockIdx.z;
+    const int tid     = threadIdx.x;
+
+    x += sample*s03 + channel*s02 + row*s01;
+
+    const uint32_t mul_row     = fastmodulo(row,     mul_nrows_packed);
+    const uint32_t mul_channel = fastmodulo(channel, mul_nchannels_packed);
+    const uint32_t mul_sample  = fastmodulo(sample,  mul_nsamples_packed);
+    mul += mul_sample*mul_s03 + mul_channel*mul_s02 + mul_row*mul_s01;
+
+    float tmp = 0.0f;
+
+    ggml_cuda_pdl_sync();
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[col];
+        tmp += xi * xi;
+    }
+
+    extern __shared__ float s_sum[];
+    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+
+    const float scale = rsqrtf(tmp/ncols + eps);
+
+    int64_t idst = sample*s3 + channel*s2 + row*s1;
+    if (set_rows_stride != 0) {
+        idst = row*s1 + row_indices[channel]*set_rows_stride;
+    }
+    dst += idst;
+
+    for (int i0 = 2*tid; i0 < ncols; i0 += 2*block_size) {
+        int ix0;
+        int ix1;
+        if (is_neox && i0 < n_dims) {
+            ix0 = i0/2;
+            ix1 = i0/2 + n_dims/2;
+        } else {
+            ix0 = i0 + 0;
+            ix1 = i0 + 1;
+        }
+
+        const float x0 = scale * x[ix0] * mul[fastmodulo(ix0, mul_ncols_packed)];
+        const float x1 = scale * x[ix1] * mul[fastmodulo(ix1, mul_ncols_packed)];
+
+        if (i0 >= n_dims) {
+            dst[ix0] = ggml_cuda_cast<D>(x0);
+            dst[ix1] = ggml_cuda_cast<D>(x1);
+            continue;
+        }
+
+        const float theta_base  = pos[channel]*powf(theta_scale, i0/2.0f);
+        const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+        float cos_theta;
+        float sin_theta;
+        rope_yarn<true>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+
+        dst[ix0] = ggml_cuda_cast<D>(x0*cos_theta - x1*sin_theta);
+        dst[ix1] = ggml_cuda_cast<D>(x0*sin_theta + x1*cos_theta);
+    }
+}
+
+template <typename D>
+static void rms_norm_mul_rope_cuda(
+        const float * x, D * dst,
+        const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s1, const int64_t s2, const int64_t s3,
+        const float eps,
+        const float * mul,
+        const int64_t mul_s01, const int64_t mul_s02, const int64_t mul_s03,
+        const uint32_t mul_ncols, const uint32_t mul_nrows,
+        const uint32_t mul_nchannels, const uint32_t mul_nsamples,
+        const int n_dims, const int32_t * pos,
+        const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims,
+        const float * freq_factors,
+        const int64_t * row_indices, const int set_rows_stride,
+        const bool is_neox, cudaStream_t stream) {
+    GGML_ASSERT(ncols % 2 == 0);
+
+    const dim3 blocks_num(nrows, nchannels, nsamples);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    const uint3 mul_ncols_packed     = init_fastdiv_values(mul_ncols);
+    const uint3 mul_nrows_packed     = init_fastdiv_values(mul_nrows);
+    const uint3 mul_nchannels_packed = init_fastdiv_values(mul_nchannels);
+    const uint3 mul_nsamples_packed  = init_fastdiv_values(mul_nsamples);
+
+    if (ncols < 1024) {
+        const dim3 block_dims(256, 1, 1);
+        const ggml_cuda_kernel_launch_params launch_params = {blocks_num, block_dims, 32*sizeof(float), stream};
+        if (freq_factors == nullptr) {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<256, false, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        } else {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<256, true, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        }
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        const ggml_cuda_kernel_launch_params launch_params = {blocks_num, block_dims, 32*sizeof(float), stream};
+        if (freq_factors == nullptr) {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<1024, false, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        } else {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<1024, true, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        }
+    }
+}
+
+void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx,
+        ggml_tensor * rms_norm, ggml_tensor * mul, ggml_tensor * rope, ggml_tensor * set_rows) {
+    const ggml_tensor * x = rms_norm->src[0];
+    const ggml_tensor * mul_src = mul->src[0] == rms_norm ? mul->src[1] : mul->src[0];
+
+    float eps = 0.0f;
+    memcpy(&eps, rms_norm->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+
+    GGML_ASSERT(x->type == GGML_TYPE_F32);
+    GGML_ASSERT(mul_src->type == GGML_TYPE_F32);
+    GGML_ASSERT(rope->type == GGML_TYPE_F32);
+
+    void *          dst_d           = rope->data;
+    ggml_type       dst_type        = rope->type;
+    const int64_t * row_indices     = nullptr;
+    int             set_rows_stride = 0;
+
+    if (set_rows != nullptr) {
+        dst_d           = set_rows->data;
+        dst_type        = set_rows->type;
+        row_indices     = (const int64_t *) set_rows->src[1]->data;
+        set_rows_stride = set_rows->nb[1] / ggml_type_size(set_rows->type);
+    }
+
+    const int n_dims     = ((const int32_t *) rope->op_params)[1];
+    const int mode       = ((const int32_t *) rope->op_params)[2];
+    const int n_ctx_orig = ((const int32_t *) rope->op_params)[4];
+
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+
+    memcpy(&freq_base,   (const int32_t *) rope->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (const int32_t *) rope->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (const int32_t *) rope->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (const int32_t *) rope->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (const int32_t *) rope->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (const int32_t *) rope->op_params + 10, sizeof(float));
+
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+    const int32_t * pos = (const int32_t *) rope->src[1]->data;
+
+    const float * freq_factors = rope->src[2] != nullptr ? (const float *) rope->src[2]->data : nullptr;
+
+    rope_corr_dims corr_dims;
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
+
+    const size_t ts0 = ggml_type_size(x->type);
+    GGML_ASSERT(x->nb[0] == ts0);
+    const int64_t s01 = x->nb[1] / ts0;
+    const int64_t s02 = x->nb[2] / ts0;
+    const int64_t s03 = x->nb[3] / ts0;
+
+    const size_t ts_mul = ggml_type_size(mul_src->type);
+    GGML_ASSERT(mul_src->nb[0] == ts_mul);
+    const int64_t mul_s01 = mul_src->nb[1] / ts_mul;
+    const int64_t mul_s02 = mul_src->nb[2] / ts_mul;
+    const int64_t mul_s03 = mul_src->nb[3] / ts_mul;
+
+    const size_t ts_dst = ggml_type_size(rope->type);
+    const int64_t s1 = rope->nb[1] / ts_dst;
+    const int64_t s2 = rope->nb[2] / ts_dst;
+    const int64_t s3 = rope->nb[3] / ts_dst;
+
+    cudaStream_t stream = ctx.stream();
+
+    if (dst_type == GGML_TYPE_F32) {
+        rms_norm_mul_rope_cuda((const float *) x->data, (float *) dst_d,
+            x->ne[0], x->ne[1], x->ne[2], x->ne[3], s01, s02, s03, s1, s2, s3, eps,
+            (const float *) mul_src->data, mul_s01, mul_s02, mul_s03,
+            mul_src->ne[0], mul_src->ne[1], mul_src->ne[2], mul_src->ne[3],
+            n_dims, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+            freq_factors, row_indices, set_rows_stride, is_neox, stream);
+    } else if (dst_type == GGML_TYPE_F16) {
+        rms_norm_mul_rope_cuda((const float *) x->data, (half *) dst_d,
+            x->ne[0], x->ne[1], x->ne[2], x->ne[3], s01, s02, s03, s1, s2, s3, eps,
+            (const float *) mul_src->data, mul_s01, mul_s02, mul_s03,
+            mul_src->ne[0], mul_src->ne[1], mul_src->ne[2], mul_src->ne[3],
+            n_dims, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+            freq_factors, row_indices, set_rows_stride, is_neox, stream);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
@@ -7,3 +7,5 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

 void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
+
+void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rms_norm, ggml_tensor * mul, ggml_tensor * rope, ggml_tensor * set_rows);
@@ -1268,8 +1268,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
        case GGML_OP_ARGSORT:
        case GGML_OP_TOP_K:
        case GGML_OP_ARANGE:
-        case GGML_OP_ROLL:
            return true;
+        case GGML_OP_ROLL:
+            return ggml_is_contiguous(op->src[0]);
        case GGML_OP_FLASH_ATTN_EXT:
            // for new head sizes, add checks here
            if (op->src[0]->ne[0] != 32 &&
@@ -3816,7 +3816,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
    }

    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-    nth = std::min(nth, args.ne00_t);
+    nth = std::min(nth, (args.ne00_t + 31)/32*32);

    const size_t smem = pipeline.smem;

@@ -11328,8 +11328,8 @@ kernel void kernel_lightning_indexer(
    const int i_kv_0   = tgpig.x*NK;            // first key of this threadgroup
    const int i_kv     = i_kv_0 + sgitg*NKPSG;  // first key of this simdgroup

-    threadgroup half4x4   sk4x4[NK*DK16];
-    threadgroup half    * sk = (threadgroup half *) sk4x4;
+    threadgroup half sk[NK * DK16 * 16];
+    threadgroup half4x4 * sk4x4 = (threadgroup half4x4 *) sk;

    for (short i = tiitg; i < NK*DK16; i += NTG) {
        const short ik  = i/DK16;
@@ -73,6 +73,7 @@ typedef const void * (*get_adreno_bin_kernel_func_t)(
 //------------------------------------------------------------------------------

 bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor);
+
 static bool ggml_cl_is_q4_0_soa(const ggml_tensor * tensor);
 static bool ggml_cl_is_q8_0_soa(const ggml_tensor * tensor);
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
@@ -4629,6 +4630,23 @@ static std::string ggml_opencl_fa_compile_opts(ggml_backend_opencl_context * bac
    if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E) {
        opts += " -D FA_C8_NO_SG_PIN";
    }
+    // Transposed K tile in local memory: the KV rows the QK loop walks together become
+    // adjacent, so a group of them is ONE 128-bit local read instead of several narrow
+    // ones. The QK loop is LDS-read-issue-bound (a wrong-math probe that kept every FMA/dp4a
+    // but removed the LDS reads ran the kernel ~40% faster), so this is worth up to +26% on
+    // fa=1 prefill. Output is bit-identical -- only the layout moves.
+    //
+    // DK <= 128 only. At DK=256 (gemma-3-4b) it measures 1-2% NEGATIVE and reproduces across
+    // rounds; padding the row stride does not recover it, so the cause is not a simple bank
+    // conflict and the wider tile does not want this layout.
+    //
+    // Default on within that gate; GGML_OPENCL_FA_K_LDS_T=0 restores the row-major tile.
+    {
+        const char * e = getenv("GGML_OPENCL_FA_K_LDS_T");
+        if ((e == nullptr || e[0] != '0') && cfg->dk <= 128) {
+            opts += " -D FA_K_LDS_T";
+        }
+    }
    return opts;
 }

@@ -211,7 +211,30 @@ __kernel void FA_TILE_NAME(

    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);

+#ifdef FA_K_LDS_T
+    // K tile transposed: [dk vec][kv row] instead of [kv row][dk vec].
+    //
+    // The QK loop walks 2 or 4 KV rows at a time against the same dk element. Row-major
+    // those are DK_VEC half4s apart, so each is its own 64-bit local read. Transposed they
+    // are adjacent, so a pair is one 128-bit read -- half the LDS issues for the same bytes,
+    // no extra registers, arithmetic untouched.
+    //
+    // This kernel looked like it should be FMA-bound (a half4 mad does ~4 ALU ops per LDS
+    // read, unlike the 1:1 of the dp4a loop), but it is NOT: a wrong-math probe that kept
+    // every FMA and removed the LDS reads ran it 38.6% faster (18.92 -> 11.62 ms/op).
+    // Explicitly 16-byte aligned: FA_LK_PAIR below reads two adjacent half4 as one float4,
+    // and the element type only obliges the compiler to align this array to 8. The indices
+    // are even so the offset is a multiple of 16, but the base has to be too, and relying
+    // on the compiler to over-align it is relying on luck.
+    __local KV_DATA_TYPE4 l_k[DK_VEC][BLOCK_N] __attribute__((aligned(16)));
+#define FA_LK(ROW, C) l_k[C][ROW]
+    // Two adjacent KV rows as one 128-bit local read (half4 pair == 16 B). j is even and
+    // BLOCK_N is even, so &l_k[c][j] is 16 B past a 16 B-aligned base.
+#define FA_LK_PAIR(C, J) as_half8(*(__local const float4 *)(&l_k[C][J]))
+#else
    __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+#define FA_LK(ROW, C) l_k[ROW][C]
+#endif
    __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

 #if N_SPLIT > 1 && !defined(HAS_SUBGROUP_SHUFFLE)
@@ -254,17 +277,17 @@ __kernel void FA_TILE_NAME(
 #ifdef FA_K_IMG
                if (use_kv_pad) {
                    const ulong k_row_offset = batch_idx * k_tile_nb3 + head_kv_idx * k_tile_nb2 + k_row_idx * k_nb1;
-                    l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
+                    FA_LK(row, col) = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
                } else {
                    const int k_row_px = batch_idx * k_pitch_px_batch + head_kv_idx * k_pitch_px_head + k_row_idx * k_pitch_px_row;
-                    l_k[row][col] = read_imageh(k_img, k_row_px + col);
+                    FA_LK(row, col) = read_imageh(k_img, k_row_px + col);
                }
 #else
                const ulong k_row_offset = batch_idx * k_tile_nb3 + head_kv_idx * k_tile_nb2 + k_row_idx * k_nb1;
-                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
+                FA_LK(row, col) = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
 #endif
            } else {
-                l_k[row][col] = (KV_DATA_TYPE4)(0.0h);
+                FA_LK(row, col) = (KV_DATA_TYPE4)(0.0h);
            }
        }
        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
@@ -292,8 +315,15 @@ __kernel void FA_TILE_NAME(
                FA_UNROLL
                for (int k = 0; k < SPLIT_DK_VEC; k++) {
                    const ACC_TYPE4 qk = q_priv[k];
+#if defined(FA_K_LDS_T)
+                    // 2 KV rows adjacent in the transposed tile: one 128-bit local read.
+                    const half8 kk = FA_LK_PAIR(dk_off + k, j);
+                    ACC_TYPE4 dot0 = qk * CONVERT_KV_ACC4(kk.lo);
+                    ACC_TYPE4 dot1 = qk * CONVERT_KV_ACC4(kk.hi);
+#else
                    ACC_TYPE4 dot0 = qk * CONVERT_KV_ACC4(l_k[j  ][dk_off + k]);
                    ACC_TYPE4 dot1 = qk * CONVERT_KV_ACC4(l_k[j+1][dk_off + k]);
+#endif
                    partial0 += dot0.s0 + dot0.s1 + dot0.s2 + dot0.s3;
                    partial1 += dot1.s0 + dot1.s1 + dot1.s2 + dot1.s3;
                }
@@ -359,7 +389,7 @@ __kernel void FA_TILE_NAME(
            ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
            FA_UNROLL
            for (int k = 0; k < SPLIT_DK_VEC; k++) {
-                dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][dk_off + k]), dot_acc);
+                dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(FA_LK(j, dk_off + k)), dot_acc);
            }
            local_partial[j][tid] =
                dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3;
@@ -452,10 +482,21 @@ __kernel void FA_TILE_NAME(
                FA_UNROLL
                for (int k = 0; k < DK_VEC; k++) {
                    const ACC_TYPE4 qk = q_priv[k];
+#if defined(FA_K_LDS_T)
+                    // 4 KV rows adjacent in the transposed tile: two 128-bit local reads
+                    // instead of four 64-bit ones.
+                    const half8 kk01 = FA_LK_PAIR(k, j);
+                    const half8 kk23 = FA_LK_PAIR(k, j + 2);
+                    dot_acc0 = mad(qk, CONVERT_KV_ACC4(kk01.lo), dot_acc0);
+                    dot_acc1 = mad(qk, CONVERT_KV_ACC4(kk01.hi), dot_acc1);
+                    dot_acc2 = mad(qk, CONVERT_KV_ACC4(kk23.lo), dot_acc2);
+                    dot_acc3 = mad(qk, CONVERT_KV_ACC4(kk23.hi), dot_acc3);
+#else
                    dot_acc0 = mad(qk, CONVERT_KV_ACC4(l_k[j][k]),   dot_acc0);
                    dot_acc1 = mad(qk, CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
                    dot_acc2 = mad(qk, CONVERT_KV_ACC4(l_k[j+2][k]), dot_acc2);
                    dot_acc3 = mad(qk, CONVERT_KV_ACC4(l_k[j+3][k]), dot_acc3);
+#endif
                }
                ACC_TYPE s0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
                ACC_TYPE s1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
@@ -1631,8 +1631,25 @@ __kernel void flash_attn_f32_q4_0(
    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);

 #ifdef FA_HAVE_INT_DOT
+// Accessors so the staging code is layout-agnostic.
+#ifdef FA_K_LDS_T
+#define FA_K_PACKED(ROW, IDX) l_k_packed[IDX][ROW]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[BLK][ROW]
+#else
+#define FA_K_PACKED(ROW, IDX) l_k_packed[ROW][IDX]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[ROW][BLK]
+#endif
+
+#ifdef FA_K_LDS_T
+    // K tile transposed: the 4 KV rows the QK loop walks together become adjacent, so each
+    // (block, group) step is ONE 128-bit local read instead of four 32-bit ones. The QK
+    // loop is LDS-read-issue-bound.
+    __local uint  l_k_packed[DK_Q4_BLOCKS_PREFILL * 8][BLOCK_N];
+    __local float l_k_scale [DK_Q4_BLOCKS_PREFILL][BLOCK_N];
+#else
    __local uint  l_k_packed[BLOCK_N][DK_Q4_BLOCKS_PREFILL * 8];
    __local float l_k_scale [BLOCK_N][DK_Q4_BLOCKS_PREFILL];
+#endif
 #else
    __local half4 l_k[BLOCK_N][DK_VEC];
 #endif
@@ -1660,17 +1677,17 @@ __kernel void flash_attn_f32_q4_0(
                    const global char * blk_ptr = k_base + k_row_off + blk * Q4_0_BLOCK_SIZE;
                    const float df = (float) vload_half(0, (const global half *) blk_ptr);
                    const global uchar * qs = (const global uchar *)(blk_ptr + 2);
-                    l_k_scale[row][blk] = df;
+                    FA_K_SCALE(row, blk) = df;
                    uint k_packed[8];
                    pack_q4_0_nibbles(qs, k_packed);
                    #pragma unroll
                    for (int j = 0; j < 8; ++j) {
-                        l_k_packed[row][blk * 8 + j] = k_packed[j];
+                        FA_K_PACKED(row, blk * 8 + j) = k_packed[j];
                    }
                } else {
-                    l_k_scale[row][blk] = 0.0f;
+                    FA_K_SCALE(row, blk) = 0.0f;
                    #pragma unroll
-                    for (int j = 0; j < 8; ++j) l_k_packed[row][blk * 8 + j] = 0u;
+                    for (int j = 0; j < 8; ++j) FA_K_PACKED(row, blk * 8 + j) = 0u;
                }
            }
 #else
@@ -1760,6 +1777,19 @@ __kernel void flash_attn_f32_q4_0(
                for (int b_local = 0; b_local < SPLIT_DK_Q4_BLOCKS; ++b_local) {
                    const int b = k_blk_base + b_local;
                    int sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+#ifdef FA_K_LDS_T
+                    // 4 KV rows are adjacent in the transposed tile: one 128-bit local
+                    // read per (block, group) instead of four 32-bit ones.
+                    #pragma unroll
+                    for (int g = 0; g < 8; ++g) {
+                        const uint  qp  = q_packed_pf[b_local * 8 + g];
+                        const uint4 kq4 = vload4(0, &l_k_packed[b * 8 + g][j]);
+                        sum0 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s0, sum0);
+                        sum1 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s1, sum1);
+                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s2, sum2);
+                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s3, sum3);
+                    }
+#else
                    #pragma unroll
                    for (int g = 0; g < 8; ++g) {
                        const uint qp = q_packed_pf[b_local * 8 + g];
@@ -1768,12 +1798,21 @@ __kernel void flash_attn_f32_q4_0(
                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+2][b * 8 + g], sum2);
                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+3][b * 8 + g], sum3);
                    }
+#endif
                    const float qd    = q_d_pf[b_local];
                    const int   q_sum = q_sum_pf[b_local];
+#ifdef FA_K_LDS_T
+                    const float4 ks4 = vload4(0, &l_k_scale[b][j]);
+                    s0 += (float)(sum0 - 8 * q_sum) * qd * ks4.s0;
+                    s1 += (float)(sum1 - 8 * q_sum) * qd * ks4.s1;
+                    s2 += (float)(sum2 - 8 * q_sum) * qd * ks4.s2;
+                    s3 += (float)(sum3 - 8 * q_sum) * qd * ks4.s3;
+#else
                    s0 += (float)(sum0 - 8 * q_sum) * qd * l_k_scale[j  ][b];
                    s1 += (float)(sum1 - 8 * q_sum) * qd * l_k_scale[j+1][b];
                    s2 += (float)(sum2 - 8 * q_sum) * qd * l_k_scale[j+2][b];
                    s3 += (float)(sum3 - 8 * q_sum) * qd * l_k_scale[j+3][b];
+#endif
                }
 #else
                ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
@@ -1393,8 +1393,31 @@ __kernel void flash_attn_f32_q8_0(
    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);

