tests : update speculative params (#26925 )

vulkan: add TQ2_0 (ternary) support (#25850 )
* vulkan: TQ2_0 (ternary) support — dequant + dedicated mul_mat_vec + matmul via dequant_funcs First Vulkan ternary type in ggml. Correctness: OM-125m TQ2_0 vs F16 top-12 logprobs identical to 4 decimals fully offloaded (float dequant path, no Q8_K activation quant). Speed at 125m ~= F16 (overhead-bound at this scale); the bandwidth win targets larger BitNet SKUs. MMQ/int-dot path intentionally not wired yet. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * tests: enable TQ2_0 in backend-ops type lists Vulkan now implements TQ2_0 (dequant, mul_mat_vec, mul_mm, get_rows); backends without support skip via not-supported as usual. TQ1_0 stays disabled. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Michael Trabalka <michael.trabalka@sqv.ai> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-12 07:49:53 +02:00 · 2026-08-12 08:08:19 +03:00 · 2026-08-12 08:07:23 +03:00 · 2026-08-12 08:06:16 +03:00 · 2026-08-12 08:05:13 +03:00 · 2026-08-12 08:02:28 +03:00
751 changed files with 25322 additions and 9329 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"
@@ -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
@@ -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 "$PR_URL" --remove-label draft
@@ -748,6 +748,135 @@ jobs:
          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

+  windows-rocm:
+    needs: [check-release]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+
+    runs-on: windows-2022
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+            build: x64
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: windows-rocm-${{ 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: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -848,6 +977,7 @@ jobs:
          name: llama-bin-win-${{ 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 +988,16 @@ 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 +1015,7 @@ jobs:
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -885,54 +1025,62 @@ 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 +1297,8 @@ 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 +1330,36 @@ 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 +1375,6 @@ 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 $(nproc)
@@ -1258,130 +1403,6 @@ jobs:
          path: llama-${{ 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: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1555,7 +1576,7 @@ jobs:
      - windows-cpu
      - windows-cuda
      #- windows-sycl
-      - windows-hip
+      - windows-rocm
      - windows-openvino
      - ubuntu-22-rocm
      - ubuntu-cpu
@@ -1667,7 +1688,7 @@ jobs:
            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ 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)
@@ -1681,10 +1702,11 @@ 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
@@ -90,23 +96,27 @@ 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"
+          ./tests.sh

      - name: Slow tests
        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
+          SLOW_TESTS=1 ./tests.sh
@@ -72,7 +72,7 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx1, backend-sampling)
        id: server_integration_tests_backend_sampling
@@ -81,7 +81,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2)
        id: server_integration_tests_gpu2
@@ -90,7 +90,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_METAL_DEVICES=2
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2, backend-sampling)
        id: server_integration_tests_gpu2_backend_sampling
@@ -99,7 +99,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

  server-cuda:
    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
@@ -132,7 +132,7 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx1, backend-sampling)
        id: server_integration_tests_backend_sampling
@@ -141,7 +141,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2)
        id: server_integration_tests_gpu2
@@ -150,7 +150,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_CUDA_DEVICES=2
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2, backend-sampling)
        id: server_integration_tests_gpu2_backend_sampling
@@ -159,7 +159,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

  server-kleidiai:
    runs-on: ah-ubuntu_22_04-c8g_8x
@@ -219,4 +219,4 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh
@@ -104,21 +104,21 @@ jobs:
        id: server_integration_tests
        run: |
          cd tools/server/tests
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
        run: |
          cd tools/server/tests
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh

      - name: Tests (Backend sampling)
        id: server_integration_tests_backend_sampling
        run: |
          cd tools/server/tests
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Slow tests (Backend sampling)
        id: server_integration_tests_slow_backend_sampling
@@ -126,7 +126,7 @@ jobs:
        run: |
          cd tools/server/tests
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh

  windows:
    runs-on: windows-2025
@@ -167,15 +167,17 @@ jobs:

      - name: Tests
        id: server_integration_tests
+        shell: bash
        run: |
          cd tools/server/tests
-          $env:PYTHONIOENCODING = ":replace"
-          pytest -v -x -m "not slow"
+          export PYTHONIOENCODING=":replace"
+          ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
+        shell: bash
        run: |
          cd tools/server/tests
-          $env:SLOW_TESTS = "1"
-          pytest -v -x
+          export SLOW_TESTS="1"
+          ./tests.sh
@@ -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>

@@ -49,6 +49,14 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")

+# gpu-rocm self-hosted runner can't upload logs to blob; keep each run's logs in
+# their own dir keyed by the GitHub run id so an Actions run URL maps to its logs.
+if [ -n "${GG_BUILD_ROCM}" ] && [ -n "${GITHUB_RUN_ID}" ]; then
+    OUT="$OUT/run-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT:-1}"
+    mkdir -p "$OUT"
+    echo "ci results dir: $OUT"
+fi
+
 rm -f $OUT/*.log
 rm -f $OUT/*.exit
 rm -f $OUT/*.md
@@ -92,7 +100,7 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
 fi

 if [ ! -z ${GG_BUILD_ROCM} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON -DGGML_HIP_ROCWMMA_FATTN=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON"
    if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
        echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
        exit 1
@@ -0,0 +1,26 @@
+# Used to cross-compile ggml-cuda for Windows ARM64 on an x64 Windows host.
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+if ( DEFINED CUDAToolkit_ROOT )
+    file( TO_CMAKE_PATH "${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 "" )
@@ -2605,14 +2605,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_env("LLAMA_ARG_DIO"));
    add_opt(common_arg(
        {"-lm", "--load-mode"}, "MODE",
-        "model loading mode (default: mmap)\n"
+        "model loading mode (default: auto)\n"
+        "- auto: mmap, unless a device does not support it\n"
        "- none: no special loading mode\n"
        "- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)\n"
        "- mlock: force system to keep model in RAM rather than swapping or compressing\n"
        "- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
        "- dio: use DirectIO if available\n",
        [](common_params & params, const std::string & value) {
-            /**/ if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
+            /**/ if (value == "auto")       { params.load_mode = LLAMA_LOAD_MODE_AUTO;       }
+            else if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
            else if (value == "mmap")       { params.load_mode = LLAMA_LOAD_MODE_MMAP;       }
            else if (value == "mlock")      { params.load_mode = LLAMA_LOAD_MODE_MLOCK;      }
            else if (value == "mmap+mlock") { params.load_mode = LLAMA_LOAD_MODE_MMAP_MLOCK; }
@@ -3308,6 +3310,17 @@ 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"
@@ -1166,6 +1166,16 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
        data.prompt += data.generation_prompt;
    }

+    std::vector<std::string> tool_call_starts = { "<tool_call>" };
+
+    // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
+    // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
+    // constraining on <function which may occur in valid content generation, e.g. #include <functional>
+    foreach_function(inputs.tools, [&](const json & tool) {
+        const std::string name = tool.at("function").at("name");
+        tool_call_starts.push_back("<function=" + name + ">");
+    });
+
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        auto generation_prompt = p.literal(GEN_PREFIX);