 #ifdef FA_HAVE_INT_DOT
+// Accessors so the staging code is layout-agnostic.
+#ifdef FA_K_LDS_T
+#define FA_K_PACKED(ROW, IDX) l_k_packed[IDX][ROW]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[BLK][ROW]
+#else
+#define FA_K_PACKED(ROW, IDX) l_k_packed[ROW][IDX]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[ROW][BLK]
+#endif
+
+#ifdef FA_K_LDS_T
+    // K tile transposed: [block*8 + g][kv row] instead of [kv row][block*8 + g].
+    //
+    // The QK loop walks 4 KV rows at a time against the same (b, g), so in the original
+    // layout those 4 values are BLOCK_N*8 uints apart and cost 4 separate 32-bit local
+    // reads. Transposed they are adjacent, so they are one 128-bit read -- 4x fewer LDS
+    // issues for the same bytes and no extra registers. That matters because the QK loop
+    // is LDS-read-issue-bound: a wrong-math probe that kept every dp4a but cut the LDS
+    // reads ran the whole kernel 41% faster (18.51 -> 10.91 ms/op), and deleting QK
+    // outright only reached 10.88 -- i.e. essentially ALL of QK's cost is these reads.
+    __local uint  l_k_packed[DK_Q8_BLOCKS_PREFILL * 8][BLOCK_N];
+    __local float l_k_scale [DK_Q8_BLOCKS_PREFILL][BLOCK_N];
+#else
    __local uint  l_k_packed[BLOCK_N][DK_Q8_BLOCKS_PREFILL * 8];
    __local float l_k_scale [BLOCK_N][DK_Q8_BLOCKS_PREFILL];
+#endif
 #else
    __local half4 l_k[BLOCK_N][DK_VEC];
 #endif
@@ -1427,7 +1450,7 @@ __kernel void flash_attn_f32_q8_0(
                    const global char * blk_ptr = k_base + k_row_off + blk * Q8_0_BLOCK_SIZE;
                    const float df = (float) vload_half(0, (const global half *) blk_ptr);
                    const global uchar * qs = (const global uchar *)(blk_ptr + 2);
-                    l_k_scale[row][blk] = df;
+                    FA_K_SCALE(row, blk) = df;
                    #pragma unroll
                    for (int j = 0; j < 8; ++j) {
                        uint k_packed =
@@ -1435,12 +1458,12 @@ __kernel void flash_attn_f32_q8_0(
                             ((uint) qs[j*4 + 1]) <<  8 |
                             ((uint) qs[j*4 + 2]) << 16 |
                             ((uint) qs[j*4 + 3]) << 24;
-                        l_k_packed[row][blk * 8 + j] = k_packed;
+                        FA_K_PACKED(row, blk * 8 + j) = k_packed;
                    }
                } else {
-                    l_k_scale[row][blk] = 0.0f;
+                    FA_K_SCALE(row, blk) = 0.0f;
                    #pragma unroll
-                    for (int j = 0; j < 8; ++j) l_k_packed[row][blk * 8 + j] = 0u;
+                    for (int j = 0; j < 8; ++j) FA_K_PACKED(row, blk * 8 + j) = 0u;
                }
            }
 #else
@@ -1556,6 +1579,19 @@ __kernel void flash_attn_f32_q8_0(
                for (int b_local = 0; b_local < SPLIT_DK_Q8_BLOCKS; ++b_local) {
                    const int b = k_blk_base + b_local;
                    int sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+#if defined(FA_K_LDS_T)
+                    // The 4 KV rows are adjacent in the transposed tile, so each (b, g)
+                    // step is ONE 128-bit local read instead of four 32-bit ones.
+                    #pragma unroll
+                    for (int g = 0; g < 8; ++g) {
+                        const uint qp = q_packed_pf[b_local * 8 + g];
+                        const uint4 kq4 = vload4(0, &l_k_packed[b * 8 + g][j]);
+                        sum0 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s0, sum0);
+                        sum1 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s1, sum1);
+                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s2, sum2);
+                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s3, sum3);
+                    }
+#else
                    #pragma unroll
                    for (int g = 0; g < 8; ++g) {
                        const uint qp = q_packed_pf[b_local * 8 + g];
@@ -1564,11 +1600,20 @@ __kernel void flash_attn_f32_q8_0(
                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+2][b * 8 + g], sum2);
                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+3][b * 8 + g], sum3);
                    }
+#endif
                    const float qd = q_d_pf[b_local];
+#ifdef FA_K_LDS_T
+                    const float4 ks4 = vload4(0, &l_k_scale[b][j]);
+                    s0 += (float)sum0 * qd * ks4.s0;
+                    s1 += (float)sum1 * qd * ks4.s1;
+                    s2 += (float)sum2 * qd * ks4.s2;
+                    s3 += (float)sum3 * qd * ks4.s3;
+#else
                    s0 += (float)sum0 * qd * l_k_scale[j  ][b];
                    s1 += (float)sum1 * qd * l_k_scale[j+1][b];
                    s2 += (float)sum2 * qd * l_k_scale[j+2][b];
                    s3 += (float)sum3 * qd * l_k_scale[j+3][b];
+#endif
                }
 #else
                ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
@@ -1022,9 +1022,20 @@ static T block_reduce(T val, T * shared_vals, int block_size_template) {
 }

 static __dpct_inline__ float ggml_sycl_ue4m3_to_fp32(uint8_t x) {
-    const uint32_t bits = x * (x != 0x7F && x != 0xFF);
-    const __nv_fp8_e4m3 xf = *reinterpret_cast<const __nv_fp8_e4m3 *>(&bits);
-    return static_cast<float>(xf) / 2;
+    // UE4M3 is unsigned: 4 exp bits (bias 7), 3 mantissa bits, no sign, no NaN.
+    // exp == 0xF is a valid exponent (256-448 range), not NaN.
+    if (x == 0 || x == 0x7F) {
+        return 0.0f;
+    }
+    const int exp = (x >> 3) & 0xF;
+    const int man = x & 0x7;
+    float raw;
+    if (exp == 0) {
+        raw = man * (1.0f / 8.0f) * sycl::pow(2.0f, -6.0f);
+    } else {
+        raw = (1.0f + man / 8.0f) * sycl::pow(2.0f, (float) exp - 7.0f);
+    }
+    return raw * 0.5f;
 }

 #endif // GGML_SYCL_COMMON_HPP
@@ -0,0 +1,280 @@
+#include "ggml-impl.h"
+#include "dsv4-hc.hpp"
+
+#include <cmath>
+
+static constexpr int DSV4_HC = 4;
+
+static void dsv4_hc_pre_f32_sycl(
+        const float * x, const float * weights, float * dst,
+        int64_t n_embd, int64_t hc, int64_t n_tokens,
+        int64_t sx0, int64_t sx1, int64_t sx2,
+        int64_t sw0, int64_t sw1,
+        int64_t sd0, int64_t sd1,
+        queue_ptr stream) {
+    const int64_t nr = n_embd * n_tokens;
+    const int64_t block_size = 256;
+    const int64_t num_blocks = (nr + block_size - 1) / block_size;
+
+    stream->parallel_for(
+        sycl::nd_range<1>(sycl::range<1>(num_blocks * block_size), sycl::range<1>(block_size)),
+        [=](sycl::nd_item<1> item) {
+            const int64_t ir = item.get_global_id(0);
+            if (ir >= nr) {
+                return;
+            }
+
+            const int64_t i0 = ir % n_embd;
+            const int64_t it = ir / n_embd;
+
+            float sum = x[i0*sx0 + it*sx2] * weights[it*sw1];
+            for (int64_t ih = 1; ih < hc; ++ih) {
+                const float xv = x[i0*sx0 + ih*sx1 + it*sx2];
+                const float wv = weights[ih*sw0 + it*sw1];
+                sum += xv * wv;
+            }
+
+            dst[i0*sd0 + it*sd1] = sum;
+        });
+}
+
+static void dsv4_hc_comb_norm_cols(float * comb, float eps) {
+    for (int idst = 0; idst < DSV4_HC; ++idst) {
+        float sum = eps;
+        for (int isrc = 0; isrc < DSV4_HC; ++isrc) {
+            sum += comb[idst + DSV4_HC*isrc];
+        }
+
+        const float inv_sum = 1.0f / sum;
+        for (int isrc = 0; isrc < DSV4_HC; ++isrc) {
+            comb[idst + DSV4_HC*isrc] *= inv_sum;
+        }
+    }
+}
+
+static void dsv4_hc_comb_norm_rows(float * comb, float eps) {
+    for (int isrc = 0; isrc < DSV4_HC; ++isrc) {
+        float sum = eps;
+        for (int idst = 0; idst < DSV4_HC; ++idst) {
+            sum += comb[idst + DSV4_HC*isrc];
+        }
+
+        const float inv_sum = 1.0f / sum;
+        for (int idst = 0; idst < DSV4_HC; ++idst) {
+            comb[idst + DSV4_HC*isrc] *= inv_sum;
+        }
+    }
+}
+
+static void dsv4_hc_comb_f32_sycl(
+        const float * mixes,
+        const float * scale,
+        const float * base,
+        float * dst,
+        int64_t n_tokens,
+        int64_t sm0,
+        int64_t sm1,
+        int64_t ss0,
+        int64_t sb0,
+        int64_t sd0,
+        int64_t sd1,
+        int64_t sd2,
+        float eps,
+        int32_t n_iter,
+        queue_ptr stream) {
+    constexpr int comb_offset = 2*DSV4_HC;
+
+    const int64_t block_size = 256;
+    const int64_t num_blocks = (n_tokens + block_size - 1) / block_size;
+
+    stream->parallel_for(
+        sycl::nd_range<1>(sycl::range<1>(num_blocks * block_size), sycl::range<1>(block_size)),
+        [=](sycl::nd_item<1> item_ct1) {
+            const int64_t it = item_ct1.get_global_id(0);
+
+            if (it >= n_tokens) {
+                return;
+            }
+
+            const float scale_comb = scale[2*ss0];
+            float comb[DSV4_HC*DSV4_HC];
+
+            for (int isrc = 0; isrc < DSV4_HC; ++isrc) {
+                float max = -INFINITY;
+                for (int idst = 0; idst < DSV4_HC; ++idst) {
+                    const int idx = idst + DSV4_HC*isrc;
+                    const float v = mixes[(comb_offset + idx)*sm0 + it*sm1] * scale_comb + base[(comb_offset + idx)*sb0];
+                    comb[idx] = v;
+                    max = fmaxf(max, v);
+                }
+
+                float sum = 0.0f;
+                for (int idst = 0; idst < DSV4_HC; ++idst) {
+                    const int idx = idst + DSV4_HC*isrc;
+                    const float v = expf(comb[idx] - max);
+                    comb[idx] = v;
+                    sum += v;
+                }
+
+                const float inv_sum = 1.0f / sum;
+                for (int idst = 0; idst < DSV4_HC; ++idst) {
+                    const int idx = idst + DSV4_HC*isrc;
+                    comb[idx] = comb[idx] * inv_sum + eps;
+                }
+            }
+
+            dsv4_hc_comb_norm_cols(comb, eps);
+            for (int32_t i = 1; i < n_iter; ++i) {
+                dsv4_hc_comb_norm_rows(comb, eps);
+                dsv4_hc_comb_norm_cols(comb, eps);
+            }
+
+            for (int isrc = 0; isrc < DSV4_HC; ++isrc) {
+                for (int idst = 0; idst < DSV4_HC; ++idst) {
+                    const int idx = idst + DSV4_HC*isrc;
+                    dst[idst*sd0 + isrc*sd1 + it*sd2] = comb[idx];
+                }
+            }
+        });
+}
+
+static void dsv4_hc_post_f32_sycl(
+        const float * x, const float * residual, const float * post, const float * comb, float * dst,
+        int64_t n_embd, int64_t hc, int64_t n_tokens,
+        int64_t sx0, int64_t sx1,
+        int64_t sr0, int64_t sr1, int64_t sr2,
+        int64_t sp0, int64_t sp1,
+        int64_t sc0, int64_t sc1, int64_t sc2,
+        int64_t sd0, int64_t sd1, int64_t sd2,
+        queue_ptr stream) {
+    const int64_t nr = n_embd * hc * n_tokens;
+    const int64_t block_size = 256;
+    const int64_t num_blocks = (nr + block_size - 1) / block_size;
+
+    stream->parallel_for(
+        sycl::nd_range<1>(sycl::range<1>(num_blocks * block_size), sycl::range<1>(block_size)),
+        [=](sycl::nd_item<1> item) {
+            const int64_t ir = item.get_global_id(0);
+            if (ir >= nr) {
+                return;
+            }
+
+            const int64_t i0   = ir % n_embd;
+            const int64_t idst = (ir / n_embd) % hc;
+            const int64_t it   = ir / (n_embd * hc);
+
+            float sum = x[i0*sx0 + it*sx1] * post[idst*sp0 + it*sp1];
+            for (int64_t isrc = 0; isrc < hc; ++isrc) {
+                sum += residual[i0*sr0 + isrc*sr1 + it*sr2] * comb[idst*sc0 + isrc*sc1 + it*sc2];
+            }
+
+            dst[i0*sd0 + idst*sd1 + it*sd2] = sum;
+        });
+}
+
+void ggml_sycl_op_dsv4_hc_pre(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * x       = dst->src[0];
+    const ggml_tensor * weights = dst->src[1];
+
+    GGML_ASSERT(x->type == GGML_TYPE_F32);
+    GGML_ASSERT(weights->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_LOCALS(size_t, nbx, x,       nb);
+    GGML_TENSOR_LOCALS(size_t, nbw, weights, nb);
+    GGML_TENSOR_LOCALS(size_t, nbd, dst,     nb);
+
+    const int64_t n_embd   = x->ne[0];
+    const int64_t hc       = x->ne[1];
+    const int64_t n_tokens = x->ne[2];
+
+    queue_ptr stream = ctx.stream();
+
+    dsv4_hc_pre_f32_sycl(
+            (const float *) x->data, (const float *) weights->data, (float *) dst->data,
+            n_embd, hc, n_tokens,
+            nbx0 / sizeof(float), nbx1 / sizeof(float), nbx2 / sizeof(float),
+            nbw0 / sizeof(float), nbw1 / sizeof(float),
+            nbd0 / sizeof(float), nbd1 / sizeof(float),
+            stream);
+}
+
+void ggml_sycl_op_dsv4_hc_comb(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/3);
+
+    const ggml_tensor * mixes = dst->src[0];
+    const ggml_tensor * scale = dst->src[1];
+    const ggml_tensor * base  = dst->src[2];
+
+    GGML_ASSERT(mixes->type == GGML_TYPE_F32);
+    GGML_ASSERT(scale->type == GGML_TYPE_F32);
+    GGML_ASSERT(base->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    constexpr int64_t hc_mix_dim = (2 + DSV4_HC)*DSV4_HC;
+
+    GGML_ASSERT(mixes->ne[0] == hc_mix_dim);
+    GGML_ASSERT(dst->ne[0] == DSV4_HC);
+    GGML_ASSERT(dst->ne[1] == DSV4_HC);
+    GGML_ASSERT(dst->ne[2] == mixes->ne[1]);
+    GGML_ASSERT(scale->ne[0] >= 3);
+    GGML_ASSERT(base->ne[0] == hc_mix_dim);
+
+    GGML_TENSOR_LOCALS(size_t, nbm, mixes, nb);
+    GGML_TENSOR_LOCALS(size_t, nbs, scale, nb);
+    GGML_TENSOR_LOCALS(size_t, nbb, base,  nb);
+    GGML_TENSOR_LOCALS(size_t, nbd, dst,   nb);
+
+    const int64_t n_tokens = mixes->ne[1];
+    const float eps = ggml_get_op_params_f32(dst, 0);
+    const int32_t n_iter = ggml_get_op_params_i32(dst, 1);
+
+    queue_ptr stream = ctx.stream();
+
+    dsv4_hc_comb_f32_sycl(
+            (const float *) mixes->data, (const float *) scale->data, (const float *) base->data, (float *) dst->data,
+            n_tokens,
+            nbm0 / sizeof(float), nbm1 / sizeof(float),
+            nbs0 / sizeof(float),
+            nbb0 / sizeof(float),
+            nbd0 / sizeof(float), nbd1 / sizeof(float), nbd2 / sizeof(float),
+            eps, n_iter, stream);
+}
+
+void ggml_sycl_op_dsv4_hc_post(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/4);
+    const ggml_tensor * x        = dst->src[0];
+    const ggml_tensor * residual = dst->src[1];
+    const ggml_tensor * post     = dst->src[2];
+    const ggml_tensor * comb     = dst->src[3];
+
+    GGML_ASSERT(x->type == GGML_TYPE_F32);
+    GGML_ASSERT(residual->type == GGML_TYPE_F32);
+    GGML_ASSERT(post->type == GGML_TYPE_F32);
+    GGML_ASSERT(comb->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_LOCALS(size_t, nbx, x,        nb);
+    GGML_TENSOR_LOCALS(size_t, nbr, residual, nb);
+    GGML_TENSOR_LOCALS(size_t, nbp, post,     nb);
+    GGML_TENSOR_LOCALS(size_t, nbc, comb,     nb);
+    GGML_TENSOR_LOCALS(size_t, nbd, dst,      nb);
+
+    const int64_t n_embd   = x->ne[0];
+    const int64_t n_tokens = x->ne[1];
+    const int64_t hc       = residual->ne[1];
+
+    queue_ptr stream = ctx.stream();
+
+    dsv4_hc_post_f32_sycl(
+            (const float *) x->data, (const float *) residual->data,
+            (const float *) post->data, (const float *) comb->data, (float *) dst->data,
+            n_embd, hc, n_tokens,
+            nbx0 / sizeof(float), nbx1 / sizeof(float),
+            nbr0 / sizeof(float), nbr1 / sizeof(float), nbr2 / sizeof(float),
+            nbp0 / sizeof(float), nbp1 / sizeof(float),
+            nbc0 / sizeof(float), nbc1 / sizeof(float), nbc2 / sizeof(float),
+            nbd0 / sizeof(float), nbd1 / sizeof(float), nbd2 / sizeof(float),
+            stream);
+}
@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_DSV4_HC_HPP
+#define GGML_SYCL_DSV4_HC_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_dsv4_hc_pre(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_op_dsv4_hc_comb(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_op_dsv4_hc_post(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_DSV4_HC_HPP
@@ -420,53 +420,31 @@ static void clamp(const T * x, T * dst, const float min, const float max, const
    }
 }

-template<typename T>
-static void gated_op_fused_geglu(const T * x, const T * g, T * dst, const uint64_t k, const sycl::uint3 n_fd, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+template<typename T, typename F>
+static void unary_gated_op_flat_kernel(const T * x, const T * g, T * dst, const uint64_t k, const sycl::nd_item<1> & item_ct1, F func) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = func(x[i]) * g[i];
+    }
+}
+
+template<typename T, typename F>
+static void unary_gated_op_generic_kernel(
+        const T * x,
+        const T * g,
+        T * dst,
+        const uint64_t k,
+        const sycl::uint3 n_fd,
+        const uint64_t o0,
+        const uint64_t o1,
+        const sycl::nd_item<1> & item_ct1,
+        F func) {
+
+    // rows of n columns at strides o0 and o1: two halves of one fused tensor, or two tensors
    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
        const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
        const int64_t j0 = rc.x() * o0 + rc.y();
        const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
-        dst[i] = op_gelu(x[j0]) * g[j1];
-    }
-}
-
-template<typename T>
-static void gated_op_fused_reglu(const T * x, const T * g, T * dst, const uint64_t k, const sycl::uint3 n_fd, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
-        const int64_t j0 = rc.x() * o0 + rc.y();
-        const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
-        dst[i] = op_relu(x[j0]) * g[j1];
-    }
-}
-
-template<typename T>
-static void gated_op_fused_swiglu(const T * x, const T * g, T * dst, const uint64_t k, const sycl::uint3 n_fd, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1)  {
-        const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
-        const int64_t j0 = rc.x() * o0 + rc.y();
-        const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
-        dst[i] = op_silu(x[j0]) * g[j1];
-    }
-}
-
-template<typename T>
-static void gated_op_fused_geglu_erf(const T * x, const T * g, T * dst, const uint64_t k, const sycl::uint3 n_fd, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
-        const int64_t j0 = rc.x() * o0 + rc.y();
-        const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
-        dst[i] = op_gelu_erf(x[j0]) * g[j1];
-    }
-}
-
-template<typename T>
-static void gated_op_fused_geglu_quick(const T * x, const T * g, T * dst, const uint64_t k, const sycl::uint3 n_fd, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
-        const int64_t j0 = rc.x() * o0 + rc.y();
-        const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
-        dst[i] = op_gelu_quick(x[j0]) * g[j1];
+        dst[i] = func(x[j0]) * g[j1];
    }
 }

@@ -670,6 +648,35 @@ static inline void ggml_sycl_op_unary(
        });
 }

+template<typename F>
+static inline void ggml_sycl_op_unary_gated(
+        ggml_backend_sycl_context & ctx, ggml_tensor * dst, F func) {
+
+    dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [func](const auto * x_ptr, const auto * g_ptr, auto * dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+
+            const uint32_t num_blocks = (uint32_t) ceil_div(k, SYCL_GLU_BLOCK_SIZE);
+            const sycl::nd_range<1> launch_range(num_blocks * sycl::range<1>(SYCL_GLU_BLOCK_SIZE),
+                                                 sycl::range<1>(SYCL_GLU_BLOCK_SIZE));
+
+            // o0 == n and o1 == n make the index math the identity, so index flat
+            // note: not ggml_is_contiguous - a fused [gate|up] src0 is contiguous with o0 == 2n
+            if (o0 == n && o1 == n) {
+                main_stream->parallel_for(launch_range,
+                    [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        unary_gated_op_flat_kernel(x_ptr, g_ptr, dst_ptr, k, item_ct1, func);
+                    });
+            } else {
+                // launch-invariant divisor, and only this path needs it
+                const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
+                main_stream->parallel_for(launch_range,
+                    [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        unary_gated_op_generic_kernel(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1, func);
+                    });
+            }
+        });
+}
+

 static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    GGML_ASSERT(dst->type == GGML_TYPE_F32);
@@ -967,42 +974,21 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
 }

 static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
-        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
-            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
-            const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
-            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
-                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1);
-            });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
+        return op_gelu(x);
+    });
 }

 static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
-        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
-            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_RELU_BLOCK_SIZE); // Using RELU block size for reglu
-            const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
-            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
-                    sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1);
-            });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
+        return op_relu(x);
+    });
 }

 static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
-        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
-            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_SILU_BLOCK_SIZE); // Using SILU block size for swiglu
-            const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
-            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
-                    sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1);
-            });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
+        return op_silu(x);
+    });
 }

 __dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {
@@ -1097,29 +1083,15 @@ void ggml_sycl_op_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
 }

 static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
-        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
-            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
-            const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
-            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
-                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1);
-            });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
+        return op_gelu_erf(x);
+    });
 }

 static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
-        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
-            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
-            const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
-            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
-                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1);
-            });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
+        return op_gelu_quick(x);
+    });
 }


@@ -73,6 +73,7 @@ static void flash_attn_ext_vec(const char* __restrict__ Q,
                        const int32_t nb31,
                        const int32_t nb32,
                        const int64_t nb33) {
+
 #ifdef SYCL_FLASH_ATTN
    // Skip unused kernel variants for faster compilation:

@@ -469,7 +470,6 @@ static void flash_attn_ext_vec(const char* __restrict__ Q,
        }
    }

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

 #pragma unroll
@@ -591,22 +591,24 @@ void ggml_sycl_flash_attn_ext_vec_case_impl(ggml_backend_sycl_context & ctx, ggm

    const auto arch = ggml_sycl_info().devices[ctx.device].hw_info.arch;
    const int nthreads = ggml_sycl_fattn_vec_get_nthreads_device(arch);
-    // 256 threads would overflow the 64 KB work-group local memory at D == 512, so keep 128 there.
-    if (D <= 256 && nthreads == 256) {
-        constexpr int nthreads_hw = 256;
-        constexpr int nwarps = nthreads_hw / warp_size;
-        launch_fattn<D, cols_per_block, 1,
-                     flash_attn_ext_vec<D, cols_per_block, type_K, type_V,
-                                        use_logit_softcap, warp_size, nthreads_hw>, warp_size>(
-            ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
-    } else {
-        constexpr int nthreads_hw = 128;
-        constexpr int nwarps = nthreads_hw / warp_size;
-        launch_fattn<D, cols_per_block, 1,
-                     flash_attn_ext_vec<D, cols_per_block, type_K, type_V,
-                                        use_logit_softcap, warp_size, nthreads_hw>, warp_size>(
-            ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+    if constexpr (D <= 256) {
+        if (nthreads == 256) {
+            constexpr int nthreads_hw = 256;
+            constexpr int nwarps = nthreads_hw / warp_size;
+            launch_fattn<D, cols_per_block, 1,
+                         flash_attn_ext_vec<D, cols_per_block, type_K, type_V,
+                                            use_logit_softcap, warp_size, nthreads_hw>, warp_size>(
+                ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+            return;
+        }
    }
+
+    constexpr int nthreads_hw = 128;
+    constexpr int nwarps = nthreads_hw / warp_size;
+    launch_fattn<D, cols_per_block, 1,
+                 flash_attn_ext_vec<D, cols_per_block, type_K, type_V,
+                                    use_logit_softcap, warp_size, nthreads_hw>, warp_size>(
+        ctx, dst, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }

 template <int D, int type_K, int type_V>
@@ -62,6 +62,8 @@
 #include "ggml-sycl/repeat_back.hpp"
 #include "ggml-sycl/set_rows.hpp"
 #include "ggml-sycl/set.hpp"
+#include "ggml-sycl/dsv4-hc.hpp"
+#include "ggml-sycl/lightning-indexer.hpp"
 #include "ggml-sycl/conv2d.hpp"
 #include "ggml-sycl/conv2d-dw.hpp"
 #include "ggml-sycl/conv2d-transpose.hpp"
@@ -4942,6 +4944,18 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
        case GGML_OP_SET_ROWS:
            ggml_sycl_op_set_rows(ctx, dst);
            break;
+        case GGML_OP_DSV4_HC_PRE:
+            ggml_sycl_op_dsv4_hc_pre(ctx, dst);
+            break;
+        case GGML_OP_DSV4_HC_COMB:
+            ggml_sycl_op_dsv4_hc_comb(ctx, dst);
+            break;
+        case GGML_OP_DSV4_HC_POST:
+            ggml_sycl_op_dsv4_hc_post(ctx, dst);
+            break;
+        case GGML_OP_LIGHTNING_INDEXER:
+            ggml_sycl_op_lightning_indexer(ctx, dst);
+            break;
        case GGML_OP_DUP:
            ggml_sycl_dup(ctx, dst);
            break;
@@ -5795,17 +5809,33 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons

        case GGML_OP_SET_ROWS:
            {
-
-                auto res = ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
-                         op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q5_0 ||
-                         op->type == GGML_TYPE_Q1_0 ||
-                         op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_IQ4_NL ||
-                         op->type == GGML_TYPE_MXFP4 || op->type == GGML_TYPE_NVFP4) &&
-                        op->src[0]->type == GGML_TYPE_F32 &&
-                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32));
+                auto res = (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 ||
+                            op->src[0]->type == GGML_TYPE_BF16) &&
+                           (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
                return res;
            }
            break;
+        case GGML_OP_DSV4_HC_PRE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
+                op->type == GGML_TYPE_F32;
+        case GGML_OP_DSV4_HC_COMB:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_DSV4_HC_POST:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 && op->src[3]->type == GGML_TYPE_F32 &&
+                op->type == GGML_TYPE_F32;
+        case GGML_OP_LIGHTNING_INDEXER:
+            return op->src[0]->type == GGML_TYPE_F32 &&
+                (op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32 ||
+                 op->src[1]->type == GGML_TYPE_BF16 || op->src[1]->type == GGML_TYPE_Q8_0 ||
+                 op->src[1]->type == GGML_TYPE_Q5_1 || op->src[1]->type == GGML_TYPE_Q5_0 ||
+                 op->src[1]->type == GGML_TYPE_Q4_1 || op->src[1]->type == GGML_TYPE_Q4_0 ||
+                 op->src[1]->type == GGML_TYPE_IQ4_NL) &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->src[3]->type == GGML_TYPE_F16 &&
+                op->type == GGML_TYPE_F32 &&
+                op->src[0]->ne[0] == WARP_SIZE * 8;
        case GGML_OP_CPY:
            {
                ggml_type src0_type = op->src[0]->type;
@@ -0,0 +1,197 @@
+#include "lightning-indexer.hpp"
+#include "dequantize.hpp"
+
+static void lightning_indexer_f32_sycl(
+        const char * q, const char * k, const char * w, const char * m, float * dst,
+        int64_t n_embd, int64_t n_head, int64_t n_batch, int64_t n_stream, int64_t n_kv,
+        int64_t nem3,
+        int64_t nbq1, int64_t nbq2, int64_t nbq3,
+        int64_t nbk2, int64_t nbk3,
+        int64_t nbw1, int64_t nbw3,
+        int64_t nbm1, int64_t nbm3,
+        int64_t nb1, int64_t nb3,
+        ggml_type k_type,
+        queue_ptr stream) {
+
+    constexpr int64_t LANES = WARP_SIZE;
+    constexpr int64_t ELEMS_PER_LANE = 8;
+    constexpr int64_t ROWS_PER_BLOCK = 4;
+    constexpr int64_t BLOCK_SIZE = ROWS_PER_BLOCK * LANES;
+
+    const int64_t n_rows = n_batch * n_stream * n_kv;
+    const int64_t n_blocks = (n_rows + ROWS_PER_BLOCK - 1) / ROWS_PER_BLOCK;
+
+    stream->parallel_for(
+        sycl::nd_range<1>(
+            sycl::range<1>(n_blocks * BLOCK_SIZE),
+            sycl::range<1>(BLOCK_SIZE)),
+        [=](sycl::nd_item<1> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+            const int64_t ir   = item.get_global_id(0);
+            const int64_t lane = ir % LANES;
+            const int64_t row  = ir / LANES;
+            if (row >= n_rows) {
+                return;
+            }
+
+            const int64_t i_bs     = row / n_kv;
+            const int64_t i_kv     = row % n_kv;
+            const int64_t i_batch  = i_bs / n_stream;
+            const int64_t i_stream = i_bs % n_stream;
+
+            // load K row slice into registers (row is contiguous, nbk0 == type size)
+            const char * k_base = k + i_kv*nbk2 + i_stream*nbk3;
+            float k_local[ELEMS_PER_LANE];
+            if (k_type == GGML_TYPE_F16) {
+                const sycl::half * k_row = (const sycl::half *) k_base;
+#pragma unroll
+                for (int64_t j = 0; j < ELEMS_PER_LANE; ++j) {
+                    k_local[j] = static_cast<float>(k_row[lane*ELEMS_PER_LANE + j]);
+                }
+            } else if (k_type == GGML_TYPE_F32) {
+                const float * k_row = (const float *) k_base;
+#pragma unroll
+                for (int64_t j = 0; j < ELEMS_PER_LANE; ++j) {
+                    k_local[j] = k_row[lane*ELEMS_PER_LANE + j];
+                }
+            } else {
+                const int64_t lane_base = lane * ELEMS_PER_LANE;
+                switch (k_type) {
+                    case GGML_TYPE_BF16: {
+                        const sycl::ext::oneapi::bfloat16 * k_row = (const sycl::ext::oneapi::bfloat16 *) k_base;
+#pragma unroll
+                        for (int64_t j = 0; j < ELEMS_PER_LANE; ++j) {
+                            k_local[j] = static_cast<float>(k_row[lane_base + j]);
+                        }
+                    } break;
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1: {
+#pragma unroll
+                        for (int64_t j = 0; j < ELEMS_PER_LANE; ++j) {
+                            const int64_t idx = lane_base + j;
+                            const int64_t ib  = idx / QK4_0;
+                            const int iqs     = idx % (QK4_0/2);
+                            dfloat2 kv;
+                            if (k_type == GGML_TYPE_Q4_0) {
+                                dequantize_q4_0(k_base, ib, iqs, kv);
+                            } else if (k_type == GGML_TYPE_Q4_1) {
+                                dequantize_q4_1(k_base, ib, iqs, kv);
+                            } else if (k_type == GGML_TYPE_Q5_0) {
+                                dequantize_q5_0(k_base, ib, iqs, kv);
+                            } else {
+                                dequantize_q5_1(k_base, ib, iqs, kv);
+                            }
+                            k_local[j] = (idx % QK4_0) < (QK4_0/2) ? static_cast<float>(kv.x()) : static_cast<float>(kv.y());
+                        }
+                    } break;
+                    case GGML_TYPE_Q8_0: {
+#pragma unroll
+                        for (int64_t pair = 0; pair < ELEMS_PER_LANE / 2; ++pair) {
+                            const int64_t elem0 = lane_base + 2 * pair;
+                            dfloat2 kv;
+                            dequantize_q8_0(k_base, elem0 / QK8_0, elem0 % QK8_0, kv);
+                            k_local[2 * pair + 0] = static_cast<float>(kv.x());
+                            k_local[2 * pair + 1] = static_cast<float>(kv.y());
+                        }
+                    } break;
+                    case GGML_TYPE_IQ4_NL: {
+#pragma unroll
+                        for (int64_t pair = 0; pair < ELEMS_PER_LANE / 2; ++pair) {
+                            const int64_t elem0 = lane_base + 2 * pair;
+                            dfloat2 kv;
+                            dequantize_iq4_nl(k_base, elem0 / QK4_NL, elem0 % QK4_NL, kv);
+                            k_local[2 * pair + 0] = static_cast<float>(kv.x());
+                            k_local[2 * pair + 1] = static_cast<float>(kv.y());
+                        }
+                    } break;
+                    default:
+#pragma unroll
+                        for (int64_t j = 0; j < ELEMS_PER_LANE; ++j) {
+                            k_local[j] = 0.0f;
+                        }
+                        break;
+                }
+            }
+
+            const char  * q_base = q + i_batch*nbq2 + i_stream*nbq3;
+            const float * w_base = (const float *) (w + i_batch*nbw1 + i_stream*nbw3);
+
+            float score = 0.0f;
+            for (int64_t h = 0; h < n_head; ++h) {
+                const float * q_row = (const float *) (q_base + h*nbq1);
+                float dot = 0.0f;
+#pragma unroll
+                for (int64_t j = 0; j < ELEMS_PER_LANE; ++j) {
+                    const int64_t i = lane*ELEMS_PER_LANE + j;
+                    if (i < n_embd) {
+                        dot += q_row[i] * k_local[j];
+                    }
+                }
+                dot = sycl::reduce_over_group(item.get_sub_group(), dot, sycl::plus<float>());
+                if (lane == 0) {
+                    score += sycl::max(dot, 0.0f) * w_base[h];
+                }
+            }
+
+            if (lane == 0) {
+                const sycl::half * m_base = (const sycl::half *) (m + i_batch*nbm1 + (i_stream % nem3)*nbm3);
+                // flat-index store: storing through a strided base pointer
+                // hangs/misroutes writes on this stack when n_batch*n_stream > 1
+                const int64_t dst_idx = i_kv + i_batch*(nb1/sizeof(float)) + i_stream*(nb3/sizeof(float));
+                dst[dst_idx] = score + static_cast<float>(m_base[i_kv]);
+            }
+        });
+}
+
+void ggml_sycl_op_lightning_indexer(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/4);
+    const ggml_tensor * q = dst->src[0];
+    const ggml_tensor * k = dst->src[1];
+    const ggml_tensor * w = dst->src[2]; // weights
+    const ggml_tensor * m = dst->src[3]; // mask
+
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(  q->type == GGML_TYPE_F32);
+    GGML_ASSERT(  w->type == GGML_TYPE_F32);
+    GGML_ASSERT(  m->type == GGML_TYPE_F16);
+    GGML_ASSERT(k->type == GGML_TYPE_F16 || k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_BF16 ||
+                k->type == GGML_TYPE_Q8_0 || k->type == GGML_TYPE_Q5_1 || k->type == GGML_TYPE_Q5_0 ||
+                k->type == GGML_TYPE_Q4_1 || k->type == GGML_TYPE_Q4_0 || k->type == GGML_TYPE_IQ4_NL);
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbq, q, nb);
+    GGML_TENSOR_LOCALS(int64_t, nek, k, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbk, k, nb);
+    GGML_TENSOR_LOCALS(size_t,  nbw, w, nb);
+    GGML_TENSOR_LOCALS(int64_t, nem, m, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbm, m, nb);
+    GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+    GGML_TENSOR_LOCALS(size_t,  nb, dst, nb);
+
+    // input rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbm0 == ggml_type_size(m->type));
+    GGML_ASSERT(nb0  == ggml_type_size(dst->type));
+
+    const int64_t n_embd   = neq0;
+    const int64_t n_head   = neq1;
+    const int64_t n_batch  = neq2;
+    const int64_t n_stream = neq3;
+    const int64_t n_kv     = nek2;
+
+    GGML_ASSERT(n_embd == WARP_SIZE * 8);
+
+    lightning_indexer_f32_sycl(
+            (const char *) q->data, (const char *) k->data,
+            (const char *) w->data, (const char *) m->data, (float *) dst->data,
+            n_embd, n_head, n_batch, n_stream, n_kv, nem3,
+            nbq1, nbq2, nbq3,
+            nbk2, nbk3,
+            nbw1, nbw3,
+            nbm1, nbm3,
+            nb1, nb3,
+            k->type,
+            ctx.stream());
+}
@@ -0,0 +1,8 @@
+#ifndef GGML_SYCL_LIGHTNING_INDEXER_HPP
+#define GGML_SYCL_LIGHTNING_INDEXER_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_lightning_indexer(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_LIGHTNING_INDEXER_HPP
@@ -20,8 +20,6 @@
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses

 #define SYCL_COL2IM_1D_BLOCK_SIZE 256
-#define SYCL_GELU_BLOCK_SIZE 256
-#define SYCL_SILU_BLOCK_SIZE 256
 #define SYCL_TANH_BLOCK_SIZE 256
 #define SYCL_RELU_BLOCK_SIZE 256
 #define SYCL_HARDSIGMOID_BLOCK_SIZE 256
@@ -1,6 +1,10 @@
 #include "set_rows.hpp"
 #include "cpy.hpp"

+#include "ggml-quants.h"
+
+#include <vector>
+
 namespace utils {
 template<typename T>
 static constexpr bool is_arithmetic_v() {
@@ -20,7 +24,17 @@ convert (const char* src, char* dst) {
   *reinterpret_cast<TOut*>(dst) = dst_val;
 }

-template <typename TIdx, typename blockType, int qk, cpy_kernel_t cpyblck>
+#ifdef GGML_SYCL_HAS_BF16
+// sycl::vec::convert does not provide a half -> bfloat16 path, so route through float.
+template<>
+inline void convert<sycl::half, sycl::ext::oneapi::bfloat16>(const char* src, char* dst) {
+    const float tmp = sycl::vec<sycl::half, 1>(*reinterpret_cast<const sycl::half*>(src))
+                          .template convert<float, sycl::rounding_mode::automatic>()[0];
+    *reinterpret_cast<sycl::ext::oneapi::bfloat16*>(dst) = sycl::ext::oneapi::bfloat16(tmp);
+}
+#endif
+
+template <typename TIn, typename TIdx, typename blockType, int qk, cpy_kernel_t cpyblck>
 static void set_rows_sycl_q(const char * __restrict__ src0_d,
                            const TIdx * __restrict__ src1_d,
                            blockType * __restrict__ dst_d,
@@ -68,13 +82,22 @@ static void set_rows_sycl_q(const char * __restrict__ src0_d,
        const int64_t i11         = i02 % ne11;
        const int64_t i10         = i01;
        const size_t  src_offset  = calculate_offset<3>({ nb01, nb02, nb03 }, { i01, i02, i03 });
-        const char *  src_block   = src0_d + src_offset + i00 * sizeof(float);
+        const char *  src_block   = src0_d + src_offset + i00 * sizeof(TIn);
        const size_t  src1_offset = calculate_offset<3>({ nb10, nb11, nb12 }, { i10, i11, i12 });
        const int64_t dst_row     = src1_d[src1_offset / sizeof(TIdx)];
        const size_t  dst_offset =
            calculate_offset<3>({ nb1, nb2, nb3 }, { dst_row, i02, i03 }) + (i00 / qk) * sizeof(blockType);
        char * dst_block = reinterpret_cast<char *>(reinterpret_cast<char *>(dst_d) + dst_offset);
-        cpyblck(src_block, dst_block);
+        if constexpr (std::is_same_v<TIn, float>) {
+            cpyblck(src_block, dst_block);
+        } else {
+            float src_block_f32[qk];
+            const TIn * src_block_t = reinterpret_cast<const TIn *>(src_block);
+            for (int j = 0; j < qk; ++j) {
+                src_block_f32[j] = (float) src_block_t[j];
+            }
+            cpyblck(reinterpret_cast<const char *>(src_block_f32), dst_block);
+        }
    });
    GGML_UNUSED(ne10);
    GGML_UNUSED(ne13);
@@ -82,6 +105,139 @@ static void set_rows_sycl_q(const char * __restrict__ src0_d,
    GGML_UNUSED(nb13);
 }