@@ -1238,7 +1248,7 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));

            return generation_prompt +
-                   (reasoning << p.content(p.until_one_of({ "<tool_call>", "<function=" })) << tool_calls);
+                   (reasoning << p.content(p.until_one_of(tool_call_starts)) << tool_calls);
        }

        // Content only parser
@@ -1264,12 +1274,9 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
        });

        if (data.grammar_lazy) {
-            data.grammar_triggers = {
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>" },
-                // Trigger on "<function" and not "<function=" because the trailing "=" is part of
-                // the token with the function name e.g. "=read"
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"   },
-            };
+            for (const auto & start : tool_call_starts) {
+                data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, start });
+            }
        }
    }

@@ -3086,6 +3093,153 @@ 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_one_of({ "<|eot|>", "<|eom|>" })));
+
+        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;
+            }
+            auto trailing_calls = p.optional(p.literal("<|eom|>") + start + tool_calls);
+            return p.zero_or_more(start + analysis) + start + (tool_calls | (final_msg + trailing_calls));
+        }
+
+        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 +3268,12 @@ 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)
@@ -472,7 +473,7 @@ struct common_params {
    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);

    enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_MMAP; // how to load the model
+    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model

    common_cpu_params cpuparams;
    common_cpu_params cpuparams_batch;
@@ -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
@@ -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,
@@ -570,23 +570,34 @@ struct parser_executor {
    }

    static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
+        auto save = pos;
+
        ++pos; // consume '\'
        if (pos >= ctx.input.size()) {
            if (!ctx.is_lenient()) {
                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
            }
+            pos = save; // suppress unmatched '\'
            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
        }

        char c = ctx.input[pos];
+
        if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
            ++pos;
            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
-        } else if (c == 'u') {
-            return handle_unicode_escape(ctx, start, pos);
-        } else {
-            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
        }
+
+        if (c == 'u') {
+            auto result = handle_unicode_escape(ctx, start, pos);
+            if (result.need_more_input()) {
+                pos = save; // suppress incomplete sequence
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
+            }
+            return result;
+        }
+
+        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
    }

    static common_peg_parse_result handle_unicode_escape(common_peg_parse_context & ctx, size_t start, size_t & pos) {
@@ -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);
@@ -171,12 +171,6 @@ struct common_speculative_impl {
    // (optional) serialize/restore per-seq internal state (e.g. eagle3's deferred boundary).
    virtual bool get_state(llama_seq_id /*seq_id*/, std::vector<uint8_t> & /*data*/) const { return false; }
    virtual void set_state(llama_seq_id /*seq_id*/, const std::vector<uint8_t> & /*data*/) {}
-
-    // true if this implementation requires the target context to extract post-norm embeddings
-    virtual bool need_embd() const = 0;
-
-    // true if this implementation requires the target context to extract pre-norm embeddings
-    virtual bool need_embd_nextn() const { return false; }
 };

 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -193,6 +187,10 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
        auto * ctx_dft = this->params.ctx_dft;
        auto * ctx_tgt = this->params.ctx_tgt;

+        if (!ctx_dft) {
+            throw std::runtime_error("draft-simple requires a draft context");
+        }
+
        SPC_TRC("%s", "adding speculative implementation 'draft-simple'\n");
        SPC_TRC("- n_max=%d, n_min=%d, p_min=%f\n", this->params.n_max, this->params.n_min, this->params.p_min);
        SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n",
@@ -385,10 +383,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };


@@ -907,10 +901,6 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        pending_g_last[seq_id].resize(n_embd_dec);
        std::memcpy(pending_g_last[seq_id].data(), data.data() + sizeof(llama_pos), (size_t) n_embd_dec * sizeof(float));
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 // DFlash: block-diffusion drafting with a draft-side KV cache injection
@@ -1032,7 +1022,14 @@ 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;
        }

@@ -1240,10 +1237,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_draft_mtp : public common_speculative_impl {
@@ -1682,14 +1675,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
        const size_t row_bytes = (size_t) n_embd * sizeof(float);
        std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
    }
-
-    bool need_embd() const override {
-        return false;
-    }
-
-    bool need_embd_nextn() const override {
-        return true;
-    }
 };

 // state of self-speculation (simple implementation, not ngram-map)
@@ -1736,10 +1721,6 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
@@ -1794,10 +1775,6 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {

        common_ngram_map_accept(config[seq_id], n_accepted);
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_mod : public common_speculative_impl {
@@ -1973,10 +1950,6 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
            }
        }
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_cache : public common_speculative_impl {
@@ -2116,10 +2089,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative {
@@ -2292,6 +2261,7 @@ 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;
 }
@@ -2314,7 +2284,6 @@ common_speculative_init_result::common_speculative_init_result(
    const bool spec_mtp = std::find(params.speculative.types.begin(),
                                    params.speculative.types.end(),
                                    COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-    GGML_ASSERT(has_draft || spec_mtp);

    auto mparams = common_model_params_to_llama(params);
    auto cparams = common_context_params_to_llama(params);
@@ -2377,6 +2346,17 @@ 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) {
@@ -2541,34 +2521,6 @@ bool common_speculative_process(common_speculative * spec, const llama_batch & b
    return result;
 }

-bool common_speculative_need_embd(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
-bool common_speculative_need_embd_nextn(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd_nextn()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
 void common_speculative_draft(common_speculative * spec) {
    if (spec == nullptr) {
        return;
@@ -2653,7 +2605,10 @@ void common_speculative_draft(common_speculative * spec) {
 void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, uint16_t n_accepted) {
    common_speculative_impl * impl = spec->impl_last[seq_id];

-    GGML_ASSERT(impl);
+    if (impl == nullptr) {
+        GGML_ASSERT(n_accepted == 0);
+        return;
+    }

    {
        common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
@@ -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);
@@ -58,12 +67,6 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);

-// true if any implementation requires target post-norm embeddings to be extracted
-bool common_speculative_need_embd(common_speculative * spec);
-
-// true if any implementation requires target nextn embeddings to be extracted
-bool common_speculative_need_embd_nextn(common_speculative * spec);
-
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);

@@ -103,6 +103,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteMoeForCausalLM": "granite",
    "GraniteMoeHybridForCausalLM": "granite",
    "GraniteMoeSharedForCausalLM": "granite",
+    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
    "Grok1ForCausalLM": "grok",
@@ -182,6 +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",
@@ -211,6 +214,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3MoeForCausalLM": "qwen",
    "Qwen3NextForCausalLM": "qwen",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "PocketTTSModel": "pockettts",
    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
@@ -297,6 +301,7 @@ 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",
@@ -306,6 +311,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "Qwen2_5_VLForConditionalGeneration": "qwenvl",
    "Qwen3ASRForConditionalGeneration": "qwen3vl",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "PocketTTSModel": "pockettts",
    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
@@ -58,6 +58,11 @@ logger = logging.getLogger("hf-to-gguf")
 AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")


+# for checkpoints that ship no config.json, we will try to provide a synthetic one
+HparamsMatcher = Callable[[Path], bool]
+HparamsLoader = Callable[[Path], dict[str, Any]]
+
+
 class SentencePieceTokenTypes(IntEnum):
    NORMAL = 1
    UNKNOWN = 2
@@ -77,6 +82,7 @@ class ModelBase:
        ModelType.TEXT: {},
        ModelType.MMPROJ: {},
    }
+    _hparams_loaders: list[tuple[HparamsMatcher, HparamsLoader]] = []

    dir_model: Path
    ftype: gguf.LlamaFileType
@@ -823,7 +829,7 @@ class ModelBase:
            elif any(str(v.get("quant_algo")).endswith("NVFP4") for v in quant_layers.values() if isinstance(v, dict)):
                quant_algo = "NVFP4"

-        self._is_nvfp4 = quant_algo == "NVFP4"
+        self._is_nvfp4 = quant_algo in ("NVFP4", "W4A16_NVFP4")
        self._is_mxfp4 = quant_method == "mxfp4"

        # NVFP4 weights are repacked and written directly to gguf_writer.
@@ -1040,6 +1046,24 @@ class ModelBase:

        return part_names

+    @staticmethod
+    def load_hparams_guess(dir_model: Path) -> dict[str, Any] | None:
+        # some models ship no config.json, will try to guess them
+        from conversion import load_all_models
+        load_all_models()
+
+        for matcher, loader in ModelBase._hparams_loaders:
+            if matcher(dir_model):
+                return loader(dir_model)
+        return None
+
+    @classmethod
+    def register_hparams_loader(cls, matcher: HparamsMatcher) -> Callable[[HparamsLoader], HparamsLoader]:
+        def inner(loader: HparamsLoader) -> HparamsLoader:
+            cls._hparams_loaders.append((matcher, loader))
+            return loader
+        return inner
+
    @staticmethod
    def load_hparams(dir_model: Path, is_mistral_format: bool):
        if is_mistral_format:
@@ -1053,6 +1077,10 @@ class ModelBase:
            config = AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
        except Exception as e:
            logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            if not (dir_model / "config.json").is_file():
+                config = ModelBase.load_hparams_guess(dir_model)
+                if config is not None:
+                    return config
            logger.warning("Trying to load config.json instead")
            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                config = json.load(f)
@@ -665,7 +665,18 @@ class Gemma4Model(Gemma3Model):
        swa_layers = [t == "sliding_attention" for t in self.hparams["layer_types"]]
        self.gguf_writer.add_sliding_window_pattern(swa_layers)

-        head_dim_full = self.hparams["global_head_dim"]
+        per_layer_config = self.hparams.get("per_layer_config")
+        layer_types = self.hparams.get("layer_types", [])
+        if (head_dim_full := self.hparams.get("global_head_dim")) is None and per_layer_config is not None:
+            for layer_idx, layer_config in per_layer_config.items():
+                layer_idx = int(layer_idx)
+                if layer_idx < len(layer_types):
+                    if layer_types[layer_idx] == "full_attention" and "head_dim" in layer_config:
+                        head_dim_full = layer_config["head_dim"]
+                        break
+
+        assert head_dim_full is not None
+
        head_dim_swa = self.hparams["head_dim"]
        # correct the head dim for global/swa layers
        self.gguf_writer.add_key_length(head_dim_full)
@@ -685,8 +696,14 @@ class Gemma4Model(Gemma3Model):
            n_ff_arr = [n_ff if il < first_kv_shared_layer_idx else n_ff * 2 for il in range(self.block_count)]
            self.gguf_writer.add_feed_forward_length(n_ff_arr)