+template<typename blockType>
+using quantize_row_qk_t = void (*)(const float *, blockType *, int64_t);
+
+using quantize_rows_f_t = size_t (*)(const float *, void *, int64_t, int64_t, const float *);
+
+template <typename TIn, typename TIdx, typename blockType, int qk, quantize_row_qk_t<blockType> quantize_row>
+static void set_rows_sycl_qk_host(
+        const ggml_tensor * src0,
+        const ggml_tensor * src1,
+        ggml_tensor * dst,
+        const int64_t ne00,
+        const int64_t ne01,
+        const int64_t ne02,
+        const int64_t ne03,
+        const int64_t ne11,
+        const int64_t ne12,
+        const size_t nb01,
+        const size_t nb02,
+        const size_t nb03,
+        const size_t nb10,
+        const size_t nb11,
+        const size_t nb12,
+        const size_t nb1,
+        const size_t nb2,
+        const size_t nb3,
+        queue_ptr stream) {
+    GGML_ASSERT(ne00 % qk == 0);
+
+    const size_t src0_bytes = ggml_nbytes(src0);
+    const size_t src1_bytes = ggml_nbytes(src1);
+
+    std::vector<char> src0_host(src0_bytes);
+    std::vector<char> src1_host(src1_bytes);
+
+    stream->memcpy(src0_host.data(), src0->data, src0_bytes);
+    stream->memcpy(src1_host.data(), src1->data, src1_bytes);
+    stream->wait();
+
+    std::vector<float> src_row_f32(ne00);
+    const int64_t nblocks = ne00 / qk;
+    std::vector<blockType> dst_row_q(nblocks);
+
+    for (int64_t i03 = 0; i03 < ne03; ++i03) {
+        for (int64_t i02 = 0; i02 < ne02; ++i02) {
+            for (int64_t i01 = 0; i01 < ne01; ++i01) {
+                const int64_t i12 = i03 % ne12;
+                const int64_t i11 = i02 % ne11;
+                const int64_t i10 = i01;
+
+                const size_t src1_offset = calculate_offset<3>({ nb10, nb11, nb12 }, { i10, i11, i12 });
+                const int64_t dst_row = *(const TIdx *) (src1_host.data() + src1_offset);
+
+                const size_t src0_row_offset = calculate_offset<3>({ nb01, nb02, nb03 }, { i01, i02, i03 });
+                const TIn * src_row = reinterpret_cast<const TIn *>(src0_host.data() + src0_row_offset);
+
+                for (int64_t i00 = 0; i00 < ne00; ++i00) {
+                    src_row_f32[i00] = (float) src_row[i00];
+                }
+
+                quantize_row(src_row_f32.data(), dst_row_q.data(), ne00);
+
+                const size_t dst_offset = calculate_offset<3>({ nb1, nb2, nb3 }, { dst_row, i02, i03 });
+                stream->memcpy((char *) dst->data + dst_offset, dst_row_q.data(), nblocks * sizeof(blockType));
+                stream->wait();
+            }
+        }
+    }
+}
+
+template <typename TIn, typename TIdx, typename blockType, int qk, quantize_rows_f_t quantize_rows>
+static void set_rows_sycl_iq_host(
+        const ggml_tensor * src0,
+        const ggml_tensor * src1,
+        ggml_tensor * dst,
+        const int64_t ne00,
+        const int64_t ne01,
+        const int64_t ne02,
+        const int64_t ne03,
+        const int64_t ne11,
+        const int64_t ne12,
+        const size_t nb01,
+        const size_t nb02,
+        const size_t nb03,
+        const size_t nb10,
+        const size_t nb11,
+        const size_t nb12,
+        const size_t nb1,
+        const size_t nb2,
+        const size_t nb3,
+        queue_ptr stream) {
+    GGML_ASSERT(ne00 % qk == 0);
+
+    const size_t src0_bytes = ggml_nbytes(src0);
+    const size_t src1_bytes = ggml_nbytes(src1);
+
+    std::vector<char> src0_host(src0_bytes);
+    std::vector<char> src1_host(src1_bytes);
+
+    stream->memcpy(src0_host.data(), src0->data, src0_bytes);
+    stream->memcpy(src1_host.data(), src1->data, src1_bytes);
+    stream->wait();
+
+    std::vector<float> src_row_f32(ne00);
+    const int64_t nblocks = ne00 / qk;
+    std::vector<blockType> dst_row_q(nblocks);
+
+    for (int64_t i03 = 0; i03 < ne03; ++i03) {
+        for (int64_t i02 = 0; i02 < ne02; ++i02) {
+            for (int64_t i01 = 0; i01 < ne01; ++i01) {
+                const int64_t i12 = i03 % ne12;
+                const int64_t i11 = i02 % ne11;
+                const int64_t i10 = i01;
+
+                const size_t src1_offset = calculate_offset<3>({ nb10, nb11, nb12 }, { i10, i11, i12 });
+                const int64_t dst_row = *(const TIdx *) (src1_host.data() + src1_offset);
+
+                const size_t src0_row_offset = calculate_offset<3>({ nb01, nb02, nb03 }, { i01, i02, i03 });
+                const TIn * src_row = reinterpret_cast<const TIn *>(src0_host.data() + src0_row_offset);
+
+                for (int64_t i00 = 0; i00 < ne00; ++i00) {
+                    src_row_f32[i00] = (float) src_row[i00];
+                }
+
+                quantize_rows(src_row_f32.data(), dst_row_q.data(), 1, ne00, nullptr);
+
+                const size_t dst_offset = calculate_offset<3>({ nb1, nb2, nb3 }, { dst_row, i02, i03 });
+                stream->memcpy((char *) dst->data + dst_offset, dst_row_q.data(), nblocks * sizeof(blockType));
+                stream->wait();
+            }
+        }
+    }
+}
+
 template<typename TIn, typename TIdx, typename TOut>
 static void k_set_rows(
        const char * __restrict__ src0, const TIdx * __restrict__ src1, char * __restrict__ dst,
@@ -200,31 +356,194 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
            break;
 #endif
        case GGML_TYPE_Q8_0:
-            set_rows_sycl_q<TIdx, block_q8_0, QK8_0, cpy_blck_f32_q8_0>(src0_d, src1_d, (block_q8_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_q8_0, QK8_0, cpy_blck_f32_q8_0>(
+                src0_d, src1_d, (block_q8_0 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q1_0:
-            set_rows_sycl_q<TIdx, block_q1_0, QK1_0, cpy_blck_f32_q1_0>(src0_d, src1_d, (block_q1_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_q1_0, QK1_0, cpy_blck_f32_q1_0>(
+                src0_d, src1_d, (block_q1_0 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q2_0:
+            set_rows_sycl_q<TIn, TIdx, block_q2_0, QK2_0, cpy_blck_f32_q2_0>(
+                src0_d, src1_d, (block_q2_0 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q5_1:
-            set_rows_sycl_q<TIdx, block_q5_1, QK5_1, cpy_blck_f32_q5_1>(src0_d, src1_d, (block_q5_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_q5_1, QK5_1, cpy_blck_f32_q5_1>(
+                src0_d, src1_d, (block_q5_1 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q5_0:
-            set_rows_sycl_q<TIdx, block_q5_0, QK5_0, cpy_blck_f32_q5_0>(src0_d, src1_d, (block_q5_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_q5_0, QK5_0, cpy_blck_f32_q5_0>(
+                src0_d, src1_d, (block_q5_0 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q4_1:
-            set_rows_sycl_q<TIdx, block_q4_1, QK4_1, cpy_blck_f32_q4_1>(src0_d, src1_d, (block_q4_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_q4_1, QK4_1, cpy_blck_f32_q4_1>(
+                src0_d, src1_d, (block_q4_1 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q4_0:
-            set_rows_sycl_q<TIdx, block_q4_0, QK4_0, cpy_blck_f32_q4_0>(src0_d, src1_d, (block_q4_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_q4_0, QK4_0, cpy_blck_f32_q4_0>(
+                src0_d, src1_d, (block_q4_0 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_IQ4_NL:
-            set_rows_sycl_q<TIdx, block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>(src0_d, src1_d, (block_iq4_nl *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>(
+                src0_d, src1_d, (block_iq4_nl *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_MXFP4:
-            set_rows_sycl_q<TIdx, block_mxfp4, QK_MXFP4, cpy_blck_f32_mxfp4>(src0_d, src1_d, (block_mxfp4 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_mxfp4, QK_MXFP4, cpy_blck_f32_mxfp4>(
+                src0_d, src1_d, (block_mxfp4 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_NVFP4:
-            set_rows_sycl_q<TIdx, block_nvfp4, QK_NVFP4, cpy_blck_f32_nvfp4>(src0_d, src1_d, (block_nvfp4 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIn, TIdx, block_nvfp4, QK_NVFP4, cpy_blck_f32_nvfp4>(
+                src0_d, src1_d, (block_nvfp4 *) dst->data, ne00, ne01, ne02, ne03,
+                ne10, ne11, ne12, ne13, nb00, nb01,
+                nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            set_rows_sycl_qk_host<TIn, TIdx, block_q2_K, QK_K, quantize_row_q2_K_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            set_rows_sycl_qk_host<TIn, TIdx, block_q3_K, QK_K, quantize_row_q3_K_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            set_rows_sycl_qk_host<TIn, TIdx, block_q4_K, QK_K, quantize_row_q4_K_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            set_rows_sycl_qk_host<TIn, TIdx, block_q5_K, QK_K, quantize_row_q5_K_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            set_rows_sycl_qk_host<TIn, TIdx, block_q6_K, QK_K, quantize_row_q6_K_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            set_rows_sycl_iq_host<TIn, TIdx, block_iq2_xxs, QK_K, quantize_iq2_xxs>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            set_rows_sycl_iq_host<TIn, TIdx, block_iq2_xs, QK_K, quantize_iq2_xs>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            set_rows_sycl_iq_host<TIn, TIdx, block_iq2_s, QK_K, quantize_iq2_s>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            set_rows_sycl_qk_host<TIn, TIdx, block_iq3_xxs, QK_K, quantize_row_iq3_xxs_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            set_rows_sycl_qk_host<TIn, TIdx, block_iq3_s, QK_K, quantize_row_iq3_s_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            set_rows_sycl_iq_host<TIn, TIdx, block_iq1_s, QK_K, quantize_iq1_s>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            set_rows_sycl_iq_host<TIn, TIdx, block_iq1_m, QK_K, quantize_iq1_m>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            set_rows_sycl_qk_host<TIn, TIdx, block_iq4_xs, QK_K, quantize_row_iq4_xs_ref>(
+                src0, src1, dst,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                stream);
            break;
        default:
            GGML_ABORT("Unsupported tensor type!");
@@ -237,12 +556,21 @@ void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];

-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I64 || dst->src[1]->type == GGML_TYPE_I32);

-    if (src1->type == GGML_TYPE_I64) {
-        set_rows_sycl<float, int64_t>(ctx, src0, src1, dst);
+    // dispatch on the index type (src1) and the source value type (src0)
+    if (src0->type == GGML_TYPE_F16) {
+        if (src1->type == GGML_TYPE_I64) {
+            set_rows_sycl<sycl::half, int64_t>(ctx, src0, src1, dst);
+        } else {
+            set_rows_sycl<sycl::half, int32_t>(ctx, src0, src1, dst);
+        }
    } else {
-        set_rows_sycl<float, int32_t>(ctx, src0, src1, dst);
+        if (src1->type == GGML_TYPE_I64) {
+            set_rows_sycl<float, int64_t>(ctx, src0, src1, dst);
+        } else {
+            set_rows_sycl<float, int32_t>(ctx, src0, src1, dst);
+        }
    }
 }
@@ -36,9 +36,13 @@ static void kernel_ssm_conv(
                    return;
                }

-                const int channel = static_cast<int>(idx % d_inner);
-                const int token   = static_cast<int>((idx / d_inner) % n_t);
-                const int seq     = static_cast<int>(idx / (static_cast<size_t>(d_inner) * static_cast<size_t>(n_t)));
+                // src has the tokens of one channel contiguous, dst has the channels of one
+                // token contiguous, so either the loads or the store must be strided. Indexing
+                // token-fastest coalesces the d_conv loads, which measured faster except for
+                // short, cache-resident rows.
+                const int token   = static_cast<int>(idx % n_t);
+                const int channel = static_cast<int>((idx / n_t) % d_inner);
+                const int seq     = static_cast<int>(idx / (static_cast<size_t>(n_t) * static_cast<size_t>(d_inner)));

                const float *s = src_data
                    + static_cast<size_t>(seq) * static_cast<size_t>(src_stride_seq)
@@ -186,13 +186,22 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }

 #define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256

-#define VK_CHECK(err, msg)                                          \
+#define VK_CHECK(err, msg, dev)                                     \
    do {                                                            \
-        vk::Result err_ = (err);                                    \
+        vk::Result err_;                                            \
+        try {                                                       \
+            err_ = (err);                                           \
+        } catch (vk::DeviceLostError &) {                           \
+            ggml_vk_print_device_lost_info(dev);                    \
+            GGML_LOG_ERROR("ggml_vulkan: %s at %s:%d\n",            \
+                #err, __FILE__, __LINE__);                          \
+            throw;                                                  \
+        }                                                           \
        if (err_ != vk::Result::eSuccess) {                         \
-            fprintf(stderr, "ggml_vulkan: %s error %s at %s:%d\n",  \
+            GGML_LOG_ERROR("ggml_vulkan: %s error %s at %s:%d\n",   \
                #err, to_string(err_).c_str(), __FILE__, __LINE__); \
-            exit(1);                                                \
+            throw vk::SystemError(vk::make_error_code(err_),        \
+                "ggml_vulkan: " msg);                               \
        }                                                           \
    } while (0)

@@ -302,9 +311,13 @@ struct vk_command_pool {
    }
 };

+static void ggml_vk_print_device_fault_info(const vk_device& device);
+static void ggml_vk_print_device_lost_info(const vk_device& device);
+
 // Prevent simultaneous submissions to the same queue.
 struct vk_queue_handle {
    vk::Queue queue;
+    vk_device_ref device;
    virtual void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) = 0;
    virtual void lock()   {}   // no-op by default (internally synchronized case)
    virtual void unlock() {}
@@ -315,7 +328,14 @@ struct vk_queue_handle_synchronized : vk_queue_handle {
    std::mutex mutex;
    void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
        std::lock_guard<std::mutex> guard(mutex);
-        queue.submit(submits, fence);
+        try {
+            queue.submit(submits, fence);
+        } catch (vk::DeviceLostError &) {
+            if (auto dev = device.lock()) {
+                ggml_vk_print_device_lost_info(dev);
+            }
+            throw;
+        }
    }
    void lock()   override { mutex.lock(); }
    void unlock() override { mutex.unlock(); }
@@ -324,7 +344,14 @@ struct vk_queue_handle_synchronized : vk_queue_handle {
 struct vk_queue_handle_unsynchronized : vk_queue_handle {
    void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
        // Driver guarantees internal synchronization via VK_KHR_internally_synchronized_queues
-        queue.submit(submits, fence);
+        try {
+            queue.submit(submits, fence);
+        } catch (vk::DeviceLostError &) {
+            if (auto dev = device.lock()) {
+                ggml_vk_print_device_lost_info(dev);
+            }
+            throw;
+        }
    }
    // lock()/unlock() inherited no-ops
 };
@@ -835,6 +862,15 @@ struct vk_device_struct {

    bool pipeline_executable_properties_support {};

+    bool device_fault {};
+    PFN_vkGetDeviceFaultInfoEXT pfn_vkGetDeviceFaultInfoEXT {};
+
+    bool serialize_submissions {};
+
+    const ggml_cgraph * diag_cgraph {};
+    int diag_prev_start = -1;
+    int diag_prev_end = -1;
+
    size_t idx;

    bool mul_mat_l[GGML_TYPE_COUNT];
@@ -1118,6 +1154,57 @@ void vk_command_pool::destroy(vk::Device& device) {
    cmd_buffers.clear();
 }

+static void ggml_vk_print_device_fault_info(const vk_device& device) {
+    if (!device->device_fault || !device->pfn_vkGetDeviceFaultInfoEXT) {
+        return;
+    }
+
+    VkDeviceFaultCountsEXT fault_counts {};
+    fault_counts.sType = VK_STRUCTURE_TYPE_DEVICE_FAULT_COUNTS_EXT;
+    VkResult res = device->pfn_vkGetDeviceFaultInfoEXT(device->device, &fault_counts, nullptr);
+    if (res != VK_SUCCESS) {
+        GGML_LOG_ERROR("ggml_vulkan: vkGetDeviceFaultInfoEXT (counts) failed: %d\n", res);
+        return;
+    }
+
+    std::vector<VkDeviceFaultAddressInfoEXT> address_infos(fault_counts.addressInfoCount);
+    std::vector<VkDeviceFaultVendorInfoEXT> vendor_infos(fault_counts.vendorInfoCount);
+
+    VkDeviceFaultInfoEXT fault_info {};
+    fault_info.sType = VK_STRUCTURE_TYPE_DEVICE_FAULT_INFO_EXT;
+    fault_info.pAddressInfos = address_infos.data();
+    fault_info.pVendorInfos = vendor_infos.data();
+
+    res = device->pfn_vkGetDeviceFaultInfoEXT(device->device, &fault_counts, &fault_info);
+    if (res != VK_SUCCESS) {
+        GGML_LOG_ERROR("ggml_vulkan: vkGetDeviceFaultInfoEXT (info) failed: %d\n", res);
+        return;
+    }
+
+    if (fault_counts.addressInfoCount == 0 && fault_counts.vendorInfoCount == 0 && fault_info.description[0] == '\0') {
+        return;
+    }
+
+    if (fault_info.description[0] != '\0') {
+        GGML_LOG_ERROR("ggml_vulkan: device fault on %s: %s\n", device->name.c_str(), fault_info.description);
+    }
+
+    for (uint32_t i = 0; i < fault_counts.addressInfoCount; i++) {
+        const auto& info = address_infos[i];
+        GGML_LOG_CONT("  address fault %u: type=%d address=0x%llx precision=0x%llx\n",
+                i, (int)info.addressType,
+                (unsigned long long)info.reportedAddress,
+                (unsigned long long)info.addressPrecision);
+    }
+    for (uint32_t i = 0; i < fault_counts.vendorInfoCount; i++) {
+        const auto& info = vendor_infos[i];
+        GGML_LOG_CONT("  vendor fault %u: %s (code=0x%llx data=0x%llx)\n",
+                i, info.description,
+                (unsigned long long)info.vendorFaultCode,
+                (unsigned long long)info.vendorFaultData);
+    }
+}
+
 struct vk_buffer_struct {
    vk::Buffer buffer = VK_NULL_HANDLE;
    vk::DeviceMemory device_memory = VK_NULL_HANDLE;
@@ -2059,6 +2146,36 @@ static uint64_t ggml_vk_get_node_flops(const ggml_tensor * node) {
    return 0;
 }

+static void ggml_vk_print_node_list(const ggml_cgraph * cgraph, int start, int end) {
+    uint64_t total_flops = 0;
+    int n_ops = 0;
+    for (int j = start; j <= end && j < cgraph->n_nodes; j++) {
+        uint64_t flops = ggml_vk_get_node_flops(cgraph->nodes[j]);
+        total_flops += flops;
+        n_ops++;
+        if (flops > 0) {
+            GGML_LOG_CONT("  node %d: %s (%s) [%.2f GFLOP]\n",
+                    j, cgraph->nodes[j]->name, ggml_op_name(cgraph->nodes[j]->op),
+                    flops / 1e9);
+        } else {
+            GGML_LOG_CONT("  node %d: %s (%s)\n",
+                    j, cgraph->nodes[j]->name, ggml_op_name(cgraph->nodes[j]->op));
+        }
+    }
+    GGML_LOG_CONT("  total: %d ops, %.2f GFLOP\n", n_ops, total_flops / 1e9);
+}
+
+static void ggml_vk_print_device_lost_info(const vk_device& device) {
+    ggml_vk_print_device_fault_info(device);
+    if (device->serialize_submissions && device->diag_cgraph != nullptr && device->diag_prev_start >= 0) {
+        GGML_LOG_ERROR("ggml_vulkan: device lost on %s, likely caused by previous submission (nodes %d to %d):\n",
+                device->name.c_str(), device->diag_prev_start, device->diag_prev_end);
+        ggml_vk_print_node_list(device->diag_cgraph, device->diag_prev_start, device->diag_prev_end);
+    } else {
+        GGML_LOG_ERROR("ggml_vulkan: device lost on %s\n", device->name.c_str());
+    }
+}
+
 class vk_perf_logger {
  public:
    void print_timings(bool force = false) {
@@ -2471,17 +2588,27 @@ static void ggml_vk_wait_for_fence(ggml_backend_vk_context * ctx) {
    // Use waitForFences while most of the graph executes. Hopefully the CPU can sleep
    // during this wait.
    if (ctx->almost_ready_fence_pending) {
-        VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence");
+        VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence", ctx->device);
        ctx->device->device.resetFences({ ctx->almost_ready_fence });
        ctx->almost_ready_fence_pending = false;
    }

    // Spin (w/pause) waiting for the graph to finish executing.
    vk::Result result;
-    while ((result = ctx->device->device.getFenceStatus(ctx->fence)) != vk::Result::eSuccess) {
+    for (;;) {
+        try {
+            result = ctx->device->device.getFenceStatus(ctx->fence);
+        } catch (vk::DeviceLostError &) {
+            ggml_vk_print_device_lost_info(ctx->device);
+            GGML_LOG_ERROR("ggml_vulkan: getFenceStatus at %s:%d\n", __FILE__, __LINE__);
+            throw;
+        }
+        if (result == vk::Result::eSuccess) {
+            break;
+        }
        if (result != vk::Result::eNotReady) {
-            fprintf(stderr, "ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__);
-            exit(1);
+            GGML_LOG_ERROR("ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__);
+            throw vk::SystemError(vk::make_error_code(result), "ggml_vulkan: getFenceStatus");
        }
        for (uint32_t i = 0; i < 100; ++i) {
            YIELD();
@@ -3172,6 +3299,7 @@ static std::unique_ptr<vk_queue> ggml_vk_create_queue(vk_device& device, uint32_
    }

    h->queue = device->device.getQueue2(queue_info2);
+    h->device = device;
    q->handle = h;

    q->cmd_pool.init(device, q.get());
@@ -6117,6 +6245,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
 #endif
            } else if (strcmp(VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME, properties.extensionName) == 0) {
                internally_sync_support = true;
+            } else if (strcmp("VK_EXT_device_fault", properties.extensionName) == 0) {
+                device->device_fault = true;
            }
        }

@@ -6471,8 +6601,18 @@ static vk_device ggml_vk_get_device(size_t idx) {
        }
 #endif

+        VkPhysicalDeviceFaultFeaturesEXT fault_features {};
+        fault_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FAULT_FEATURES_EXT;
+        if (device->device_fault) {
+            last_struct->pNext = (VkBaseOutStructure *)&fault_features;
+            last_struct = (VkBaseOutStructure *)&fault_features;
+            device_extensions.push_back("VK_EXT_device_fault");
+        }
+
        vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);

+        device->device_fault = device->device_fault && fault_features.deviceFault;
+
        device->has_internally_synchronized_queues = internally_synchronized_queues_features.internallySynchronizedQueues;

        // Build queue create infos only after querying whether internally synchronized queues are enabled.
@@ -6771,6 +6911,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
        device_create_info.setPNext(&device_features2);
        device->device = device->physical_device.createDevice(device_create_info);

+        if (device->device_fault) {
+            device->pfn_vkGetDeviceFaultInfoEXT = (PFN_vkGetDeviceFaultInfoEXT)
+                vkGetDeviceProcAddr(device->device, "vkGetDeviceFaultInfoEXT");
+        }
+
        // Queues
        device->compute_queue = ggml_vk_create_queue(device, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);

@@ -6893,6 +7038,8 @@ static vk_device ggml_vk_get_device(size_t idx) {

        device->idx = idx;

+        device->serialize_submissions = getenv("GGML_VK_SERIALIZE_SUBMISSIONS") != nullptr;
+
        device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;

        device->add_rms_fusion = !device->disable_fusion &&
@@ -8319,7 +8466,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
        }

        ggml_vk_submit(subctx, dst->device->fence);
-        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
+        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences", dst->device);
        dst->device->device.resetFences({ dst->device->fence });
        ggml_vk_queue_command_pools_cleanup(dst->device);
    }
@@ -8431,7 +8578,7 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
        ggml_vk_ctx_end(subctx);
        ggml_vk_submit(subctx, src->device->fence);
        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX),
-                 "vk_buffer_read_2d uma waitForFences");
+                 "vk_buffer_read_2d uma waitForFences", src->device);
        src->device->device.resetFences({ src->device->fence });
        ggml_vk_queue_command_pools_cleanup(src->device);

@@ -8452,7 +8599,7 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
        ggml_vk_ctx_end(subctx);

        ggml_vk_submit(subctx, src->device->fence);
-        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read_2d waitForFences");
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read_2d waitForFences", src->device);
        src->device->device.resetFences({ src->device->fence });
        ggml_vk_queue_command_pools_cleanup(src->device);

@@ -8487,7 +8634,7 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
        ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
        ggml_vk_ctx_end(subctx);
        ggml_vk_submit(subctx, src->device->fence);
-        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences", src->device);
        src->device->device.resetFences({ src->device->fence });
        ggml_vk_queue_command_pools_cleanup(src->device);
    } else {
@@ -8531,7 +8678,7 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
    ggml_vk_ctx_end(subctx);

    ggml_vk_submit(subctx, dst->device->fence);
-    VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences");
+    VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences", dst->device);
    dst->device->device.resetFences({ dst->device->fence });
    ggml_vk_queue_command_pools_cleanup(dst->device);
 }
@@ -14266,7 +14413,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t

    auto begin = std::chrono::high_resolution_clock::now();
    ggml_vk_submit(subctx, ctx->fence);
-    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences");
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences", ctx->device);
    ctx->device->device.resetFences({ ctx->fence });
    ggml_vk_queue_command_pools_cleanup(ctx->device);

@@ -14468,7 +14615,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
    auto begin = std::chrono::high_resolution_clock::now();

    ggml_vk_submit(subctx, ctx->fence);
-    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences", ctx->device);
    ctx->device->device.resetFences({ ctx->fence });
    ggml_vk_queue_command_pools_cleanup(ctx->device);

@@ -14754,7 +14901,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
    auto begin = std::chrono::high_resolution_clock::now();

    ggml_vk_submit(subctx, ctx->fence);
-    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences", ctx->device);
    ctx->device->device.resetFences({ ctx->fence });
    ggml_vk_queue_command_pools_cleanup(ctx->device);

@@ -15553,7 +15700,9 @@ static void ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
            memset(mset.dst, mset.val, mset.n);
        }

-        if (almost_ready && !ctx->almost_ready_fence_pending) {
+        if (ctx->device->serialize_submissions) {
+            ggml_vk_submit(subctx, ctx->fence);
+        } else if (almost_ready && !ctx->almost_ready_fence_pending) {
            ggml_vk_submit(subctx, ctx->almost_ready_fence);
            ctx->almost_ready_fence_pending = true;
        } else {
@@ -16164,12 +16313,20 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
            memcpy(cpy.dst, cpy.src, cpy.n);
        }

-        ggml_vk_submit(compute_ctx, {});
+        if (ctx->device->serialize_submissions) {
+            ggml_vk_submit(compute_ctx, ctx->fence);
+            VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "synchronize waitForFences", ctx->device);
+            ctx->device->device.resetFences({ ctx->fence });
+        } else {
+            ggml_vk_submit(compute_ctx, {});
+        }
        ctx->submit_pending = true;
    }

    if (ctx->submit_pending) {
-        if (ctx->device->async_use_transfer_queue && ctx->transfer_semaphore_last_submitted < ctx->transfer_semaphore.value) {
+        if (ctx->device->serialize_submissions) {
+            ctx->submit_pending = false;
+        } else if (ctx->device->async_use_transfer_queue && ctx->transfer_semaphore_last_submitted < ctx->transfer_semaphore.value) {
            vk::TimelineSemaphoreSubmitInfo tl_info{
                1, &ctx->transfer_semaphore.value,
                0, nullptr,
@@ -16186,7 +16343,9 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
        } else {
            ctx->device->compute_queue->handle->submit({}, ctx->fence);
        }
-        ggml_vk_wait_for_fence(ctx);
+        if (!ctx->device->serialize_submissions) {
+            ggml_vk_wait_for_fence(ctx);
+        }
        ctx->submit_pending = false;
        if (cmd_buf) {
            cmd_buf->in_use = false;
@@ -16758,6 +16917,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;

+    ctx->device->diag_cgraph = nullptr;
+    ctx->device->diag_prev_start = -1;
+    ctx->device->diag_prev_end = -1;
+
    if (vk_instance.debug_utils_support) {
        vk::DebugUtilsLabelEXT dul = {};
        dul.pLabelName = "ggml_backend_vk_graph_compute";
@@ -16849,6 +17012,36 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    }
    uint64_t flops_per_submit = std::min(flops_cap, ctx->last_total_flops / 40u);

+    auto const submit_after = [&](int start, int end) {
+        if (ctx->device->serialize_submissions) {
+            try {
+                auto res = ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX);
+                if (res != vk::Result::eSuccess) {
+                    GGML_LOG_ERROR("ggml_vulkan: waitForFences error during serialized submission\n");
+                    throw vk::SystemError(vk::make_error_code(res), "ggml_vulkan: waitForFences during serialized submission");
+                }
+            } catch (vk::DeviceLostError &) {
+                ggml_vk_print_device_fault_info(ctx->device);
+                GGML_LOG_ERROR("ggml_vulkan: device lost on %s waiting for submission (nodes %d to %d):\n",
+                        ctx->device->name.c_str(), start, end);
+                ggml_vk_print_node_list(cgraph, start, end);
+                throw;
+            }
+            ctx->device->device.resetFences({ ctx->fence });
+            ctx->submit_pending = false;
+            ctx->device->diag_cgraph = cgraph;
+            ctx->device->diag_prev_start = start;
+            ctx->device->diag_prev_end = end;
+        }
+        first_node_in_batch = true;
+        submitted_nodes = 0;
+        batch_flops = 0;
+        if (submit_count < 3) {
+            flops_per_submit *= 2;
+        }
+        submit_count++;
+    };
+
    for (int i = 0; i < cgraph->n_nodes; i++) {
        if (first_node_in_batch) {
            submit_node_idx = i;
@@ -16856,8 +17049,20 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg

        {
            auto node_flops = ggml_vk_get_node_flops(cgraph->nodes[i]);
-            batch_flops += node_flops;
            total_flops += node_flops;
+
+            // Flush the current batch before recording a node that would push it over the flop threshold
+            if (flops_per_submit != 0 && submitted_nodes > 0 && batch_flops + node_flops >= flops_per_submit) {
+                vk_context flush_ctx = ggml_vk_get_compute_ctx(ctx);
+                ggml_vk_ctx_end(flush_ctx);
+                flush_ctx->exit_tensor_idx = -1;
+                ctx->compute_ctx.reset();
+                ggml_vk_compute_forward(ctx, cgraph, cgraph->nodes[submit_node_idx], submit_node_idx, false);
+                submit_after(submit_node_idx, i - 1);
+                submit_node_idx = i;
+            }
+
+            batch_flops += node_flops;
        }

        // op_srcs_fused_elementwise indicates whether an op's srcs all contribute to
@@ -17111,13 +17316,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
        }

        if (submit && enqueued) {
-            first_node_in_batch = true;
-            submitted_nodes = 0;
-            batch_flops = 0;
-            if (submit_count < 3) {
-                flops_per_submit *= 2;
-            }
-            submit_count++;
+            submit_after(submit_node_idx, i + (int)ctx->num_additional_fused_ops);
        }
        i += ctx->num_additional_fused_ops;
        ctx->num_additional_fused_ops = 0;
@@ -17133,13 +17332,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
        ggml_vk_ctx_end(compute_ctx);

        ggml_vk_submit(compute_ctx, ctx->device->fence);
-        VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences");
+        VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences", ctx->device);
        ctx->device->device.resetFences({ ctx->device->fence });
        ctx->compute_ctx.reset();

        // Get the results and pass them to the logger
        std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
-        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->query_pool, 0, ctx->query_idx, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results");
+        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->query_pool, 0, ctx->query_idx, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results", ctx->device);
        if (!vk_perf_logger_concurrent) {
            // Log each op separately
            for (int i = 1; i < ctx->query_idx; i++) {
@@ -18366,7 +18565,7 @@ static void ggml_backend_vk_device_event_synchronize(ggml_backend_dev_t dev, ggm
        vk::Semaphore sem = vkev->tl_semaphore.s;
        uint64_t val = vkev->tl_semaphore.value;
        vk::SemaphoreWaitInfo swi{vk::SemaphoreWaitFlags{}, sem, val};
-        VK_CHECK(device->device.waitSemaphores(swi, UINT64_MAX), "event_synchronize");
+        VK_CHECK(device->device.waitSemaphores(swi, UINT64_MAX), "event_synchronize", device);

        // Reset and move submitted events
        for (auto& event : vkev->events_submitted) {
@@ -2815,11 +2815,25 @@ class ggml_webgpu_shader_lib {
        key.common.v_direct &= decisions.use_sg_matrix && key.common.v_type == GGML_TYPE_F16;
        key.use_sg_matrix = decisions.use_sg_matrix;

-        const uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
+        uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
            context.wg_mem_limit_bytes, decisions.q_tile, decisions.use_sg_matrix ? context.sg_mat_n : 1u,
            key.common.head_dim_qk, key.common.head_dim_v, key.common.has_mask,
            key.common.k_direct || key.common.v_direct);
-        GGML_ASSERT(max_kv_tile > 0);
+
+        // WorkGroup storage size isn't enough for some params with subgroup matrices path (ref. https://github.com/ggml-org/llama.cpp/pull/26566)
+        if (max_kv_tile == 0) {
+            GGML_ASSERT(decisions.use_sg_matrix);
+            // switch to flash_attn_reg_tile path
+            decisions.use_sg_matrix = false;
+            decisions.q_tile        = GGML_WEBGPU_FLASH_ATTN_TILE_Q_TILE;
+            key.common.k_direct     = false;
+            key.common.v_direct     = false;
+            key.use_sg_matrix       = false;
+            max_kv_tile             = ggml_webgpu_flash_attn_max_kv_tile(
+                context.wg_mem_limit_bytes, decisions.q_tile, 1u, key.common.head_dim_qk, key.common.head_dim_v,
+                key.common.has_mask, key.common.k_direct || key.common.v_direct);
+            GGML_ASSERT(max_kv_tile > 0);
+        }

        decisions.kv_tile = decisions.use_sg_matrix ?
                                std::min(max_kv_tile, context.sg_mat_n * GGML_WEBGPU_FLASH_ATTN_PREFERRED_KV_SG_TILES) :
@@ -2993,6 +3007,10 @@ class ggml_webgpu_shader_lib {
                defines.push_back("SRC_F16");
                variant += "_f16";
                break;
+            case GGML_TYPE_I32:
+                defines.push_back("SRC_I32");
+                variant += "_i32";
+                break;
            default:
                GGML_ABORT("Unsupported src type for cpy shader");
        }
@@ -3221,17 +3239,17 @@ class ggml_webgpu_shader_lib {
        auto push_type_defines = [&](const char * prefix, ggml_type type) {
            std::string s_prefix = prefix;
            if (type == GGML_TYPE_F32) {
-                defines.push_back(s_prefix + "_F32");
+                defines.push_back(s_prefix + "=f32");
            } else if (type == GGML_TYPE_F16) {
-                defines.push_back(s_prefix + "_F16");
+                defines.push_back(s_prefix + "=f16");
            } else {
                GGML_ABORT("Unsupported type for CONV_2D shader");
            }
        };

-        push_type_defines("WEIGHT", key.weight_type);
-        push_type_defines("INPUT", key.input_type);
-        push_type_defines("OUTPUT", key.output_type);
+        push_type_defines("WEIGHT_TYPE", key.weight_type);
+        push_type_defines("INPUT_TYPE", key.input_type);
+        push_type_defines("OUTPUT_TYPE", key.output_type);

        defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));

@@ -3263,17 +3281,18 @@ class ggml_webgpu_shader_lib {
        auto push_type_defines = [&](const char * prefix, ggml_type type) {
            std::string s_prefix = prefix;
            if (type == GGML_TYPE_F32) {
-                defines.push_back(s_prefix + "_F32");
+                defines.push_back(s_prefix + "=f32");
            } else if (type == GGML_TYPE_F16) {
-                defines.push_back(s_prefix + "_F16");
+                defines.push_back(s_prefix + "=f16");
            } else {
-                GGML_ABORT("Unsupported type for CONV_2D_DW shader");
+                GGML_ABORT("Unsupported type for CONV_2D shader");
            }
        };

-        push_type_defines("WEIGHT", key.weight_type);
-        push_type_defines("INPUT", key.input_type);
-        push_type_defines("OUTPUT", key.output_type);
+        push_type_defines("WEIGHT_TYPE", key.weight_type);
+        push_type_defines("INPUT_TYPE", key.input_type);
+        push_type_defines("OUTPUT_TYPE", key.output_type);
+
        if (whcn) {
            defines.push_back("WHCN");
        }
@@ -3304,16 +3323,16 @@ class ggml_webgpu_shader_lib {
        auto push_type_defines = [&](const char * prefix, ggml_type type) {
            std::string s_prefix = prefix;
            if (type == GGML_TYPE_F32) {
-                defines.push_back(s_prefix + "_F32");
+                defines.push_back(s_prefix + "=f32");
            } else if (type == GGML_TYPE_F16) {
-                defines.push_back(s_prefix + "_F16");
+                defines.push_back(s_prefix + "=f16");
            } else {
                GGML_ABORT("Unsupported type for IM2COL shader");
            }
        };

-        push_type_defines("INPUT", key.input_type);
-        push_type_defines("OUTPUT", key.output_type);
+        push_type_defines("INPUT_TYPE", key.input_type);
+        push_type_defines("OUTPUT_TYPE", key.output_type);

        defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));

@@ -930,7 +930,6 @@ static webgpu_encoded_op ggml_webgpu_solve_tri(webgpu_context & ctx,

        (uint32_t) src1->ne[0],
        (uint32_t) dst->ne[2],
-        (uint32_t) dst->ne[3],
    };

    std::vector<wgpu::BindGroupEntry> entries = {
@@ -1039,7 +1038,6 @@ static webgpu_encoded_op ggml_webgpu_conv_2d_dw(webgpu_context & ctx,

        (uint32_t) ggml_nelements(dst),
        (uint32_t) dst->ne[2],
-        (uint32_t) dst->ne[3],
        (uint32_t) dst->ne[0],
        (uint32_t) dst->ne[1],
        (uint32_t) src1->ne[0],
@@ -1328,7 +1326,6 @@ static webgpu_encoded_op ggml_webgpu_ssm_scan(webgpu_context & ctx,
        (uint32_t) src0->ne[2],
        (uint32_t) src4->ne[1],
        (uint32_t) src1->ne[2],
-        (uint32_t) src1->ne[3],
        (uint32_t) ggml_nelements(src1),
    };

@@ -1921,25 +1918,20 @@ static bool ggml_webgpu_flash_attn_use_vec_path(const webgpu_global_context & gl
                                                const ggml_tensor *           K,
                                                const ggml_tensor *           V) {
    const size_t storage_offset_alignment = global_ctx->capabilities.limits.minStorageBufferOffsetAlignment;
-    const bool   k_float_vec4_aligned     = (K->type != GGML_TYPE_F16 && K->type != GGML_TYPE_F32) ||
-                                            ggml_webgpu_flash_attn_float_vec4_aligned(K, storage_offset_alignment);
-    const bool   v_float_vec4_aligned     = (V->type != GGML_TYPE_F16 && V->type != GGML_TYPE_F32) ||
-                                            ggml_webgpu_flash_attn_float_vec4_aligned(V, storage_offset_alignment);
-    const bool   k_vec_type_supported =
-        K->type == GGML_TYPE_F32 || K->type == GGML_TYPE_F16 || K->type == GGML_TYPE_Q4_0 || K->type == GGML_TYPE_Q8_0;
-    const bool v_vec_type_supported =
-        V->type == GGML_TYPE_F32 || V->type == GGML_TYPE_F16 || V->type == GGML_TYPE_Q4_0 || V->type == GGML_TYPE_Q8_0;
-    const uint32_t k_vec_head_align         = (K->type == GGML_TYPE_F32 || K->type == GGML_TYPE_F16) ?
-                                                  GGML_WEBGPU_FLASH_ATTN_TILE_KV_VEC_WIDTH :
-                                                  (uint32_t) ggml_blck_size(K->type);
-    const uint32_t v_vec_head_align         = (V->type == GGML_TYPE_F32 || V->type == GGML_TYPE_F16) ?
-                                                  GGML_WEBGPU_FLASH_ATTN_TILE_KV_VEC_WIDTH :
-                                                  (uint32_t) ggml_blck_size(V->type);
-    const bool     kv_vec_head_dims_aligned = Q->ne[0] % k_vec_head_align == 0 && V->ne[0] % v_vec_head_align == 0;
+
+    const bool k_float_vec4_aligned = (K->type != GGML_TYPE_F16 && K->type != GGML_TYPE_F32) ||
+                                      ggml_webgpu_flash_attn_float_vec4_aligned(K, storage_offset_alignment);
+    const bool v_float_vec4_aligned = (V->type != GGML_TYPE_F16 && V->type != GGML_TYPE_F32) ||
+                                      ggml_webgpu_flash_attn_float_vec4_aligned(V, storage_offset_alignment);
+
+    const uint32_t k_vec_head_align =
+        ggml_is_quantized(K->type) ? ggml_blck_size(K->type) : GGML_WEBGPU_FLASH_ATTN_TILE_KV_VEC_WIDTH;
+    const uint32_t v_vec_head_align =
+        ggml_is_quantized(V->type) ? ggml_blck_size(V->type) : GGML_WEBGPU_FLASH_ATTN_TILE_KV_VEC_WIDTH;
+    const bool kv_vec_head_dims_aligned = Q->ne[0] % k_vec_head_align == 0 && V->ne[0] % v_vec_head_align == 0;

    return global_ctx->capabilities.supports_subgroups && (Q->ne[1] < GGML_WEBGPU_FLASH_ATTN_VEC_MAX_SEQ_LEN) &&
-           kv_vec_head_dims_aligned && k_vec_type_supported && v_vec_type_supported && k_float_vec4_aligned &&
-           v_float_vec4_aligned;
+           kv_vec_head_dims_aligned && k_float_vec4_aligned && v_float_vec4_aligned;
 }

 static ggml_webgpu_flash_attn_op ggml_webgpu_flash_attn_prepare(webgpu_context & ctx,
@@ -2514,7 +2506,6 @@ static webgpu_encoded_op ggml_webgpu_concat(webgpu_context & ctx,
                                     (uint32_t) dst->ne[0],
                                     (uint32_t) dst->ne[1],
                                     (uint32_t) dst->ne[2],
-                                     (uint32_t) dst->ne[3],
                                     dim,
                                     (uint32_t) src0->ne[dim] };

@@ -2610,7 +2601,6 @@ static std::optional<webgpu_encoded_op> ggml_webgpu_rms_norm_mul(webgpu_context
        (uint32_t) dst->ne[0],
        (uint32_t) dst->ne[1],
        (uint32_t) dst->ne[2],
-        (uint32_t) dst->ne[3],
        ggml_webgpu_u32_from_f32(ggml_get_op_params_f32(rn_dst, 0))  // epsilon, treated as f32 in the shader
    };

@@ -2666,7 +2656,6 @@ static webgpu_encoded_op ggml_webgpu_row_norm(webgpu_context & ctx, ggml_tensor
        (uint32_t) src->ne[0],
        (uint32_t) src->ne[1],
        (uint32_t) src->ne[2],
-        (uint32_t) src->ne[3],
        ggml_webgpu_u32_from_f32(ggml_get_op_params_f32(dst, 0))  // epsilon, treated as f32 in the shader
    };

@@ -2925,7 +2914,6 @@ static webgpu_encoded_op ggml_webgpu_soft_max(webgpu_context & ctx,
        (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
        (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
-        (uint32_t) ggml_nelements(dst),
        (uint32_t) src0->ne[0],
        (uint32_t) src0->ne[1],
        (uint32_t) src0->ne[2],
@@ -4295,9 +4283,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
            break;
        case GGML_OP_CPY:
        case GGML_OP_CONT:
-            supports_op = ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
-                           (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) ||
-                          (op->type == GGML_TYPE_I32 && src0->type == GGML_TYPE_F32);
+            supports_op = (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_I32) &&
+                          (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32);
            break;
        case GGML_OP_SET:
            supports_op = src0->type == src1->type && src0->type == op->type &&
@@ -18,7 +18,6 @@ struct Params {
    ne0: u32,
    ne1: u32,
    ne2: u32,
-    ne3: u32,

    dim: u32,
    src0_nedim: u32
@@ -2,25 +2,11 @@
 enable f16;

@group(0) @binding(0)
-#if defined(WEIGHT_F32)
-var<storage, read_write> weights: array<f32>;
-#elif defined(WEIGHT_F16)
-var<storage, read_write> weights: array<f16>;
-#endif
-
+var<storage, read_write> weights: array<WEIGHT_TYPE>;
@group(0) @binding(1)
-#if defined(INPUT_F32)
-var<storage, read_write> input: array<f32>;
-#elif defined(INPUT_F16)
-var<storage, read_write> input: array<f16>;
-#endif
-
+var<storage, read_write> input: array<INPUT_TYPE>;
@group(0) @binding(2)
-#if defined(OUTPUT_F32)
-var<storage, read_write> output: array<f32>;
-#elif defined(OUTPUT_F16)
-var<storage, read_write> output: array<f16>;
-#endif
+var<storage, read_write> output: array<OUTPUT_TYPE>;

 struct Params {
    offset_w: u32,
@@ -50,30 +36,6 @@ struct Params {
@group(0) @binding(3)
 var<uniform> params: Params;

-fn load_weight(idx: u32) -> f32 {
-    #if defined(WEIGHT_F32)
-        return weights[idx];
-    #elif defined(WEIGHT_F16)
-        return f32(weights[idx]);
-    #endif
-}
-
-fn load_input(idx: u32) -> f32 {
-    #if defined(INPUT_F32)
-        return input[idx];
-    #elif defined(INPUT_F16)
-        return f32(input[idx]);
-    #endif
-}
-
-fn store_output(idx: u32, val: f32) {
-    #if defined(OUTPUT_F32)
-        output[idx] = val;
-    #elif defined(OUTPUT_F16)
-        output[idx] = f16(val);
-    #endif
-}
-
 fn ceil_div_u32(x: u32, y: u32) -> u32 {
    return (x + y - 1) / y;
 }
@@ -136,7 +98,7 @@ fn main(
    // entire receptive field is out of bounds
    if (kw_begin >= kw_end || kh_begin >= kh_end) {
        let out_idx = params.offset_o + ow * params.so0 + oh * params.so1 + oc * params.so2 + n * params.so3;
-        store_output(out_idx, 0.0);
+        output[out_idx] = OUTPUT_TYPE(0.0);
        return;
    }

@@ -155,11 +117,11 @@ fn main(
                let iw = u32(ow_base + i32(kw * params.d0));
                let w_idx = w_row_base + kw * params.sw0;
                let in_idx = in_row_base + iw * params.si0;
-                sum += load_weight(w_idx) * load_input(in_idx);
+                sum += f32(weights[w_idx]) * f32(input[in_idx]);
            }
        }
    }

    let out_idx = params.offset_o + ow * params.so0 + oh * params.so1 + oc * params.so2 + n * params.so3;
-    store_output(out_idx, sum);
+    output[out_idx] = OUTPUT_TYPE(sum);
 }
@@ -6,25 +6,11 @@ enable f16;
 // weight (src0) is [KW,KH,1,C]; output matches the input layout.