-        # handle num_global_key_value_heads
-        num_key_value_heads_full = self.hparams.get("num_global_key_value_heads")
+        if (num_key_value_heads_full := self.hparams.get("num_global_key_value_heads")) is None and per_layer_config is not None:
+            for layer_idx, layer_config in per_layer_config.items():
+                layer_idx = int(layer_idx)
+                if layer_idx < len(layer_types):
+                    if layer_types[layer_idx] == "full_attention" and "num_key_value_heads" in layer_config:
+                        num_key_value_heads_full = layer_config["num_key_value_heads"]
+                        break
+
        num_key_value_heads_swa = self.hparams.get("num_key_value_heads")
        if num_key_value_heads_full is not None and num_key_value_heads_swa is not None:
            value_arr = [num_key_value_heads_swa if is_swa else num_key_value_heads_full for is_swa in swa_layers]
@@ -708,7 +725,19 @@ class Gemma4Model(Gemma3Model):
        # IMPORTANT: this ROPE_FREQS tensor is ONLY used by the full_attention layers
        rope_params_full = self.hparams["rope_parameters"]["full_attention"]
        assert rope_params_full["rope_type"] == "proportional"
-        head_dim_full = (self.hparams["global_head_dim"])
+
+        per_layer_config = self.hparams.get("per_layer_config")
+        if (head_dim_full := self.hparams.get("global_head_dim")) is None and per_layer_config is not None:
+            layer_types = self.hparams.get("layer_types", [])
+            for layer_idx, layer_config in per_layer_config.items():
+                layer_idx = int(layer_idx)
+                if layer_idx < len(layer_types):
+                    if layer_types[layer_idx] == "full_attention" and "head_dim" in layer_config:
+                        head_dim_full = layer_config["head_dim"]
+                        break
+
+        assert head_dim_full is not None
+
        partial_rotary_factor_full = rope_params_full["partial_rotary_factor"]
        n_rot_full = int(head_dim_full * partial_rotary_factor_full / 2)
        n_unrot_full = int(head_dim_full / 2) - n_rot_full
@@ -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 +266,44 @@ 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
+            # Include lm_head scale sidecars so NVFP4 packing sees them.
+            keep = name in (
+                "backbone.embeddings.weight",
+                "backbone.norm_f.weight",
+                "lm_head.weight",
+                "lm_head.weight_scale",
+                "lm_head.weight_scale_2",
+                "lm_head.weight_scale_inv",
+                "lm_head.input_scale",
+                "lm_head.input_global_scale",
+                "lm_head.weight_global_scale",
+                "lm_head.weight_packed",
+            )
+            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 +342,10 @@ 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 +412,24 @@ 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
@@ -0,0 +1,378 @@
+from __future__ import annotations
+
+import re
+from pathlib import Path
+from typing import Any, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, MmprojModel, SentencePieceTokenTypes, TextModel, gguf, logger
+
+# Pocket TTS is a CALM: the backbone conditions a flow-matching decoder that generates one
+# continuous 32-d latent per frame. There is no codebook in this model.
+# The checkpoint ships no config.json, hparams come from _load_hparams() below.
+#
+# Tricks being used to support this model via existing llama.cpp code paths:
+# - bos_before_voice and bos_emb are learned input vectors, not tokens
+#   they are appended to the embedding table as extra tokens, to be looked up like any other row
+# - bos_emb lives in latent space, so input_linear is folded into it here
+# - the backbone has no lm_head, the embedding table is reused as output for the unused logits
+#
+# pipeline stage mapping:
+#   mimi encoder + speaker_proj --> mapped to normal mtmd audio encoder
+#   flow_lm.transformer         --> mapped to normal libllama text model (autoregressive)
+#   flow_lm.flow_net + out_eos  --> MTMD_GEN_PROCESS_TYPE_GEN_CODE
+#   mimi decoder                --> MTMD_GEN_PROCESS_TYPE_GEN_WAV
+
+# indices into mimi.encoder.model / mimi.decoder.model for stage i, see SEANetEncoder/SEANetDecoder
+_ENC_RES_IDX   = lambda i: 1 + 3 * i  # noqa: E731
+_ENC_SCALE_IDX = lambda i: 3 + 3 * i  # noqa: E731
+_DEC_SCALE_IDX = lambda i: 2 + 3 * i  # noqa: E731
+_DEC_RES_IDX   = lambda i: 3 + 3 * i  # noqa: E731
+
+_N_SEANET_STAGES = 3
+_SAMPLE_RATE = 24000
+
+
+def _tensor_shapes(dir_model: Path) -> dict[str, tuple[int, ...]]:
+    part_names = ModelBase.get_model_part_names(dir_model, "model", ".safetensors")
+    if len(part_names) != 1:
+        return {}
+    with gguf.utility.SafetensorsLocal(dir_model / part_names[0]) as part:
+        return {name: tuple(part[name].shape) for name in part.keys()}
+
+
+@ModelBase.register_hparams_loader(lambda dir_model: "flow_lm.bos_emb" in _tensor_shapes(dir_model))
+def _load_hparams(dir_model: Path) -> dict[str, Any]:
+    logger.info("gguf: detected pocket-tts checkpoint, deriving hparams from tensor shapes")
+    shapes = _tensor_shapes(dir_model)
+    n_vocab, n_embd = shapes["flow_lm.conditioner.embed.weight"]
+    n_layer = sum(1 for name in shapes if re.fullmatch(r"flow_lm\.transformer\.layers\.\d+\.norm1\.weight", name))
+    n_layer_a = sum(1 for name in shapes if re.fullmatch(r"mimi\.encoder_transformer\.transformer\.layers\.\d+\.norm1\.weight", name))
+    n_embd_a = shapes["mimi.encoder_transformer.transformer.layers.0.norm1.weight"][0]
+    return {
+        "architectures": ["PocketTTSModel"],
+        "model_type": "pockettts",
+        "num_hidden_layers": n_layer,
+        "hidden_size": n_embd,
+        "intermediate_size": shapes["flow_lm.transformer.layers.0.linear1.weight"][0],
+        # the transformer is fully causal with no context limit, this only bounds the KV cache
+        "max_position_embeddings": 4096,
+        # not in the checkpoint, but every released variant uses head_dim 64
+        "num_attention_heads": n_embd // 64,
+        # extra rows for the learned input vectors, see _embd_table()
+        "vocab_size": n_vocab + (2 if "flow_lm.bos_before_voice" in shapes else 1),
+        "rope_theta": 10000.0,
+        "layer_norm_eps": 1e-5,
+        "audio_config": {
+            "num_hidden_layers": n_layer_a,
+            "hidden_size": n_embd_a,
+            "intermediate_size": shapes["mimi.encoder_transformer.transformer.layers.0.linear1.weight"][0],
+            "num_attention_heads": n_embd_a // 64,
+        },
+    }
+
+
+@ModelBase.register("PocketTTSModel")
+class PocketTTSModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.POCKETTTS
+
+    _LAYER_TENSOR_MAP = {
+        "norm1":               gguf.MODEL_TENSOR.ATTN_NORM,
+        "norm2":               gguf.MODEL_TENSOR.FFN_NORM,
+        "self_attn.out_proj":  gguf.MODEL_TENSOR.ATTN_OUT,
+        "linear1":             gguf.MODEL_TENSOR.FFN_UP,
+        "linear2":             gguf.MODEL_TENSOR.FFN_DOWN,
+    }
+
+    def set_vocab(self):
+        # this is a unigram sentencepiece model, llama.cpp's SPM tokenizer cannot do
+        # unigram segmentation, so use the UGM tokenizer instead
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        proto = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        proto.ParseFromString(open(self.dir_model / "tokenizer.model", "rb").read())
+        assert proto.trainer_spec.model_type == 1, "expected a unigram tokenizer"
+
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+
+        # the last rows of the embedding table are not sentencepiece pieces
+        extra = self._extra_tokens()
+        for i, name in enumerate(extra):
+            tokens[len(tokens) - len(extra) + i] = name.encode("utf-8")
+            toktypes[len(tokens) - len(extra) + i] = SentencePieceTokenTypes.CONTROL
+            scores[len(tokens) - len(extra) + i] = -1000.0
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(proto.normalizer_spec.add_dummy_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(proto.normalizer_spec.remove_extra_whitespaces)
+        if proto.normalizer_spec.precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(proto.normalizer_spec.precompiled_charsmap)
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(False)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if not name.startswith("flow_lm."):
+            return  # mimi and the flow net go to the mmproj
+
+        if name == "flow_lm.conditioner.embed.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD), self._embd_table(data_torch))
+            return
+
+        if name.startswith("flow_lm.out_norm."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("flow_lm.transformer.layers."):
+            assert bid is not None
+            key_with_suffix = name.split(f"layers.{bid}.", 1)[1]
+            key, suffix = key_with_suffix.rsplit(".", 1)
+
+            if key == "self_attn.in_proj":
+                q, k, v = data_torch.chunk(3, dim=0)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+                return
+
+            tensor = self._LAYER_TENSOR_MAP.get(key)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid, suffix="." + suffix), data_torch)
+                return
+
+        return
+
+    def _extra_tokens(self) -> list[str]:
+        # the conditioner's padding row, then the learned vectors appended by _embd_table().
+        # bos_before_voice only exists when the pack sets insert_bos_before_voice
+        names = ["<|pad|>"]
+        if "flow_lm.bos_before_voice" in self.model_tensors:
+            names.append("<|bos_before_voice|>")
+        names.append("<|audio_bos|>")
+        return names
+
+    def _embd_table(self, embed: Tensor) -> Tensor:
+        rows = [embed]
+        if "flow_lm.bos_before_voice" in self.model_tensors:
+            rows.append(self.model_tensors["flow_lm.bos_before_voice"]().reshape(1, -1).to(embed.dtype))
+
+        # bos_emb is a latent, it only enters the backbone through input_linear
+        bos_emb = self.model_tensors["flow_lm.bos_emb"]()
+        input_linear = self.model_tensors["flow_lm.input_linear.weight"]()
+        audio_bos = torch.nn.functional.linear(bos_emb.float(), input_linear.float()).reshape(1, -1)
+        rows.append(audio_bos.to(embed.dtype))