@group(0) @binding(0)
-#if defined(WEIGHT_F32)
-var<storage, read_write> weights: array<f32>;
-#elif defined(WEIGHT_F16)
-var<storage, read_write> weights: array<f16>;
-#endif
-
+var<storage, read_write> weights: array<WEIGHT_TYPE>;
@group(0) @binding(1)
-#if defined(INPUT_F32)
-var<storage, read_write> input: array<f32>;
-#elif defined(INPUT_F16)
-var<storage, read_write> input: array<f16>;
-#endif
-
+var<storage, read_write> input: array<INPUT_TYPE>;
@group(0) @binding(2)
-#if defined(OUTPUT_F32)
-var<storage, read_write> output: array<f32>;
-#elif defined(OUTPUT_F16)
-var<storage, read_write> output: array<f16>;
-#endif
+var<storage, read_write> output: array<OUTPUT_TYPE>;

 struct Params {
    offset_w: u32,
@@ -33,7 +19,6 @@ struct Params {

    ne: u32,
    channels: u32,
-    batches: u32,
    dst_w: u32, dst_h: u32,
    src_w: u32, src_h: u32,
    knl_w: u32, knl_h: u32,
@@ -46,28 +31,6 @@ struct Params {
@group(0) @binding(3)
 var<uniform> params: Params;

-fn load_weight(idx: u32) -> f32 {
-    #if defined(WEIGHT_F32)
-        return weights[idx];
-    #elif defined(WEIGHT_F16)
-        return f32(weights[idx]);
-    #endif
-}
-fn load_input(idx: u32) -> f32 {
-    #if defined(INPUT_F32)
-        return input[idx];
-    #elif defined(INPUT_F16)
-        return f32(input[idx]);
-    #endif
-}
-fn store_output(idx: u32, val: f32) {
-    #if defined(OUTPUT_F32)
-        output[idx] = val;
-    #elif defined(OUTPUT_F16)
-        output[idx] = f16(val);
-    #endif
-}
-
 #if defined(WHCN)
 // Input/output/kernel contiguous in [W, H, C, N] order (kernel [KW,KH,C]).
 fn conv_2d_dw(idx: u32) -> f32 {
@@ -89,8 +52,8 @@ fn conv_2d_dw(idx: u32) -> f32 {
        for (var kx: u32 = 0u; kx < params.knl_w; kx += 1u) {
            let src_x = i32(dst_x) * params.stride_x + i32(kx) * params.dilation_x - params.pad_x;
            if (src_x < 0 || src_x >= i32(params.src_w)) { continue; }
-            let v = load_input(src_i + u32(src_y) * params.src_w + u32(src_x));
-            let k = load_weight(knl_i + ky * params.knl_w + kx);
+            let v = f32(input[src_i + u32(src_y) * params.src_w + u32(src_x)]);
+            let k = f32(weights[knl_i + ky * params.knl_w + kx]);
            sum += v * k;
        }
    }
@@ -117,8 +80,8 @@ fn conv_2d_dw(idx: u32) -> f32 {
        for (var kx: u32 = 0u; kx < params.knl_w; kx += 1u) {
            let src_x = i32(dst_x) * params.stride_x + i32(kx) * params.dilation_x - params.pad_x;
            if (src_x < 0 || src_x >= i32(params.src_w)) { continue; }
-            let v = load_input(src_i + u32(src_y) * src_row + u32(src_x) * params.channels + c);
-            let k = load_weight(params.offset_w + ky * knl_row + kx * params.channels + c);
+            let v = f32(input[src_i + u32(src_y) * src_row + u32(src_x) * params.channels + c]);
+            let k = f32(weights[params.offset_w + ky * knl_row + kx * params.channels + c]);
            sum += v * k;
        }
    }
@@ -133,5 +96,5 @@ fn main(
 ) {
    let idx = gid.x + (num_wg.x * u32(WG_SIZE)) * gid.y;
    if (idx >= params.ne) { return; }
-    store_output(params.offset_o + idx, conv_2d_dw(idx));
+    output[params.offset_o + idx] = OUTPUT_TYPE(conv_2d_dw(idx));
 }
@@ -4,6 +4,8 @@ enable f16;
 #define SRC_TYPE f32
 #elif defined(SRC_F16)
 #define SRC_TYPE f16
+#elif defined(SRC_I32)
+#define SRC_TYPE i32
 #endif

 #ifdef DST_F32
@@ -7,32 +7,18 @@ enable chromium_experimental_subgroup_matrix;
 #define BYTE_HELPERS
 #include "common_decls.tmpl"

-#ifdef K_F32
-#define K_TYPE f32
-#elif defined(K_Q4_0) || defined(K_Q8_0)
-#define K_TYPE u32
-#else
-#define K_TYPE f16
-#endif
-
-#ifdef V_F32
-#define V_TYPE f32
-#elif defined(V_Q4_0) || defined(V_Q8_0)
-#define V_TYPE u32
-#else
-#define V_TYPE f16
-#endif
+#define FLASH_ATTN_SCALAR_KV
+#include "flash_attn_decls.tmpl"

 // Default values
+// The actual values are defined in shader-lib.
 #define HEAD_DIM_QK 64
 #define HEAD_DIM_V 64
-
 // The number of rows/columns/k in a subgroup matrix. MxK * KxN = MxN
 // Note that the "K" here does not correspond to the K in attention's Q/K/V, it's just the common dimension.
 #define SG_MAT_M 8
 #define SG_MAT_N 8
 #define SG_MAT_K 8
-
 // Each workgroup processes one subgroup matrix of Q rows
 #define Q_TILE SG_MAT_M
 #define KV_TILE 16
@@ -41,104 +27,13 @@ enable chromium_experimental_subgroup_matrix;
 // Number of subgroup-matrix-width blocks that span the KV tile. SG_MAT_N must divide KV_TILE.
 #define KV_BLOCKS (KV_TILE / SG_MAT_N)

-struct Params {
-    offset_q: u32,
-    offset_k: u32,
-    offset_v: u32,
-    offset_mask: u32,
-    offset_sinks: u32,
-    offset_dst: u32,
-
-    // shapes of Q/K/V
-    n_heads: u32,
-    seq_len_q: u32,
-    seq_len_kv: u32,
-
-    // strides (in elements)
-    stride_q1: u32,
-    stride_q2: u32,
-    stride_q3: u32,
-    stride_k1: u32,
-    stride_k2: u32,
-    stride_k3: u32,
-    stride_v1: u32,
-    stride_v2: u32,
-    stride_v3: u32,
-    stride_mask3: u32,
-
-    // repeat factors for K/V, e.g., MHA vs. MQA vs. GQA
-    q_per_kv: u32,
-
-    // softmax params
-    scale: f32,
-    max_bias: f32,
-    logit_softcap: f32,
-    n_head_log2: f32,
-    m0: f32,
-    m1: f32,
-};
-
-@group(0) @binding(0) var<storage, read_write> Q: array<f32>;
-#ifdef KV_OVERLAP
-@group(0) @binding(1) var<storage, read_write> K: array<K_TYPE>;
-#define V K
-#else
-@group(0) @binding(1) var<storage, read_write> K: array<K_TYPE>;
-@group(0) @binding(2) var<storage, read_write> V: array<V_TYPE>;
-#endif
-
-#if defined(MASK) && defined(SINKS)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
-@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#else
-@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
-@group(0) @binding(4) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 5
-#define PARAMS_BINDING 6
-#endif
-#elif defined(MASK)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#else
-@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#endif
-#elif defined(SINKS)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#else
-@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#endif
-#else
-#ifdef KV_OVERLAP
-#define DST_BINDING 2
-#define PARAMS_BINDING 3
-#else
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#endif
-#endif
-
-@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<f32>>;
-@group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;
-
-// Just a very small float value.
-const FLOAT_MIN: f32 = -1.0e9;
-
 // The number of Q rows processed per workgroup
 var<workgroup> q_shmem: array<f16, Q_TILE * HEAD_DIM_QK>;

 #if !defined(K_DIRECT) || !defined(V_DIRECT)
+#define STAGING_SHMEM kv_shmem
+#define STAGING_OUT_TYPE f16
+#include "flash_attn_staging.tmpl"
 const kv_shmem_size = KV_TILE * max(HEAD_DIM_QK, HEAD_DIM_V);
 // we can reuse the same shmem for K and V since we only need one at a time
 var<workgroup> kv_shmem: array<f16, kv_shmem_size>;
@@ -175,50 +70,6 @@ fn calc_softmax_term(kv_idx: u32, q_tile_row: u32, slope: f32) -> f32 {
    return v;
 }

-fn load_f32x4(buf: ptr<storage, array<vec4<f32>>, read_write>, scalar_index: u32) -> vec4<f32> {
-    return (*buf)[scalar_index >> 2u];
-}
-
-fn load_kx4(buf: ptr<storage, array<vec4<K_TYPE>>, read_write>, scalar_index: u32) -> vec4<K_TYPE> {
-    return (*buf)[scalar_index >> 2u];
-}
-
-#if !defined(K_DIRECT) || !defined(V_DIRECT)
-#define QUANT_SHMEM kv_shmem
-#define QUANT_OUT_TYPE f16
-#include "flash_attn_quant_staging.tmpl"
-
-#if !defined(K_DIRECT) && !defined(K_Q4_0) && !defined(K_Q8_0)
-fn load_k_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, k_head_offset: u32) {
-    for (var elem_idx = local_x; elem_idx < KV_TILE * HEAD_DIM_QK; elem_idx += WG_SIZE) {
-        let k_row = elem_idx / HEAD_DIM_QK;
-        let k_col = elem_idx % HEAD_DIM_QK;
-        let global_k_row = kv_tile + k_row;
-        let global_k_row_offset = k_head_offset + global_k_row * params.stride_k1;
-        kv_shmem[elem_idx] = f16(select(
-            0.0,
-            K[global_k_row_offset + k_col],
-            global_k_row < params.seq_len_kv && k_col < HEAD_DIM_QK));
-    }
-}
-#endif
-
-#if !defined(V_DIRECT) && !defined(V_Q4_0) && !defined(V_Q8_0)
-fn load_v_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, v_head_offset: u32) {
-    for (var elem_idx = local_x; elem_idx < KV_TILE * HEAD_DIM_V; elem_idx += WG_SIZE) {
-        let v_row = elem_idx / HEAD_DIM_V;
-        let v_col = elem_idx % HEAD_DIM_V;
-        let global_v_row = kv_tile + v_row;
-        let global_v_row_offset = v_head_offset + global_v_row * params.stride_v1;
-        kv_shmem[elem_idx] = f16(select(
-            0.0,
-            V[global_v_row_offset + v_col],
-            global_v_row < params.seq_len_kv && v_col < HEAD_DIM_V));
-    }
-}
-#endif
-#endif
-
@compute @workgroup_size(WG_SIZE)
 fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
    @builtin(local_invocation_id) local_id: vec3<u32>,
@@ -0,0 +1,134 @@
+#ifdef Q_F32
+#define Q_TYPE f32
+#else
+#define Q_TYPE f16
+#endif
+
+#ifdef K_F32
+#define K_TYPE f32
+#elif defined(K_Q4_0) || defined(K_Q8_0)
+#define K_TYPE u32
+#else
+#define K_TYPE f16
+#endif
+
+#ifdef V_F32
+#define V_TYPE f32
+#elif defined(V_Q4_0) || defined(V_Q8_0)
+#define V_TYPE u32
+#else
+#define V_TYPE f16
+#endif
+
+#ifdef DST_F32
+#define DST_TYPE f32
+#else
+#define DST_TYPE f16
+#endif
+
+#if defined(FLASH_ATTN_SCALAR_KV) || defined(K_Q4_0) || defined(K_Q8_0)
+#define K_STORAGE_TYPE K_TYPE
+#else
+#define K_STORAGE_TYPE vec4<K_TYPE>
+#endif
+
+#if defined(FLASH_ATTN_SCALAR_KV) || defined(V_Q4_0) || defined(V_Q8_0)
+#define V_STORAGE_TYPE V_TYPE
+#else
+#define V_STORAGE_TYPE vec4<V_TYPE>
+#endif
+
+// Just a very small float value.
+const FLOAT_MIN: f32 = -1.0e9;
+
+struct Params {
+    offset_q: u32,
+    offset_k: u32,
+    offset_v: u32,
+    offset_mask: u32,
+    offset_sinks: u32,
+    offset_dst: u32,
+
+    // shapes of Q/K/V
+    n_heads: u32,
+    seq_len_q: u32,
+    seq_len_kv: u32,
+
+    // strides (in elements)
+    stride_q1: u32,
+    stride_q2: u32,
+    stride_q3: u32,
+    stride_k1: u32,
+    stride_k2: u32,
+    stride_k3: u32,
+    stride_v1: u32,
+    stride_v2: u32,
+    stride_v3: u32,
+    stride_mask3: u32,
+
+    // repeat factors for K/V, e.g., MHA vs. MQA vs. GQA
+    q_per_kv: u32,
+
+    // softmax params
+    scale: f32,
+    max_bias: f32,
+    logit_softcap: f32,
+    n_head_log2: f32,
+    m0: f32,
+    m1: f32,
+
+#ifdef FLASH_ATTN_VEC_SPLIT
+#ifdef BLK
+    blk_base: u32,
+    blk_nblk0: u32,
+    blk_nblk1: u32,
+#endif
+
+    tmp_data_base: u32,
+    tmp_stats_base: u32,
+    nwg: u32,
+#endif
+};
+
+@group(0) @binding(0) var<storage, read_write> Q: array<Q_TYPE>;
+@group(0) @binding(1) var<storage, read_write> K: array<K_STORAGE_TYPE>;
+#ifdef KV_OVERLAP
+#define V K
+#define MASK_BINDING 2
+#else
+@group(0) @binding(2) var<storage, read_write> V: array<V_STORAGE_TYPE>;
+#define MASK_BINDING 3
+#endif // KV_OVERLAP
+
+#ifdef MASK
+@group(0) @binding(MASK_BINDING) var<storage, read_write> mask: array<f16>;
+#define SINKS_BINDING (MASK_BINDING + 1)
+#else
+#define SINKS_BINDING MASK_BINDING
+#endif
+
+#ifdef SINKS
+@group(0) @binding(SINKS_BINDING) var<storage, read_write> sinks: array<f32>;
+#define BLK_BINDING (SINKS_BINDING + 1)
+#else
+#define BLK_BINDING SINKS_BINDING
+#endif
+
+#ifdef FLASH_ATTN_VEC_SPLIT
+#ifdef BLK
+@group(0) @binding(BLK_BINDING) var<storage, read_write> blk: array<u32>;
+#define TMP_BINDING (BLK_BINDING + 1)
+#else
+#define TMP_BINDING BLK_BINDING
+#endif
+
+@group(0) @binding(TMP_BINDING) var<storage, read_write> tmp: array<f32>;
+#define DST_BINDING (TMP_BINDING + 1)
+#else
+#define DST_BINDING BLK_BINDING
+#endif // FLASH_ATTN_VEC_SPLIT
+
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<DST_TYPE>>;
+
+#define PARAMS_BINDING (DST_BINDING + 1)
+@group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;
@@ -1,83 +0,0 @@
-#include "quant_inner_loops.tmpl"
-
-#define BLOCK_SIZE 32
-#define BLOCKS_K ((HEAD_DIM_QK + BLOCK_SIZE - 1) / BLOCK_SIZE)
-#define BLOCKS_V ((HEAD_DIM_V + BLOCK_SIZE - 1) / BLOCK_SIZE)
-
-#if defined(K_Q4_0)
-#define K_NQ 16
-#define K_BLOCK_SIZE_BYTES 18u
-#define K_BYTES_PER_THREAD 8u
-#define K_BYTES_PER_INNER_LOOP 4u
-#elif defined(K_Q8_0)
-#define K_NQ 16
-#define K_BLOCK_SIZE_BYTES 34u
-#define K_BYTES_PER_THREAD 16u
-#define K_BYTES_PER_INNER_LOOP 4u
-#endif
-
-#if defined(V_Q4_0)
-#define V_NQ 16
-#define V_BLOCK_SIZE_BYTES 18u
-#define V_BYTES_PER_THREAD 8u
-#define V_BYTES_PER_INNER_LOOP 4u
-#elif defined(V_Q8_0)
-#define V_NQ 16
-#define V_BLOCK_SIZE_BYTES 34u
-#define V_BYTES_PER_THREAD 16u
-#define V_BYTES_PER_INNER_LOOP 4u
-#endif
-
-#if defined(K_Q4_0) || defined(K_Q8_0)
-fn load_k_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, k_head_offset: u32) {
-    for (var elem_idx = local_x * K_NQ; elem_idx < kv_count * HEAD_DIM_QK; elem_idx += WG_SIZE * K_NQ) {
-        let blck_idx = elem_idx / BLOCK_SIZE;
-        let block_offset = (elem_idx % BLOCK_SIZE) / K_NQ;
-        let k_row = blck_idx / BLOCKS_K;
-        let global_k_row = kv_tile + k_row;
-        let block_k = blck_idx % BLOCKS_K;
-        let row_offset = k_row * HEAD_DIM_QK;
-        let global_block_idx = k_head_offset + global_k_row * params.stride_k1 + block_k;
-        let block_byte_base = global_block_idx * K_BLOCK_SIZE_BYTES;
-        let d = f16_from_u16(load_k_u16_at(block_byte_base));
-        let thread_byte_offset = block_offset * K_BYTES_PER_THREAD;
-        let shmem_idx = row_offset + block_k * BLOCK_SIZE + thread_byte_offset;
-        for (var j = 0u; j < K_BYTES_PER_THREAD / K_BYTES_PER_INNER_LOOP; j += 1u) {
-            let q_byte_offset = block_byte_base + 2u + thread_byte_offset + j * K_BYTES_PER_INNER_LOOP;
-            let q_packed = load_k_u32_at(q_byte_offset);
-#if defined(K_Q4_0)
-            dequant_q4_0_packed_to_shmem(q_packed, d, shmem_idx + j * K_BYTES_PER_INNER_LOOP);
-#elif defined(K_Q8_0)
-            dequant_q8_0_packed_to_shmem(q_packed, d, shmem_idx + j * K_BYTES_PER_INNER_LOOP);
-#endif
-        }
-    }
-}
-#endif
-
-#if defined(V_Q4_0) || defined(V_Q8_0)
-fn load_v_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, v_head_offset: u32) {
-    for (var elem_idx = local_x * V_NQ; elem_idx < kv_count * HEAD_DIM_V; elem_idx += WG_SIZE * V_NQ) {
-        let blck_idx = elem_idx / BLOCK_SIZE;
-        let block_offset = (elem_idx % BLOCK_SIZE) / V_NQ;
-        let v_row = blck_idx / BLOCKS_V;
-        let global_v_row = kv_tile + v_row;
-        let block_k = blck_idx % BLOCKS_V;
-        let row_offset = v_row * HEAD_DIM_V;
-        let global_block_idx = v_head_offset + global_v_row * params.stride_v1 + block_k;
-        let block_byte_base = global_block_idx * V_BLOCK_SIZE_BYTES;
-        let d = f16_from_u16(load_v_u16_at(block_byte_base));
-        let thread_byte_offset = block_offset * V_BYTES_PER_THREAD;
-        let shmem_idx = row_offset + block_k * BLOCK_SIZE + thread_byte_offset;
-        for (var j = 0u; j < V_BYTES_PER_THREAD / V_BYTES_PER_INNER_LOOP; j += 1u) {
-            let q_byte_offset = block_byte_base + 2u + thread_byte_offset + j * V_BYTES_PER_INNER_LOOP;
-            let q_packed = load_v_u32_at(q_byte_offset);
-#if defined(V_Q4_0)
-            dequant_q4_0_packed_to_shmem(q_packed, d, shmem_idx + j * V_BYTES_PER_INNER_LOOP);
-#elif defined(V_Q8_0)
-            dequant_q8_0_packed_to_shmem(q_packed, d, shmem_idx + j * V_BYTES_PER_INNER_LOOP);
-#endif
-        }
-    }
-}
-#endif
@@ -0,0 +1,136 @@
+#if defined(K_Q4_0) || defined(K_Q8_0) || defined(V_Q4_0) || defined(V_Q8_0)
+#define QUANT_SHMEM STAGING_SHMEM
+#define QUANT_OUT_TYPE STAGING_OUT_TYPE
+#include "quant_inner_loops.tmpl"
+#undef QUANT_SHMEM
+#undef QUANT_OUT_TYPE
+#define BLOCK_SIZE 32
+#define BLOCKS_K ((HEAD_DIM_QK + BLOCK_SIZE - 1) / BLOCK_SIZE)
+#define BLOCKS_V ((HEAD_DIM_V + BLOCK_SIZE - 1) / BLOCK_SIZE)
+#endif
+
+#if defined(K_Q4_0)
+#define K_NQ 16
+#define K_BLOCK_SIZE_BYTES 18u
+#define K_BYTES_PER_THREAD 8u
+#define K_BYTES_PER_INNER_LOOP 4u
+#define DEQUANT_K_PACKED_TO_SHMEM dequant_q4_0_packed_to_shmem
+#elif defined(K_Q8_0)
+#define K_NQ 16
+#define K_BLOCK_SIZE_BYTES 34u
+#define K_BYTES_PER_THREAD 16u
+#define K_BYTES_PER_INNER_LOOP 4u
+#define DEQUANT_K_PACKED_TO_SHMEM dequant_q8_0_packed_to_shmem
+#endif
+
+#if defined(V_Q4_0)
+#define V_NQ 16
+#define V_BLOCK_SIZE_BYTES 18u
+#define V_BYTES_PER_THREAD 8u
+#define V_BYTES_PER_INNER_LOOP 4u
+#define DEQUANT_V_PACKED_TO_SHMEM dequant_q4_0_packed_to_shmem
+#elif defined(V_Q8_0)
+#define V_NQ 16
+#define V_BLOCK_SIZE_BYTES 34u
+#define V_BYTES_PER_THREAD 16u
+#define V_BYTES_PER_INNER_LOOP 4u
+#define DEQUANT_V_PACKED_TO_SHMEM dequant_q8_0_packed_to_shmem
+#endif
+
+#ifndef K_DIRECT
+fn load_k_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, k_head_offset: u32) {
+#if defined(K_Q4_0) || defined(K_Q8_0)
+    for (var elem_idx = local_x * K_NQ; elem_idx < kv_count * HEAD_DIM_QK; elem_idx += WG_SIZE * K_NQ) {
+        let blck_idx = elem_idx / BLOCK_SIZE;
+        let block_offset = (elem_idx % BLOCK_SIZE) / K_NQ;
+        let k_row = blck_idx / BLOCKS_K;
+        let global_k_row = kv_tile + k_row;
+        let block_k = blck_idx % BLOCKS_K;
+        let row_offset = k_row * HEAD_DIM_QK;
+        let global_block_idx = k_head_offset + global_k_row * params.stride_k1 + block_k;
+        let block_byte_base = global_block_idx * K_BLOCK_SIZE_BYTES;
+        let d = f16_from_u16(load_k_u16_at(block_byte_base));
+        let thread_byte_offset = block_offset * K_BYTES_PER_THREAD;
+        let shmem_idx = row_offset + block_k * BLOCK_SIZE + thread_byte_offset;
+        for (var j = 0u; j < K_BYTES_PER_THREAD / K_BYTES_PER_INNER_LOOP; j += 1u) {
+            let q_byte_offset = block_byte_base + 2u + thread_byte_offset + j * K_BYTES_PER_INNER_LOOP;
+            let q_packed = load_k_u32_at(q_byte_offset);
+            DEQUANT_K_PACKED_TO_SHMEM(q_packed, d, shmem_idx + j * K_BYTES_PER_INNER_LOOP);
+        }
+    }
+#elif defined(FLASH_ATTN_SCALAR_KV)
+    for (var elem_idx = local_x; elem_idx < KV_TILE * HEAD_DIM_QK; elem_idx += WG_SIZE) {
+        let k_row = elem_idx / HEAD_DIM_QK;
+        let k_col = elem_idx % HEAD_DIM_QK;
+        let global_k_row = kv_tile + k_row;
+        let global_k_row_offset = k_head_offset + global_k_row * params.stride_k1;
+        STAGING_SHMEM[elem_idx] = STAGING_OUT_TYPE(select(
+            0.0,
+            K[global_k_row_offset + k_col],
+            global_k_row < params.seq_len_kv && k_col < HEAD_DIM_QK));
+    }
+#else
+    for (var vec_idx_local = local_x; vec_idx_local < kv_count * Q_CHUNKS; vec_idx_local += WG_SIZE) {
+        let kv_local = vec_idx_local / Q_CHUNKS;
+        let chunk = vec_idx_local % Q_CHUNKS;
+        let global_k_row = kv_tile + kv_local;
+        let k_vec_index = (k_head_offset + global_k_row * params.stride_k1 + chunk * 4u) >> 2u;
+        let k4 = K[k_vec_index];
+        let kv_off = kv_local * HEAD_DIM_QK + chunk * 4u;
+        STAGING_SHMEM[kv_off + 0u] = STAGING_OUT_TYPE(k4.x);
+        STAGING_SHMEM[kv_off + 1u] = STAGING_OUT_TYPE(k4.y);
+        STAGING_SHMEM[kv_off + 2u] = STAGING_OUT_TYPE(k4.z);
+        STAGING_SHMEM[kv_off + 3u] = STAGING_OUT_TYPE(k4.w);
+    }
+#endif
+}
+#endif // !defined(K_DIRECT)
+
+#ifndef V_DIRECT
+fn load_v_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, v_head_offset: u32) {
+#if defined(V_Q4_0) || defined(V_Q8_0)
+    for (var elem_idx = local_x * V_NQ; elem_idx < kv_count * HEAD_DIM_V; elem_idx += WG_SIZE * V_NQ) {
+        let blck_idx = elem_idx / BLOCK_SIZE;
+        let block_offset = (elem_idx % BLOCK_SIZE) / V_NQ;
+        let v_row = blck_idx / BLOCKS_V;
+        let global_v_row = kv_tile + v_row;
+        let block_k = blck_idx % BLOCKS_V;
+        let row_offset = v_row * HEAD_DIM_V;
+        let global_block_idx = v_head_offset + global_v_row * params.stride_v1 + block_k;
+        let block_byte_base = global_block_idx * V_BLOCK_SIZE_BYTES;
+        let d = f16_from_u16(load_v_u16_at(block_byte_base));
+        let thread_byte_offset = block_offset * V_BYTES_PER_THREAD;
+        let shmem_idx = row_offset + block_k * BLOCK_SIZE + thread_byte_offset;
+        for (var j = 0u; j < V_BYTES_PER_THREAD / V_BYTES_PER_INNER_LOOP; j += 1u) {
+            let q_byte_offset = block_byte_base + 2u + thread_byte_offset + j * V_BYTES_PER_INNER_LOOP;
+            let q_packed = load_v_u32_at(q_byte_offset);
+            DEQUANT_V_PACKED_TO_SHMEM(q_packed, d, shmem_idx + j * V_BYTES_PER_INNER_LOOP);
+        }
+    }
+#elif defined(FLASH_ATTN_SCALAR_KV)
+    for (var elem_idx = local_x; elem_idx < KV_TILE * HEAD_DIM_V; elem_idx += WG_SIZE) {
+        let v_row = elem_idx / HEAD_DIM_V;
+        let v_col = elem_idx % HEAD_DIM_V;
+        let global_v_row = kv_tile + v_row;
+        let global_v_row_offset = v_head_offset + global_v_row * params.stride_v1;
+        STAGING_SHMEM[elem_idx] = STAGING_OUT_TYPE(select(
+            0.0,
+            V[global_v_row_offset + v_col],
+            global_v_row < params.seq_len_kv && v_col < HEAD_DIM_V));
+    }
+#else
+    for (var vec_idx_local = local_x; vec_idx_local < kv_count * V_CHUNKS; vec_idx_local += WG_SIZE) {
+        let kv_local = vec_idx_local / V_CHUNKS;
+        let chunk = vec_idx_local % V_CHUNKS;
+        let global_v_row = kv_tile + kv_local;
+        let v_vec_index = (v_head_offset + global_v_row * params.stride_v1 + chunk * 4u) >> 2u;
+        let v4 = V[v_vec_index];
+        let kv_off = kv_local * HEAD_DIM_V + chunk * 4u;
+        STAGING_SHMEM[kv_off + 0u] = STAGING_OUT_TYPE(v4.x);
+        STAGING_SHMEM[kv_off + 1u] = STAGING_OUT_TYPE(v4.y);
+        STAGING_SHMEM[kv_off + 2u] = STAGING_OUT_TYPE(v4.z);
+        STAGING_SHMEM[kv_off + 3u] = STAGING_OUT_TYPE(v4.w);
+    }
+#endif
+}
+#endif // !defined(V_DIRECT)
@@ -3,192 +3,32 @@ enable subgroups;

 #define BYTE_HELPERS
 #include "common_decls.tmpl"
+#include "flash_attn_decls.tmpl"

-#ifdef Q_F16
-#define Q_TYPE f16
-#else
-#define Q_TYPE f32
-#endif
-
-#ifdef K_F32
-#define K_TYPE f32
-#elif defined(K_Q4_0) || defined(K_Q8_0)
-#define K_TYPE u32
-#else
-#define K_TYPE f16
-#endif
-
-#ifdef V_F32
-#define V_TYPE f32
-#elif defined(V_Q4_0) || defined(V_Q8_0)
-#define V_TYPE u32
-#else
-#define V_TYPE f16
-#endif
-
-#ifdef DST_F16
-#define DST_TYPE f16
-#else
-#define DST_TYPE f32
-#endif
-
+// Default values
+// The actual values are defined in shader-lib.
 #define HEAD_DIM_QK 64
 #define HEAD_DIM_V 64
 #define Q_TILE 4
 #define KV_TILE 64
 #define WG_SIZE 128
-#ifndef MIN_SUBGROUP_SIZE
-#define MIN_SUBGROUP_SIZE MAX_SUBGROUP_SIZE
-#endif

-struct Params {
-    offset_q: u32,
-    offset_k: u32,
-    offset_v: u32,
-    offset_mask: u32,
-    offset_sinks: u32,
-    offset_dst: u32,
-
-    n_heads: u32,
-    seq_len_q: u32,
-    seq_len_kv: u32,
-
-    stride_q1: u32,
-    stride_q2: u32,
-    stride_q3: u32,
-    stride_k1: u32,
-    stride_k2: u32,
-    stride_k3: u32,
-    stride_v1: u32,
-    stride_v2: u32,
-    stride_v3: u32,
-    stride_mask3: u32,
-
-    q_per_kv: u32,
-
-    scale: f32,
-    max_bias: f32,
-    logit_softcap: f32,
-    n_head_log2: f32,
-    m0: f32,
-    m1: f32,
-};
-
-@group(0) @binding(0) var<storage, read_write> Q: array<Q_TYPE>;
-#ifdef KV_OVERLAP
-#if defined(K_Q4_0) || defined(K_Q8_0)
-@group(0) @binding(1) var<storage, read_write> K: array<K_TYPE>;
-#else
-@group(0) @binding(1) var<storage, read_write> K: array<vec4<K_TYPE>>;
-#endif
-#define V K
-#else
-#if defined(K_Q4_0) || defined(K_Q8_0)
-@group(0) @binding(1) var<storage, read_write> K: array<K_TYPE>;
-#else
-@group(0) @binding(1) var<storage, read_write> K: array<vec4<K_TYPE>>;
-#endif
-#if defined(V_Q4_0) || defined(V_Q8_0)
-@group(0) @binding(2) var<storage, read_write> V: array<V_TYPE>;
-#else
-@group(0) @binding(2) var<storage, read_write> V: array<vec4<V_TYPE>>;
-#endif
-#endif
-
-#if defined(MASK) && defined(SINKS)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
-@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#else
-@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
-@group(0) @binding(4) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 5
-#define PARAMS_BINDING 6
-#endif
-#elif defined(MASK)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#else
-@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#endif
-#elif defined(SINKS)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#else
-@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#endif
-#else
-#ifdef KV_OVERLAP
-#define DST_BINDING 2
-#define PARAMS_BINDING 3
-#else
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#endif
-#endif
-
-@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<DST_TYPE>>;
-@group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;
-
-const FLOAT_MIN: f32 = -1.0e9;
 const Q_CHUNKS: u32 = HEAD_DIM_QK / 4u;
 const V_CHUNKS: u32 = HEAD_DIM_V / 4u;
 const SCORE_REGS_PER_LANE: u32 = (KV_TILE + MIN_SUBGROUP_SIZE - 1u) / MIN_SUBGROUP_SIZE;
 const OUT_REGS_PER_LANE: u32 = (V_CHUNKS + MIN_SUBGROUP_SIZE - 1u) / MIN_SUBGROUP_SIZE;
+
+#if !defined(K_DIRECT) || !defined(V_DIRECT)
+#define STAGING_SHMEM kv_shmem
+#define STAGING_OUT_TYPE f16
+#include "flash_attn_staging.tmpl"
 const kv_shmem_size = KV_TILE * max(HEAD_DIM_QK, HEAD_DIM_V);
+var<workgroup> kv_shmem: array<f16, kv_shmem_size>;
+#endif

 var<workgroup> q_shmem: array<Q_TYPE, Q_TILE * HEAD_DIM_QK>;
-var<workgroup> kv_shmem: array<f16, kv_shmem_size>;
 var<workgroup> p_shmem: array<f16, Q_TILE * KV_TILE>;

-#define QUANT_SHMEM kv_shmem
-#define QUANT_OUT_TYPE f16
-#include "flash_attn_quant_staging.tmpl"
-
-#if !defined(K_Q4_0) && !defined(K_Q8_0)
-fn load_k_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, k_head_offset: u32) {
-    for (var vec_idx_local = local_x; vec_idx_local < kv_count * Q_CHUNKS; vec_idx_local += WG_SIZE) {
-        let kv_local = vec_idx_local / Q_CHUNKS;
-        let chunk = vec_idx_local % Q_CHUNKS;
-        let global_k_row = kv_tile + kv_local;
-        let k_vec_index = (k_head_offset + global_k_row * params.stride_k1 + chunk * 4u) >> 2u;
-        let k4 = K[k_vec_index];
-        let kv_off = kv_local * HEAD_DIM_QK + chunk * 4u;
-        kv_shmem[kv_off + 0u] = f16(k4.x);
-        kv_shmem[kv_off + 1u] = f16(k4.y);
-        kv_shmem[kv_off + 2u] = f16(k4.z);
-        kv_shmem[kv_off + 3u] = f16(k4.w);
-    }
-}
-#endif
-
-#if !defined(V_Q4_0) && !defined(V_Q8_0)
-fn load_v_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, v_head_offset: u32) {
-    for (var vec_idx_local = local_x; vec_idx_local < kv_count * V_CHUNKS; vec_idx_local += WG_SIZE) {
-        let kv_local = vec_idx_local / V_CHUNKS;
-        let chunk = vec_idx_local % V_CHUNKS;
-        let global_v_row = kv_tile + kv_local;
-        let v_vec_index = (v_head_offset + global_v_row * params.stride_v1 + chunk * 4u) >> 2u;
-        let v4 = V[v_vec_index];
-        let kv_off = kv_local * HEAD_DIM_V + chunk * 4u;
-        kv_shmem[kv_off + 0u] = f16(v4.x);
-        kv_shmem[kv_off + 1u] = f16(v4.y);
-        kv_shmem[kv_off + 2u] = f16(v4.z);
-        kv_shmem[kv_off + 3u] = f16(v4.w);
-    }
-}
-#endif
-
@compute @workgroup_size(WG_SIZE)
 fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
        @builtin(local_invocation_id) local_id: vec3<u32>,
@@ -4,200 +4,35 @@ enable subgroups;

 #define BYTE_HELPERS
 #include "common_decls.tmpl"
+#define FLASH_ATTN_VEC_SPLIT
+#include "flash_attn_decls.tmpl"

-#ifdef K_F32
-#define K_TYPE f32
-#elif defined(K_Q4_0) || defined(K_Q8_0)
-#define K_TYPE u32
-#else
-#define K_TYPE f16
-#endif
-
-#ifdef V_F32
-#define V_TYPE f32
-#elif defined(V_Q4_0) || defined(V_Q8_0)
-#define V_TYPE u32
-#else
-#define V_TYPE f16
-#endif
-
-#ifdef Q_F16
-#define Q_TYPE f16
-#else
-#define Q_TYPE f32
-#endif
-
-#ifdef DST_F16
-#define DST_TYPE f16
-#else
-#define DST_TYPE f32
-#endif
-
+// Default values
+// The actual values are defined in shader-lib.
 #define HEAD_DIM_QK 64
 #define HEAD_DIM_V 64
-
-#define KV_GRANULARITY 8
 #define KV_TILE 16
 #define WG_SIZE 64

-#define KV_BLOCKS (KV_TILE / KV_GRANULARITY)
-
-struct Params {
-    offset_q: u32,
-    offset_k: u32,
-    offset_v: u32,
-    offset_mask: u32,
-    offset_sinks: u32,
-    offset_dst: u32,
-
-    // shapes of Q/K/V
-    n_heads: u32,
-    seq_len_q: u32,
-    seq_len_kv: u32,
-
-    // strides (in elements)
-    stride_q1: u32,
-    stride_q2: u32,
-    stride_q3: u32,
-    stride_k1: u32,
-    stride_k2: u32,
-    stride_k3: u32,
-    stride_v1: u32,
-    stride_v2: u32,
-    stride_v3: u32,
-    stride_mask3: u32,
-
-    // repeat factors for K/V, e.g., MHA vs. MQA vs. GQA
-    q_per_kv: u32,
-
-    // softmax params
-    scale: f32,
-    max_bias: f32,
-    logit_softcap: f32,
-    n_head_log2: f32,
-    m0: f32,
-    m1: f32,
-
-#ifdef BLK
-    blk_base: u32,
-    blk_nblk0: u32,
-    blk_nblk1: u32,
-#endif
-
-    tmp_data_base: u32,
-    tmp_stats_base: u32,
-    nwg: u32,
-};
-
-@group(0) @binding(0) var<storage, read_write> Q: array<Q_TYPE>;
-#ifdef KV_OVERLAP
-#if defined(K_Q4_0) || defined(K_Q8_0)
-@group(0) @binding(1) var<storage, read_write> K: array<K_TYPE>;
-#else
-@group(0) @binding(1) var<storage, read_write> K: array<vec4<K_TYPE>>;
-#endif
-#define V K
-#else
-#if defined(K_Q4_0) || defined(K_Q8_0)
-@group(0) @binding(1) var<storage, read_write> K: array<K_TYPE>;
-#else
-@group(0) @binding(1) var<storage, read_write> K: array<vec4<K_TYPE>>;
-#endif
-#if defined(V_Q4_0) || defined(V_Q8_0)
-@group(0) @binding(2) var<storage, read_write> V: array<V_TYPE>;
-#else
-@group(0) @binding(2) var<storage, read_write> V: array<vec4<V_TYPE>>;
-#endif
-#endif
-#if defined(MASK) && defined(SINKS)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
-@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
-#ifdef BLK
-#define BLK_BINDING 4
-#define TMP_BINDING 5
-#define DST_BINDING 6
-#define PARAMS_BINDING 7
-#else
-#define TMP_BINDING 4
-#define DST_BINDING 5
-#define PARAMS_BINDING 6
-#endif
-#else
-@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
-@group(0) @binding(4) var<storage, read_write> sinks: array<f32>;
-#ifdef BLK
-#define BLK_BINDING 5
-#define TMP_BINDING 6
-#define DST_BINDING 7
-#define PARAMS_BINDING 8
-#else
-#define TMP_BINDING 5
-#define DST_BINDING 6
-#define PARAMS_BINDING 7
-#endif
-#endif
-#elif defined(MASK)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
-#ifdef BLK
-#define BLK_BINDING 3
-#define TMP_BINDING 4
-#define DST_BINDING 5
-#define PARAMS_BINDING 6
-#else
-#define TMP_BINDING 3
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#endif
-#else
-@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
-#ifdef BLK
-#define BLK_BINDING 4
-#define TMP_BINDING 5
-#define DST_BINDING 6
-#define PARAMS_BINDING 7
-#else
-#define TMP_BINDING 4
-#define DST_BINDING 5
-#define PARAMS_BINDING 6
-#endif
-#endif
-#elif defined(SINKS)
-#ifdef KV_OVERLAP
-@group(0) @binding(2) var<storage, read_write> sinks: array<f32>;
-#define TMP_BINDING 3
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#else
-@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
-#define TMP_BINDING 4
-#define DST_BINDING 5
-#define PARAMS_BINDING 6
-#endif
-#else
-#ifdef KV_OVERLAP
-#define TMP_BINDING 2
-#define DST_BINDING 3
-#define PARAMS_BINDING 4
-#else
-#define TMP_BINDING 3
-#define DST_BINDING 4
-#define PARAMS_BINDING 5
-#endif
-#endif
-
-#ifdef BLK
-@group(0) @binding(BLK_BINDING) var<storage, read_write> blk: array<u32>;
-#endif
-@group(0) @binding(TMP_BINDING) var<storage, read_write> tmp: array<f32>;
-@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<DST_TYPE>>;
-@group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;
-
-// Just a very small float value.
-const FLOAT_MIN: f32 = -1.0e9;
+const Q_CHUNKS: u32 = HEAD_DIM_QK / 4u;
+const V_CHUNKS: u32 = HEAD_DIM_V / 4u;
 const kv_shmem_size = KV_TILE * max(HEAD_DIM_QK, HEAD_DIM_V);

+#if defined(K_DIRECT) || defined(V_DIRECT)
+// Shared memory for scale factor (d) in quantized K/V. Multiple threads use the same value,
+// so caching it is more efficient, even on the direct path.
+var<workgroup> d_shmem: array<f32, kv_shmem_size / 32>;
+#endif
+
+// K/V shared memory handling
+#if !defined(K_DIRECT) || !defined(V_DIRECT)
+#define STAGING_SHMEM kv_shmem
+#define STAGING_OUT_TYPE f32
+#include "flash_attn_staging.tmpl"
+// we can reuse the same shmem for K and V since we only need one at a time
+var<workgroup> kv_shmem: array<f32, kv_shmem_size>;
+#endif
+
 var<workgroup> q_shmem: array<f32, HEAD_DIM_QK>;
 var<workgroup> o_shmem: array<f32, HEAD_DIM_V>;
 // note that we reuse the same storage for both since we only need one at a time
@@ -208,59 +43,6 @@ var<workgroup> inter_shmem: array<f32, KV_TILE>;
 var<workgroup> mask_shmem: array<f32, KV_TILE>;
 #endif

-#if defined(K_DIRECT) || defined(V_DIRECT)
-// Shared memory for scale factor (d) in quantized K/V. Multiple threads use the same value,
-// so caching it is more efficient, even on the direct path.
-var<workgroup> d_shmem: array<f32, kv_shmem_size / 32>;
-#endif
-
-// K/V shared memory handling
-#if !defined(K_DIRECT) || !defined(V_DIRECT)
-
-// we can reuse the same shmem for K and V since we only need one at a time
-var<workgroup> kv_shmem: array<f32, kv_shmem_size>;
-
-#define QUANT_SHMEM kv_shmem
-#define QUANT_OUT_TYPE f32
-#include "flash_attn_quant_staging.tmpl"
-
-#if !defined(K_DIRECT) && !defined(K_Q4_0) && !defined(K_Q8_0)
-fn load_k_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, k_head_offset: u32) {
-    for (var elem_idx = local_x * 4u; elem_idx < KV_TILE * HEAD_DIM_QK; elem_idx += WG_SIZE * 4u) {
-        let k_row = elem_idx / HEAD_DIM_QK;
-        let k_col = elem_idx % HEAD_DIM_QK;
-        let global_k_row = kv_tile + k_row;
-        let global_k_row_offset = k_head_offset + global_k_row * params.stride_k1;
-        let in_bounds = global_k_row < params.seq_len_kv && (k_col + 3u) < HEAD_DIM_QK;
-        let vec_idx = (global_k_row_offset + k_col) >> 2u;
-        let k4 = select(vec4<K_TYPE>(0.0), K[vec_idx], in_bounds);
-        kv_shmem[elem_idx + 0u] = f32(k4.x);
-        kv_shmem[elem_idx + 1u] = f32(k4.y);
-        kv_shmem[elem_idx + 2u] = f32(k4.z);
-        kv_shmem[elem_idx + 3u] = f32(k4.w);
-    }
-}
-#endif
-
-#if !defined(V_DIRECT) && !defined(V_Q4_0) && !defined(V_Q8_0)
-fn load_v_tile_block(local_x: u32, kv_count: u32, kv_tile: u32, v_head_offset: u32) {
-    for (var elem_idx = local_x * 4u; elem_idx < KV_TILE * HEAD_DIM_V; elem_idx += WG_SIZE * 4u) {
-        let v_row = elem_idx / HEAD_DIM_V;
-        let v_col = elem_idx % HEAD_DIM_V;
-        let global_v_row = kv_tile + v_row;
-        let global_v_row_offset = v_head_offset + global_v_row * params.stride_v1;
-        let in_bounds = global_v_row < params.seq_len_kv && (v_col + 3u) < HEAD_DIM_V;
-        let vec_idx = (global_v_row_offset + v_col) >> 2u;
-        let v4 = select(vec4<V_TYPE>(0.0), V[vec_idx], in_bounds);
-        kv_shmem[elem_idx + 0u] = f32(v4.x);
-        kv_shmem[elem_idx + 1u] = f32(v4.y);
-        kv_shmem[elem_idx + 2u] = f32(v4.z);
-        kv_shmem[elem_idx + 3u] = f32(v4.w);
-    }
-}
-#endif
-#endif // !defined(K_DIRECT) || !defined(V_DIRECT)
-
 // Storage for row max and exp sum during online softmax
 fn calc_softmax_term(kv_idx: u32, slope: f32, has_bias: bool, apply_mask: bool) -> f32 {
    var v = select(FLOAT_MIN,
@@ -1,19 +1,9 @@
-#include "common_decls.tmpl"
 enable f16;

@group(0) @binding(0)
-#if defined(INPUT_F32)
-var<storage, read_write> input: array<f32>;
-#elif defined(INPUT_F16)
-var<storage, read_write> input: array<f16>;
-#endif
-
+var<storage, read_write> input: array<INPUT_TYPE>;
@group(0) @binding(1)
-#if defined(OUTPUT_F32)
-var<storage, read_write> output: array<f32>;
-#elif defined(OUTPUT_F16)
-var<storage, read_write> output: array<f16>;
-#endif
+var<storage, read_write> output: array<OUTPUT_TYPE>;

 struct Params {
    offset_i: u32,
@@ -38,22 +28,6 @@ struct Params {
@group(0) @binding(2)
 var<uniform> params: Params;