+
+        return torch.cat(rows, dim=0)
+
+
+@ModelBase.register("PocketTTSModel")
+class PocketTTSMmprojModel(MmprojModel):
+    has_audio_encoder = True
+    has_vision_encoder = False
+
+    _MIMI_TFM_MAP = {
+        "norm1":              (gguf.MODEL_TENSOR.A_ENC_INPUT_NORM,  gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_NORM),
+        "norm2":              (gguf.MODEL_TENSOR.A_ENC_OUTPUT_NORM, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_NORM),
+        "self_attn.out_proj": (gguf.MODEL_TENSOR.A_ENC_OUTPUT,      gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_OUT),
+        "linear1":            (gguf.MODEL_TENSOR.A_ENC_FFN_UP,      gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_UP),
+        "linear2":            (gguf.MODEL_TENSOR.A_ENC_FFN_DOWN,    gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_DOWN),
+        "layer_scale_1.scale": (gguf.MODEL_TENSOR.A_ENC_ATTN_SCALE, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_SCALE),
+        "layer_scale_2.scale": (gguf.MODEL_TENSOR.A_ENC_FFN_SCALE_LS, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_SCALE),
+    }
+    _MIMI_TFM_QKV = (
+        (gguf.MODEL_TENSOR.A_ENC_ATTN_Q, gguf.MODEL_TENSOR.A_ENC_ATTN_K, gguf.MODEL_TENSOR.A_ENC_ATTN_V),
+        (gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_Q, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_K, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_V),
+    )
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+        assert self.hparams_audio is not None
+
+        # voice-prompt encoder: mimi encoder + speaker_proj
+        self.gguf_writer.add_clip_has_audio_encoder(True)
+        # note: the 24kHz sample rate is hardcoded on the clip.cpp side, like the other audio models
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_SPKENC)
+        self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_audio_block_count(self.hparams_audio["num_hidden_layers"])
+        self.gguf_writer.add_audio_embedding_length(self.hparams_audio["hidden_size"])
+        self.gguf_writer.add_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+        self.gguf_writer.add_audio_head_count(self.hparams_audio["num_attention_heads"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+        # mimi convolves the waveform directly, it is passed around as a 1-row "mel"
+        self.gguf_writer.add_audio_num_mel_bins(1)
+
+        # generation: flow-matching decoder + mimi decoder
+        # the SEANet and flow net hparams are constant across the family, clip.cpp holds them
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_GEN)
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(self.hparams_audio["hidden_size"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(self.hparams_audio["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(self.hparams_audio["num_attention_heads"])
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(1e-5)
+
+        self.gguf_writer.add_gen_audio_model_variant(self.dir_model.name)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del name, bid, n_dims
+        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
+        if ".seanet." in new_name or new_name in ("a.downsample.conv.weight", "a.gen.wav.upsample.weight"):
+            return gguf.GGMLQuantizationType.F16
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # the block index of the mimi transformers is parsed here, not by the base class
+        T = gguf.MODEL_TENSOR
+
+        if name in ("flow_lm.bos_emb", "flow_lm.bos_before_voice", "flow_lm.conditioner.embed.weight"):
+            return  # folded into the backbone embedding table
+        if name.startswith("flow_lm.transformer.") or name.startswith("flow_lm.out_norm."):
+            return  # backbone
+
+        if name == "flow_lm.speaker_proj_weight":
+            yield (self.format_tensor_name(T.A_ENC_SPEAKER_PROJ), data_torch)
+            return
+        if name == "flow_lm.input_linear.weight":
+            yield (self.format_tensor_name(T.A_GEN_INPUT_LINEAR), data_torch)
+            return
+        if name == "flow_lm.emb_mean":
+            yield (self.format_tensor_name(T.A_GEN_EMB_MEAN, suffix=""), data_torch)
+            return
+        if name == "flow_lm.emb_std":
+            yield (self.format_tensor_name(T.A_GEN_EMB_STD, suffix=""), data_torch)
+            return
+        if name.startswith("flow_lm.out_eos."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(T.A_GEN_OUT_EOS, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("flow_lm.flow_net."):
+            yield from self._flow_net_tensor(name, data_torch)
+            return
+
+        if name == "mimi.downsample.conv.conv.weight":
+            yield (self.format_tensor_name(T.A_ENC_DOWNSAMPLE_CONV), data_torch)
+            return
+        if name == "mimi.upsample.convtr.convtr.weight":
+            yield (self.format_tensor_name(T.A_GEN_WAV_UPSAMPLE), data_torch)
+            return
+        if name == "mimi.quantizer.output_proj.weight":
+            yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_OUT), data_torch.squeeze(-1))
+            return
+
+        if "_transformer.transformer.layers." in name:
+            yield from self._mimi_tfm_tensor(name, data_torch)
+            return
+
+        if name.startswith("mimi.encoder.model.") or name.startswith("mimi.decoder.model."):
+            yield from self._seanet_tensor(name, data_torch)
+            return
+
+        return
+
+    def _flow_net_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        key = name.split("flow_lm.flow_net.", 1)[1]
+        suffix = "." + key.rsplit(".", 1)[1]
+
+        simple = {
+            "input_proj":                 T.A_GEN_FLOW_INPUT_PROJ,
+            "cond_embed":                 T.A_GEN_FLOW_COND_EMBD,
+            "final_layer.linear":         T.A_GEN_FLOW_FINAL_PROJ,
+            "final_layer.adaLN_modulation.1": T.A_GEN_FLOW_FINAL_ADA,
+        }
+        tensor = simple.get(key.rsplit(".", 1)[0])
+        if tensor is not None:
+            yield (self.format_tensor_name(tensor, suffix=suffix), data_torch)
+            return
+
+        if key.startswith("time_embed."):
+            bid = int(key.split(".")[1])
+            rest = key.split(f"time_embed.{bid}.", 1)[1]
+            time_map = {
+                "freqs":       (T.A_GEN_FLOW_TIME_FREQS, ""),
+                "mlp.0":       (T.A_GEN_FLOW_TIME_UP,    suffix),
+                "mlp.2":       (T.A_GEN_FLOW_TIME_DOWN,  suffix),
+                "mlp.3.alpha": (T.A_GEN_FLOW_TIME_NORM,  ""),
+            }
+            entry = time_map.get(rest) or time_map.get(rest.rsplit(".", 1)[0])
+            if entry is not None:
+                yield (self.format_tensor_name(entry[0], bid, suffix=entry[1]), data_torch)
+            return
+
+        if key.startswith("res_blocks."):
+            bid = int(key.split(".")[1])
+            rest = key.split(f"res_blocks.{bid}.", 1)[1].rsplit(".", 1)[0]
+            blk_map = {
+                "in_ln":                T.A_GEN_FLOW_BLK_NORM,
+                "mlp.0":                T.A_GEN_FLOW_BLK_UP,
+                "mlp.2":                T.A_GEN_FLOW_BLK_DOWN,
+                "adaLN_modulation.1":   T.A_GEN_FLOW_BLK_ADA,
+            }
+            tensor = blk_map.get(rest)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid, suffix=suffix), data_torch)
+            return
+
+    def _mimi_tfm_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        is_decoder = name.startswith("mimi.decoder_transformer.")
+        bid = int(name.split("_transformer.transformer.layers.", 1)[1].split(".")[0])
+        key_with_suffix = name.split(f".layers.{bid}.", 1)[1]
+
+        if key_with_suffix == "self_attn.in_proj.weight":
+            q, k, v = data_torch.chunk(3, dim=0)
+            names = self._MIMI_TFM_QKV[1 if is_decoder else 0]
+            for tensor, part in zip(names, (q, k, v)):
+                yield (self.format_tensor_name(tensor, bid), part)
+            return
+
+        key, suffix = key_with_suffix.rsplit(".", 1)
+        entry = self._MIMI_TFM_MAP.get(key) or self._MIMI_TFM_MAP.get(key_with_suffix)
+        if entry is None:
+            return
+        tensor = entry[1 if is_decoder else 0]
+        suffix = ".weight" if key_with_suffix.endswith(".scale") else "." + suffix
+        yield (self.format_tensor_name(tensor, bid, suffix=suffix), data_torch)
+
+    def _seanet_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        is_decoder = name.startswith("mimi.decoder.")
+        idx = int(name.split(".model.", 1)[1].split(".")[0])
+        suffix = "." + name.rsplit(".", 1)[1]
+
+        conv_in, conv_out, res1, res2, scale = (
+            (T.A_GEN_WAV_SEANET_CONV_IN, T.A_GEN_WAV_SEANET_CONV_OUT, T.A_GEN_WAV_SEANET_RES_CONV1,
+             T.A_GEN_WAV_SEANET_RES_CONV2, T.A_GEN_WAV_SEANET_SCALE_CONV)
+            if is_decoder else
+            (T.A_ENC_SEANET_CONV_IN, T.A_ENC_SEANET_CONV_OUT, T.A_ENC_SEANET_RES_CONV1,
+             T.A_ENC_SEANET_RES_CONV2, T.A_ENC_SEANET_SCALE_CONV)
+        )
+
+        if idx == 0:
+            yield (self.format_tensor_name(conv_in, suffix=suffix), data_torch)
+            return
+        if idx == 3 * _N_SEANET_STAGES + 2:
+            yield (self.format_tensor_name(conv_out, suffix=suffix), data_torch)
+            return
+
+        for stage in range(_N_SEANET_STAGES):
+            res_idx = _DEC_RES_IDX(stage) if is_decoder else _ENC_RES_IDX(stage)
+            scale_idx = _DEC_SCALE_IDX(stage) if is_decoder else _ENC_SCALE_IDX(stage)
+            if idx == scale_idx:
+                yield (self.format_tensor_name(scale, stage, suffix=suffix), data_torch)
+                return
+            if idx == res_idx:
+                # block.1 is the dilated conv, block.3 the pointwise one (0 and 2 are ELU)
+                inner = int(name.split(".block.", 1)[1].split(".")[0])
+                tensor = res1 if inner == 1 else res2
+                yield (self.format_tensor_name(tensor, stage, suffix=suffix), data_torch)
+                return
@@ -647,10 +647,13 @@ class DFlashModel(Qwen3Model):
        # own tokenizer logic, not the Qwen default).
        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_hparams = json.load(f)
+            target_arch = target_hparams["architectures"][0]
        target_cls = get_model_class(target_arch)