-fn load_input(idx: u32) -> f32 {
-    #if defined(INPUT_F32)
-        return input[idx];
-    #elif defined(INPUT_F16)
-        return f32(input[idx]);
-    #endif
-}
-
-fn store_output(idx: u32, val: f32) {
-    #if defined(OUTPUT_F32)
-        output[idx] = val;
-    #elif defined(OUTPUT_F16)
-        output[idx] = f16(val);
-    #endif
-}
-
@compute @workgroup_size(WG_SIZE)
 fn main(
    @builtin(global_invocation_id) gid: vec3<u32>,
@@ -90,12 +64,14 @@ fn main(
    let iw_i32 = i32(ow * params.s0 + kw * params.d0) - i32(params.p0);
    let ih_i32 = i32(oh * params.s1 + kh * params.d1) - i32(params.p1);

+    let output_idx = params.offset_o + k * params.so0 + ow * params.so1 + oh * params.so2 + n * params.so3;
+
    if (iw_i32 >= 0 && iw_i32 < i32(params.IW) && ih_i32 >= 0 && ih_i32 < i32(params.IH)) {
        let iw = u32(iw_i32);
        let ih = u32(ih_i32);
        let in_idx = params.offset_i + iw * params.si0 + ih * params.si1 + ic * params.si2 + n * params.si3;
-        store_output(params.offset_o + k * params.so0 + ow * params.so1 + oh * params.so2 + n * params.so3, load_input(in_idx));
+        output[output_idx] = OUTPUT_TYPE(input[in_idx]);
    } else {
-        store_output(params.offset_o + k * params.so0 + ow * params.so1 + oh * params.so2 + n * params.so3, 0.0);
+        output[output_idx] = OUTPUT_TYPE(0.0);
    }
 }
@@ -88,7 +88,6 @@ struct Params {
    ne0: u32,
    ne1: u32,
    ne2: u32,
-    ne3: u32,

    eps: f32
 };
@@ -31,7 +31,6 @@ struct Params {
    ne0: u32,
    ne1: u32,
    ne2: u32,
-    ne3: u32,

    eps: f32
 };
@@ -27,7 +27,6 @@ struct Params {
    stride_dst3: u32,

    // shape of src0/dst
-    ne: u32,
    ne0: u32,
    ne1: u32,
    ne2: u32,
@@ -43,71 +42,38 @@ struct Params {
    m1: f32,
 };

-@group(0) @binding(0)
+#define SRC_BINDING 0
+@group(0) @binding(SRC_BINDING)
 var<storage, read_write> src: array<f32>;

 #ifdef HAS_MASK
-#ifdef HAS_SINK
-@group(0) @binding(1)
+#define MASK_BINDING SRC_BINDING + 1
+@group(0) @binding(MASK_BINDING)
 var<storage, read_write> mask: array<MaskType>;
-@group(0) @binding(2)
-var<storage, read_write> sinks: array<f32>;
-
-#ifdef INPLACE
-@group(0) @binding(3)
-var<uniform> params: Params;
-
 #else
-@group(0) @binding(3)
-var<storage, read_write> dst: array<f32>;
-@group(0) @binding(4)
-var<uniform> params: Params;
+#define MASK_BINDING SRC_BINDING
 #endif

-#else
-@group(0) @binding(1)
-var<storage, read_write> mask: array<MaskType>;
-
-#ifdef INPLACE
-@group(0) @binding(2)
-var<uniform> params: Params;
-
-#else
-@group(0) @binding(2)
-var<storage, read_write> dst: array<f32>;
-@group(0) @binding(3)
-var<uniform> params: Params;
-#endif
-#endif
-
-#else
 #ifdef HAS_SINK
-@group(0) @binding(1)
+#define SINKS_BINDING MASK_BINDING + 1
+@group(0) @binding(SINKS_BINDING)
 var<storage, read_write> sinks: array<f32>;
+#else
+#define SINKS_BINDING MASK_BINDING
+#endif
+
+#define DST_BINDING SINKS_BINDING + 1
+@group(0) @binding(DST_BINDING)
+var<storage, read_write> dst: array<f32>;

 #ifdef INPLACE
-@group(0) @binding(2)
-var<uniform> params: Params;
-
+#define PARAMS_BINDING DST_BINDING
 #else
-@group(0) @binding(2)
-var<storage, read_write> dst: array<f32>;
-@group(0) @binding(3)
-var<uniform> params: Params;
+#define PARAMS_BINDING (DST_BINDING + 1)
 #endif

-#else
-#ifdef INPLACE
-@group(0) @binding(1)
+@group(0) @binding(PARAMS_BINDING)
 var<uniform> params: Params;
-#else
-@group(0) @binding(1)
-var<storage, read_write> dst: array<f32>;
-@group(0) @binding(2)
-var<uniform> params: Params;
-#endif
-#endif
-#endif

 #ifdef INPLACE
 fn inter_value(i: u32) -> f32 {
@@ -242,4 +208,3 @@ fn main(@builtin(workgroup_id) wid: vec3<u32>,
        col += WG_SIZE;
    }
 }
-
@@ -29,7 +29,6 @@ struct Params {

    k: u32,
    ne2: u32,
-    ne3: u32,
 };

@group(0) @binding(3)
@@ -39,7 +39,6 @@ struct Params {
    n_head: u32,
    n_group: u32,
    n_seq_tokens: u32,
-    n_seqs: u32,

    y_elems: u32,
 };
@@ -7200,6 +7200,10 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
    ggml_build_forward_impl(cgraph, tensor, true, true);
 }

+void ggml_build_forward_order(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
+    ggml_build_forward_impl(cgraph, tensor, true, false);
+}
+
 void ggml_build_backward_expand(
        struct ggml_context *  ctx,
        struct ggml_cgraph  *  cgraph,
@@ -164,6 +164,13 @@ class Keys:
        NORM_BEFORE_RESIDUAL              = "{arch}.norm_before_residual"
        NORM_BEFORE_FC                    = "{arch}.norm_before_fc"

+    class Adapters:
+        COUNT                = "{arch}.adapters.count"
+        TOKEN_IDS_ACTIVATE   = "{arch}.adapters.token_ids_activate"
+        TOKEN_IDS_SUBSTITUTE = "{arch}.adapters.token_ids_substitute"
+        LORA_RANK            = "{arch}.adapters.lora_rank"
+        ROUTER_GAIN          = "{arch}.adapters.router_gain"
+
    class Attention:
        HEAD_COUNT                   = "{arch}.attention.head_count"
        HEAD_COUNT_KV                = "{arch}.attention.head_count_kv"
@@ -502,6 +509,7 @@ class MODEL_ARCH(IntEnum):
    OLMO             = auto()
    OLMO2            = auto()
    OLMOE            = auto()
+    MUSE_GLIMMER     = auto()
    OPENELM          = auto()
    ARCTIC           = auto()
    DEEPSEEK         = auto()
@@ -527,6 +535,7 @@ class MODEL_ARCH(IntEnum):
    GRANITE          = auto()
    GRANITE_MOE      = auto()
    GRANITE_HYBRID   = auto()
+    GRANITE_SWITCH   = auto()
    CHAMELEON        = auto()
    WAVTOKENIZER_DEC = auto()
    PLM              = auto()
@@ -1173,6 +1182,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.OLMO:             "olmo",
    MODEL_ARCH.OLMO2:            "olmo2",
    MODEL_ARCH.OLMOE:            "olmoe",
+    MODEL_ARCH.MUSE_GLIMMER:     "muse-glimmer",
    MODEL_ARCH.OPENELM:          "openelm",
    MODEL_ARCH.ARCTIC:           "arctic",
    MODEL_ARCH.DEEPSEEK:         "deepseek",
@@ -1198,6 +1208,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.GRANITE:          "granite",
    MODEL_ARCH.GRANITE_MOE:      "granitemoe",
    MODEL_ARCH.GRANITE_HYBRID:   "granitehybrid",
+    MODEL_ARCH.GRANITE_SWITCH:   "graniteswitch",
    MODEL_ARCH.CHAMELEON:        "chameleon",
    MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
    MODEL_ARCH.PLM:              "plm",
@@ -1553,8 +1564,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
    MODEL_TENSOR.V_MM_UP:                   "mm.up",
    MODEL_TENSOR.V_MM_DOWN:                 "mm.down",
    MODEL_TENSOR.V_MM_GATE:                 "mm.gate",
-    MODEL_TENSOR.V_MM_MERGER_FC1:            "mm.merger.fc1",
-    MODEL_TENSOR.V_MM_MERGER_FC2:            "mm.merger.fc2",
+    MODEL_TENSOR.V_MM_MERGER_FC1:           "mm.merger.fc1",
+    MODEL_TENSOR.V_MM_MERGER_FC2:           "mm.merger.fc2",
    MODEL_TENSOR.V_TOK_BOI:                 "v.boi",
    MODEL_TENSOR.V_TOK_EOI:                 "v.eoi",
    MODEL_TENSOR.V_MM_PRE_NORM:             "mm.pre_norm",
@@ -3322,6 +3333,25 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_UP_EXP,
        MODEL_TENSOR.FFN_DOWN_EXP,
    ],
+    MODEL_ARCH.MUSE_GLIMMER: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_PRE_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+    ],
    MODEL_ARCH.OPENELM: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -3837,6 +3867,12 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN_SHEXP,
        MODEL_TENSOR.FFN_UP_SHEXP,
        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        # NextN/MTP (draft head)
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
    ],
    MODEL_ARCH.EXAONE: [
        MODEL_TENSOR.TOKEN_EMBD,
@@ -3972,6 +4008,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.GRANITE_SWITCH: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
    MODEL_ARCH.CHAMELEON: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -5136,6 +5187,7 @@ class VisionProjectorType:
    MIMOVL         = "mimovl"
    MIMO_AUDIO     = "mimo_audio"
    GRANITE4_VISION = "granite4_vision"
+    MUSE_GLIMMER   = "muse-glimmer"


 # Items here are (block size, type size)
@@ -906,6 +906,21 @@ class GGUFWriter:
    def add_embedding_scale(self, value: float) -> None:
        self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)

+    def add_adapter_count(self, count: int) -> None:
+        self.add_uint32(Keys.Adapters.COUNT.format(arch=self.arch), count)
+
+    def add_adapter_token_ids_activate(self, ids: Sequence[int]) -> None:
+        self.add_array(Keys.Adapters.TOKEN_IDS_ACTIVATE.format(arch=self.arch), ids)
+
+    def add_adapter_token_ids_substitute(self, ids: Sequence[int]) -> None:
+        self.add_array(Keys.Adapters.TOKEN_IDS_SUBSTITUTE.format(arch=self.arch), ids)
+
+    def add_adapter_lora_rank(self, rank: int) -> None:
+        self.add_uint32(Keys.Adapters.LORA_RANK.format(arch=self.arch), rank)
+
+    def add_adapter_router_gain(self, gain: float) -> None:
+        self.add_float32(Keys.Adapters.ROUTER_GAIN.format(arch=self.arch), gain)
+
    def add_wkv_head_size(self, size: int) -> None:
        self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)

@@ -59,11 +59,29 @@ def byteswap_q6_k(tensor, block_offs):
    delta.byteswap(inplace=True)


+def byteswap_q1_0(tensor, block_offs):
+    # Each block_q1_0 consists of an f16 delta followed by 16 int8 quantizations.
+
+    # Byte-Swap f16 sized delta field
+    delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16)
+    delta.byteswap(inplace=True)
+
+
+def byteswap_tq2_0(tensor, block_offs):
+    # Each block_tq2_0 consists of 64 int8 values followed by 1 f16 value.
+
+    # Byte-Swap f16 sized field
+    delta = tensor.data[block_offs + 64:block_offs + 66].view(dtype=np.uint16)
+    delta.byteswap(inplace=True)
+
+
 byteswap_tensors = {
+    gguf.GGMLQuantizationType.Q1_0:  byteswap_q1_0,
    gguf.GGMLQuantizationType.Q4_0:  byteswap_q4_0,
    gguf.GGMLQuantizationType.Q8_0:  byteswap_q8_0,
    gguf.GGMLQuantizationType.Q4_K:  byteswap_q4_k,
    gguf.GGMLQuantizationType.Q6_K:  byteswap_q6_k,
+    gguf.GGMLQuantizationType.TQ2_0: byteswap_tq2_0,
    gguf.GGMLQuantizationType.MXFP4: byteswap_noop,
    gguf.GGMLQuantizationType.NVFP4: byteswap_noop,
 }
@@ -382,7 +382,7 @@ class TensorNameMap:
        ),

        MODEL_TENSOR.ATTN_GATE: (
-            "model.layers.{bid}.self_attn.gate_proj", # afmoe
+            "model.layers.{bid}.self_attn.gate_proj", # afmoe muse-glimmer
            "model.layers.{bid}.linear_attn.in_proj_z",  # qwen3.5
            "model.layers.{bid}.self_attn.g_proj",    # step3.5 head-wise attention gate
        ),
@@ -1298,10 +1298,12 @@ class TensorNameMap:
            "encoder.final_layer_norm", # t5
            "layer_norm",               # neobert
            "model.hidden_norm",        # dflash
+            "encoder.output_norm_enc",  # dflash (transformers MuseGlimmerAssistant)
        ),

        MODEL_TENSOR.FC: (
-            "model.fc", # dflash
+            "model.fc",   # dflash
+            "encoder.fc", # dflash (transformers MuseGlimmerAssistant)
        ),

        MODEL_TENSOR.DSPARK_MARKOV_W1: (
@@ -1467,6 +1469,7 @@ class TensorNameMap:
            "vision_tower.patch_embed.patchifier.proj", # dots.ocr
            "vision_model.conv1", # Step3-VL
            "model.vision_embedder.patch_dense", # gemma4 unified
+            "model.vision_tower.patch_embedder.patch_embedding", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_EMBD_NORM: (
@@ -1534,7 +1537,8 @@ class TensorNameMap:
            "siglip2.vision_model.encoder.layers.{bid}.self_attn.q_proj", # youtuvl
            "model.vision_model.transformer.layers.{bid}.self_attn.q_proj", # Deepseek-OCR CLIP, generated
            "vision_model.model.layers.{bid}.self_attn.q_proj.linear", # gemma4
-            "model.qwen2_model.model.model.layers.{bid}.self_attn.q_proj" # Deepseek-OCR-2 qwen2
+            "model.qwen2_model.model.model.layers.{bid}.self_attn.q_proj", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.layers.{bid}.attn.q_proj", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
@@ -1560,7 +1564,8 @@ class TensorNameMap:
            "model.vision_model.transformer.layers.{bid}.self_attn.k_proj", # Deepseek-OCR CLIP, generated
            "siglip2.vision_model.encoder.layers.{bid}.self_attn.k_proj",
            "vision_model.model.layers.{bid}.self_attn.k_proj.linear", # gemma4
-            "model.qwen2_model.model.model.layers.{bid}.self_attn.k_proj" # Deepseek-OCR-2 qwen2
+            "model.qwen2_model.model.model.layers.{bid}.self_attn.k_proj", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.layers.{bid}.attn.k_proj", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
@@ -1586,7 +1591,8 @@ class TensorNameMap:
            "siglip2.vision_model.encoder.layers.{bid}.self_attn.v_proj",
            "model.vision_model.transformer.layers.{bid}.self_attn.v_proj", # Deepseek-OCR CLIP, generated
            "vision_model.model.layers.{bid}.self_attn.v_proj.linear", # gemma4
-            "model.qwen2_model.model.model.layers.{bid}.self_attn.v_proj" # Deepseek-OCR-2 qwen2
+            "model.qwen2_model.model.model.layers.{bid}.self_attn.v_proj", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.layers.{bid}.attn.v_proj", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_INPUT_NORM: (
@@ -1610,6 +1616,7 @@ class TensorNameMap:
            "vision_tower.blocks.{bid}.norm1", # dots.ocr
            "vision_model.transformer.resblocks.{bid}.ln_1", # Step3-VL
            "model.qwen2_model.model.model.layers.{bid}.input_layernorm", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.layers.{bid}.norm1", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_ATTN_O: (
@@ -1635,6 +1642,7 @@ class TensorNameMap:
            "vision_model.model.layers.{bid}.self_attn.o_proj.linear", # gemma4
            "vision_tower.blocks.{bid}.attn.proj", # dots.ocr
            "vision_model.transformer.resblocks.{bid}.attn.out_proj", # Step3-VL
+            "model.vision_tower.layers.{bid}.attn.proj", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_ATTN_SINKS: (
@@ -1663,6 +1671,7 @@ class TensorNameMap:
            "vision_tower.blocks.{bid}.norm2", # dots.ocr
            "vision_model.transformer.resblocks.{bid}.ln_2", # Step3-VL
            "model.qwen2_model.model.model.layers.{bid}.post_attention_layernorm", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.layers.{bid}.norm2", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_FFN_UP: (
@@ -1687,6 +1696,7 @@ class TensorNameMap:
            "vision_model.model.layers.{bid}.mlp.up_proj", # gemma4
            "vision_model.transformer.resblocks.{bid}.mlp.c_fc", # Step3-VL
            "model.qwen2_model.model.model.layers.{bid}.mlp.up_proj", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.layers.{bid}.mlp.fc1", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_FFN_GATE: (
@@ -1719,6 +1729,7 @@ class TensorNameMap:
            "model.qwen2_model.model.model.layers.{bid}.mlp.down_proj" , # Deepseek-OCR-2 qwen2
            "vision_model.model.layers.{bid}.mlp.down_proj", # gemma4
            "vision_model.transformer.resblocks.{bid}.mlp.c_proj", # Step3-VL
+            "model.vision_tower.layers.{bid}.mlp.fc2", # muse-glimmer
        ),

        MODEL_TENSOR.V_ENC_ATTN_POST_NORM: (
@@ -1753,6 +1764,7 @@ class TensorNameMap:
            "model.vision_model.pre_layrnorm", # Deepseek-OCR CLIP
            "vision_tower.patch_embed.patchifier.norm", # dots.ocr
            "vision_model.ln_pre", # Step3-VL
+            "model.vision_tower.ln_pre", # muse-glimmer
        ),

        MODEL_TENSOR.V_POST_NORM: (
@@ -1766,6 +1778,7 @@ class TensorNameMap:
            "visual.post_layernorm", # glm4v
            "siglip2.vision_model.post_layernorm",
            "model.qwen2_model.model.model.norm", # Deepseek-OCR-2 qwen2
+            "model.vision_tower.ln_post", # muse-glimmer
        ),

        MODEL_TENSOR.V_MM_POST_NORM: (
@@ -348,14 +348,15 @@ extern "C" {
    // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
    //       https://github.com/ggml-org/llama.cpp/pull/7544
    struct llama_context_params {
-        uint32_t n_ctx;             // text context, 0 = from model
-        uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
-        uint32_t n_ubatch;          // physical maximum batch size
-        uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
-        uint32_t n_rs_seq;          // number of recurrent-state snapshots per seq for rollback (0 = no rollback) [EXPERIMENTAL]
-        uint32_t n_outputs_max;     // max outputs in a ubatch (0 = n_batch)
-        int32_t  n_threads;         // number of threads to use for generation
-        int32_t  n_threads_batch;   // number of threads to use for batch processing
+        uint32_t n_ctx;                 // text context, 0 = from model
+        uint32_t n_batch;               // logical maximum batch size that can be submitted to llama_decode
+        uint32_t n_ubatch;              // physical maximum batch size
+        uint32_t n_seq_max;             // max number of sequences (i.e. distinct states for recurrent models)
+        uint32_t n_rs_seq;              // number of recurrent-state snapshots per seq for rollback (0 = no rollback) [EXPERIMENTAL]
+        uint32_t n_outputs_max;         // max outputs in a ubatch (0 = n_batch)
+        uint32_t n_outputs_max_per_seq; // max outputs per sequence (0 = n_outputs_max)
+        int32_t  n_threads;             // number of threads to use for generation
+        int32_t  n_threads_batch;       // number of threads to use for batch processing

        enum llama_context_type      ctx_type;          // set the context type (e.g. MTP)
        enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
@@ -1054,6 +1055,9 @@ extern "C" {
    //

    // Get the backend sampled token for the ith token.
+    // With multiple outputs, sampler state advances when the token is accepted,
+    // not when it is read through this function.
+    // When accepting multiple outputs, accept a contiguous prefix in output order.
    // Returns LLAMA_TOKEN_NULL if no token was sampled.
    LLAMA_API llama_token llama_get_sampled_token_ith(struct llama_context * ctx, int32_t i);

@@ -1270,9 +1274,12 @@ extern "C" {
        // [EXPERIMENTAL]
        // backend sampling interface:

-        // return true if the backend supports all ops needed by the sampler
+        // return true if the backend supports all ops needed by the sampler and can handle up to n_outputs_max_per_seq outputs per sequence
        // note: call once per sampler
-        bool (*backend_init)(struct llama_sampler * smpl, ggml_backend_buffer_type_t buft);
+        bool (*backend_init)(
+                struct llama_sampler       * smpl,
+                ggml_backend_buffer_type_t   buft,
+                uint32_t                     n_outputs_max_per_seq);

        // call after .backend_apply()
        void (*backend_accept)(
@@ -1290,6 +1297,13 @@ extern "C" {

        // called before graph execution to set inputs for the current ubatch
        void (*backend_set_input)(struct llama_sampler * smpl);
+
+        // called before rebuilding a sampling graph to clear any internal sampler state
+        void (*backend_reset)(struct llama_sampler * smpl);
+
+        // copy mutable state from src into dst while keeping dst's references to the current sampling graph
+        // src and dst must have the same type and configuration
+        void (*copy_state)(const struct llama_sampler * src, struct llama_sampler * dst);
    };

    struct llama_sampler {
@@ -1310,6 +1324,7 @@ extern "C" {
    LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);
    LLAMA_API void                   llama_sampler_reset (      struct llama_sampler * smpl);
    LLAMA_API struct llama_sampler * llama_sampler_clone (const struct llama_sampler * smpl);
+    LLAMA_API void                   llama_sampler_copy  (const struct llama_sampler * src, struct llama_sampler * dst);
    // important: do not free if the sampler has been added to a llama_sampler_chain (via llama_sampler_chain_add)
    LLAMA_API void                   llama_sampler_free  (      struct llama_sampler * smpl);

@@ -1499,6 +1514,7 @@ extern "C" {
    LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);

    /// @details Sample and accept a token from the idx-th output of the last evaluation
+    // For multiple outputs from one sampler, call this function in output order without gaps.
    //
    // Shorthand for:
    //    const auto * logits = llama_get_logits_ith(ctx, idx);
--- a/Show More
+++ b/Show More