        if target_cls is not type(self):
+            if target_arch == "NemotronHForCausalLM":
+                setattr(self, "is_moe", "num_experts_per_tok" in target_hparams)
            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
        else:
            super().set_vocab()
@@ -688,6 +691,12 @@ class DFlashModel(Qwen3Model):
            name = "model." + name
        return super().filter_tensors((name, gen))

+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "model.embed_tokens.weight" and not self.hparams.get("has_embed_tokens", True):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+

@ModelBase.register("Qwen3DSparkModel")
 class DSparkModel(DFlashModel):
@@ -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);
@@ -1,6 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 torch
-torchvision
+torchvision; platform_machine != "s390x"
 transformers
 huggingface-hub
 accelerate
@@ -2,12 +2,15 @@

 import argparse
 import os
+import sys
 import importlib
 import torch
 import numpy as np

 from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM
-from pathlib import Path
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+from utils.common import save_output_data

 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')

@@ -54,6 +57,7 @@ print(f"Model name: {model_name}")

 prompt = "Hello world today"
 input_ids = tokenizer(prompt, return_tensors="pt").input_ids  # ty: ignore[call-non-callable]
+token_ids = input_ids[0].cpu().tolist()
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")
 print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")  # ty: ignore[unresolved-attribute]
@@ -74,21 +78,8 @@ with torch.no_grad():
    print(f"Hidden dimension: {token_embeddings.shape[-1]}")
    print(f"Number of tokens: {token_embeddings.shape[0]}")

-    # Save raw token embeddings
-    data_dir = Path("data")
-    data_dir.mkdir(exist_ok=True)
-    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
-    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
-
-    # Save all token embeddings as binary
    print(token_embeddings)
-    token_embeddings.astype(np.float32).tofile(bin_filename)
-
-    # Save as text for inspection
-    with open(txt_filename, "w") as f:
-        for i, embedding in enumerate(token_embeddings):
-            for j, val in enumerate(embedding):
-                f.write(f"{i} {j} {val:.6f}\n")
+    save_output_data(token_embeddings, token_ids, prompt, model_name, type_suffix="-embeddings")

    # Print embeddings per token in the requested format
    print("\nToken embeddings:")
@@ -110,5 +101,3 @@ with torch.no_grad():
    for i, token in enumerate(tokens):
        print(f"  Token {i}: {repr(token)}")

-    print(f"Saved bin logits to: {bin_filename}")
-    print(f"Saved txt logist to: {txt_filename}")
@@ -3,10 +3,47 @@
 Demonstration of basic greedy speculative decoding

 ```bash
+# spec-type draft-simple
 ./bin/llama-speculative-simple \
-    -m  ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
-    -md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
-    -f test.txt -c 0 -ngl 99 --color on \
-    --sampling-seq k --top-k 1 -fa on --temp 0.0 \
-    -ngld 99 --spec-draft-n-max 16 --spec-draft-n-draft-min 5 --draft-p-min 0.9
+        -hf  ggml-org/Qwen3-8B-Base-GGUF:Q8_0 \
+        -hfd ggml-org/Qwen3-0.6B-Base-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-simple --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-mtp
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3.6-27B-GGUF:Q8_0 \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-mtp (with shared KV cache)
+# note: this model needs a <s> token at the start to somewhat work without the chat template
+./bin/llama-speculative-simple \
+        -hf ggml-org/Gemma-4-31B-it-GGUF:Q8_0 \
+        -p "<s>Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-eagle3
+./bin/llama-speculative-simple \
+        -hf ggml-org/gpt-oss-20b-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-eagle3 --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-dflash
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3-8B-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-dflash --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-dspark
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3-8B-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-dspark --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
 ```
@@ -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);
@@ -45,45 +51,23 @@ int main(int argc, char ** argv) {

    const llama_vocab * vocab = llama_model_get_vocab(model_tgt);

-    // load the draft model
-    llama_model_ptr model_dft;
-    llama_context_ptr ctx_dft;
+    // load the draft model (if any) - this also creates the MTP draft context when MTP speculation is enabled
+    common_speculative_init_result_ptr spec_init;

-    // TODO: simplify this logic
    {
-        const auto & params_spec = params.speculative.draft;
+        common_params params_dft = common_base_params_to_speculative(params);

-        auto params_dft = params;
-
-        params_dft.devices      = params_spec.devices;
-        params_dft.model        = params_spec.mparams;
-        params_dft.n_gpu_layers = params_spec.n_gpu_layers;
-
-        if (params_spec.cpuparams.n_threads > 0) {
-            params_dft.cpuparams.n_threads       = params.speculative.draft.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
-        }
-
-        params_dft.tensor_buft_overrides = params.speculative.draft.tensor_buft_overrides;
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-
-        model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
-        if (model_dft == nullptr) {
-            LOG_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
-            return 1;
-        }
-
-        auto cparams = common_context_params_to_llama(params_dft);
-        ctx_dft.reset(llama_init_from_model(model_dft.get(), cparams));
+        spec_init = common_speculative_init_from_params(params_dft, model_tgt, ctx_tgt);

        params.speculative.draft.ctx_tgt = ctx_tgt;
-        params.speculative.draft.ctx_dft = ctx_dft.get();
+        params.speculative.draft.ctx_dft = spec_init->context();
    }

+    llama_context * ctx_dft = params.speculative.draft.ctx_dft;
+
    // check if the context supports partial sequence removal
-    const bool use_ckpt_tgt = (common_context_can_seq_rm(ctx_tgt)       == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
-    const bool use_ckpt_dft = (common_context_can_seq_rm(ctx_dft.get()) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
+    const bool use_ckpt_tgt = common_context_can_seq_rm(ctx_tgt) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+    const bool use_ckpt_dft = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;

    if (use_ckpt_tgt) {
        LOG_INF("speculative decoding will use checkpoints (context does not support partial sequence removal)\n");
@@ -129,9 +113,30 @@ int main(int argc, char ** argv) {
    // target model sampling context
    common_sampler_ptr smpl(common_sampler_init(model_tgt, params.sampling));

-    // eval the prompt
-    llama_decode(ctx_tgt,       llama_batch_get_one(inp.data(), inp.size() - 1));
-    llama_decode(ctx_dft.get(), llama_batch_get_one(inp.data(), inp.size() - 1));
+    // init the speculator
+    const auto & params_spec = params.speculative;
+
+    struct common_speculative * spec = common_speculative_init(params.speculative, 1);
+
+    if (spec == nullptr) {
+        LOG_ERR("%s", "failed to initialize speculative decoding\n");
+        return 1;
+    }
+
+    // eval the prompt on the target and feed it to the speculative implementation(s)
+    {
+        llama_batch batch_prompt = llama_batch_init(inp.size(), 0, 1);
+        for (size_t i = 0; i < inp.size() - 1; ++i) {
+            common_batch_add(batch_prompt, inp[i], i, { seq_id }, false);
+        }
+
+        llama_decode(ctx_tgt, batch_prompt);
+
+        if (!common_speculative_process(spec, batch_prompt)) {
+            LOG_ERR("%s", "failed to process speculative prompt\n");
+            return 1;
+        }
+    }

    // note: keep the last token separate!
    llama_token id_last = inp.back();
@@ -142,18 +147,12 @@ int main(int argc, char ** argv) {

    int n_past = inp.size() - 1;

-    // init the speculator
-    const auto & params_spec = params.speculative;
-
-    struct common_speculative * spec = common_speculative_init(params.speculative, 1);
-
    common_speculative_begin(spec, seq_id, prompt_tgt);

    llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);

-    size_t n_draft = 0;
-
    llama_tokens draft;
+
    common_prompt_checkpoint ckpt;

    const auto t_enc_end = ggml_time_us();
@@ -175,13 +174,20 @@ int main(int argc, char ** argv) {
                    llama_memory_seq_pos_max(llama_get_memory(ctx_tgt), seq_id));

            if (use_ckpt_dft) {
-                ckpt.update_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                ckpt.update_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
            }

+            // determine the max draft that fits the remaining context and generation budget
+            int n_draft_max = (int) llama_n_ctx(ctx_tgt) - n_past - 2;
+            if (params.n_predict >= 0) {
+                n_draft_max = std::min(n_draft_max, params.n_predict - n_predict - 1);
+            }
+            n_draft_max = std::max(n_draft_max, 0);
+
            // generate a new draft
            common_speculative_get_draft_params(spec, seq_id) = {
                /* .drafting   = */ true,
-                /* .n_max      = */ -1,
+                /* .n_max      = */ n_draft_max,
                /* .n_past     = */ n_past,
                /* .id_last    = */ id_last,
                /* .prompt     = */ &prompt_tgt,
@@ -189,9 +195,6 @@ int main(int argc, char ** argv) {
            };
            common_speculative_draft(spec);

-            // save the original draft size
-            n_draft = draft.size();
-
            // save a checkpoint of the target context before evaluating the draft
            // this allows us to restore the state if partial draft acceptance occurs
            if (!draft.empty()) {
@@ -200,10 +203,13 @@ int main(int argc, char ** argv) {
                }
            }

-            {
-                ckpt.load_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            // reset the draft context to the checkpoint before verification
+            if (ctx_dft) {
+                if (use_ckpt_dft) {
+                    ckpt.load_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                }

-                llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, ckpt.pos_max + 1, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, ckpt.pos_max + 1, -1);
            }
        } else {
            // we have a previous (partial) draft to reuse from checkpoint restoration
@@ -227,10 +233,10 @@ int main(int argc, char ** argv) {
            llama_decode(ctx_tgt, batch_tgt);
        }

-        // evaluate the same batch with the draft model
-        {
-            // TODO: extend to support MTP, Eagle, etc. See server code for reference
-            llama_decode(ctx_dft.get(), batch_tgt);
+        // feed the batch to the speculative implementation(s) - this drives the draft model, MTP, Eagle3, etc.
+        if (!common_speculative_process(spec, batch_tgt)) {
+            LOG_ERR("%s", "failed to process speculative batch\n");
+            break;
        }

        // only save the sampler sampler state if we use checkpoints
@@ -239,6 +245,9 @@ int main(int argc, char ** argv) {
            smpl_save.reset(common_sampler_clone(smpl.get()));
        }

+        // save the size of the draft being verified
+        const size_t n_draft = draft.size();
+
        // sample from the full target batch and return the accepted tokens based on the target sampler
        //
        // for each token to be accepted, the sampler would have to sample that same token
@@ -255,8 +264,8 @@ int main(int argc, char ** argv) {
        // check for partial draft acceptance:
        // if the context doesn't support partial sequence removal, restore the checkpoint
        // and make the accepted tokens the new partial draft for the next iteration
-        if (use_ckpt_tgt && ids.size() - 1 < draft.size()) {
-            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, draft.size());
+        if (use_ckpt_tgt && ids.size() - 1 < n_draft) {
+            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, n_draft);

            draft = std::move(ids);

@@ -266,10 +275,10 @@ int main(int argc, char ** argv) {
                llama_memory_seq_rm(llama_get_memory(ctx_tgt), seq_id, ckpt.pos_max + 1, -1);
            }

-            {
-                ckpt.load_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            if (ctx_dft) {
+                ckpt.load_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);

-                llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, ckpt.pos_max + 1, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, ckpt.pos_max + 1, -1);
            }

            prompt_tgt.resize(ckpt.n_tokens);
@@ -320,8 +329,11 @@ int main(int argc, char ** argv) {
        {
            LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);

-            llama_memory_seq_rm(llama_get_memory(ctx_tgt),       seq_id, n_past, -1);
-            llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), seq_id, n_past, -1);
+
+            if (ctx_dft) {
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, n_past, -1);
+            }
        }

        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
@@ -347,6 +359,7 @@ int main(int argc, char ** argv) {

    LOG_INF("\n");
    LOG_INF("draft:\n\n");
+    common_speculative_print_stats(spec);

    LOG_INF("\n");
    LOG_INF("target:\n\n");
@@ -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;
    }
@@ -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/")
@@ -154,6 +154,8 @@ extern "C" {
        bool buffer_from_host_ptr;
        // event synchronization
        bool events;
+        // mmap is supported for loading
+        bool mmap_support;
    };

    // all the device properties
@@ -132,6 +132,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ false, // Not implemented.
        /* .buffer_from_host_ptr  = */ false, // Not implemented.
        /* .events                = */ false, // Not implemented.
+        /* .mmap_support          = */ true,
    };
    for (ggml_backend_dev_t simple_dev : meta_dev_ctx->simple_devs) {
        ggml_backend_dev_props tmp_props;
@@ -140,6 +141,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
        props->caps.host_buffer          = props->caps.host_buffer          && tmp_props.caps.host_buffer;
        props->caps.buffer_from_host_ptr = props->caps.buffer_from_host_ptr && tmp_props.caps.buffer_from_host_ptr;
        props->caps.events               = props->caps.events               && tmp_props.caps.events;
+        props->caps.mmap_support         = props->caps.mmap_support         && tmp_props.caps.mmap_support;
    }
 }

@@ -367,6 +367,7 @@ static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct gg
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ true,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -2815,6 +2815,7 @@ static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ host_buffer,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ true,
+        /* .mmap_support          = */ true,
    };
 }

@@ -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) {
@@ -397,6 +397,7 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ true,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -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);
 }

@@ -1865,6 +1865,37 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    ggml_cuda_mul_mat_cublas(ctx, src0, src1, dst);
 }

+// returns true when ggml_cuda_mul_mat_id takes the fallback path that requires stream synchronization
+// [TAG_MUL_MAT_ID_CUDA_GRAPHS]
+static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int cc) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return true;
+    }
+
+    if (dst->ne[2] <= MMVQ_MAX_BATCH_SIZE) {
+        if (ggml_is_quantized(src0->type)) {
+            if (dst->ne[2] <= get_mmvq_mmid_max_batch(src0->type, cc)) {
+                return false;
+            }
+        } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+            return false;
+        }
+    }
+
+    if (ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[2], /*n_experts=*/src0->ne[2])) {
+        return false;
+    }
+
+    if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
+        return false;
+    }
+
+    return true;
+}
+
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
@@ -1907,7 +1938,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
    }

    // note: this path should not be reached when recording CUDA graphs, because it requires stream synchronization
-    // TODO: add asserts to verify this. should work with CUDA, HIP, etc.
+    GGML_ASSERT(ggml_cuda_mul_mat_id_needs_sync(dst, cc));
    cudaStream_t stream = ctx.stream();

    GGML_ASSERT(nb12 % nb11 == 0);
@@ -2522,10 +2553,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
        // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
        if (node->op == GGML_OP_MUL_MAT_ID) {
            const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-            const int mmvq_mmid_max = get_mmvq_mmid_max_batch(node->src[0]->type, cc);
-            if (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > mmvq_mmid_max) {
-                // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
-                // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+            if (ggml_cuda_mul_mat_id_needs_sync(node, cc)) {
+                // the mul_mat_id fallback path synchronizes the stream, so we cannot use CUDA graphs
                // ref: https://github.com/ggml-org/llama.cpp/pull/18958
                use_cuda_graph = false;
 #ifndef NDEBUG
@@ -2651,6 +2680,52 @@ 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 +3055,36 @@ 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 +3093,8 @@ 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 +3946,16 @@ 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;
@@ -4714,6 +4830,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ host_buffer,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ events,
+        /* .mmap_support          = */ props->type != GGML_BACKEND_DEVICE_TYPE_IGPU,
    };
 }

@@ -5098,7 +5215,7 @@ 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;
@@ -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);
@@ -141,6 +141,57 @@ static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, cons
    }
 }

+template <int rows_per_block>
+static __global__ void __launch_bounds__(WARP_SIZE * rows_per_block, 2)
+rwkv_wkv7_f32_t1_warp_row(const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) {
+    constexpr int head_size = CUDA_WKV_BLOCK_SIZE;
+    constexpr int half_head = head_size / 2;
+
+    const int lane = threadIdx.x;
+    const int row  = blockIdx.y * rows_per_block + threadIdx.y;
+    const int bid  = blockIdx.x;
+
+    const int batch_i = bid / H;
+    const int head_i  = bid % H;
+    const int state_size = C * head_size;
+    const int head_off = head_i * head_size;
+    const int t = batch_i * C + head_off + row;
+
+    __shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size];
+
+    if (threadIdx.y == 0) {
+        _r[lane] = r[batch_i * C + head_off + lane];
+        _w[lane] = w[batch_i * C + head_off + lane];
+        _k[lane] = k[batch_i * C + head_off + lane];
+        _a[lane] = a[batch_i * C + head_off + lane];
+        _b[lane] = b[batch_i * C + head_off + lane];
+
+        _r[lane + half_head] = r[batch_i * C + head_off + lane + half_head];
+        _w[lane + half_head] = w[batch_i * C + head_off + lane + half_head];
+        _k[lane + half_head] = k[batch_i * C + head_off + lane + half_head];
+        _a[lane + half_head] = a[batch_i * C + head_off + lane + half_head];
+        _b[lane + half_head] = b[batch_i * C + head_off + lane + half_head];
+    }
+    __syncthreads();
+
+    const int64_t state_base = batch_i * state_size + head_i * head_size * head_size + row * head_size;
+    const float s0 = s[state_base + lane];
+    const float s1 = s[state_base + lane + half_head];
+    const float sa = warp_reduce_sum(_a[lane] * s0 + _a[lane + half_head] * s1);
+
+    const float vt  = v[t];
+    const float st0 = s0 * _w[lane]             + _k[lane]             * vt + sa * _b[lane];
+    const float st1 = s1 * _w[lane + half_head] + _k[lane + half_head] * vt + sa * _b[lane + half_head];
+    const float y   = warp_reduce_sum(st0 * _r[lane] + st1 * _r[lane + half_head]);
+
+    dst[T * C + state_base + lane]             = st0;
+    dst[T * C + state_base + lane + half_head] = st1;
+
+    if (lane == 0) {
+        dst[t] = y;
+    }
+}
+
 void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const float * k_d  = (const float *)dst->src[0]->data;
    const float * v_d  = (const float *)dst->src[1]->data;
@@ -191,7 +242,10 @@ void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
    GGML_ASSERT(C % H == 0);
    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);

-    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+    if (T / B == 1 && C / H == CUDA_WKV_BLOCK_SIZE) {
+        constexpr int rows_per_block = 4;
+        rwkv_wkv7_f32_t1_warp_row<rows_per_block><<<dim3(B * H, CUDA_WKV_BLOCK_SIZE / rows_per_block), dim3(WARP_SIZE, rows_per_block), 0, stream>>>(T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    } else if (C / H == CUDA_WKV_BLOCK_SIZE) {
        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
    } else {
        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
@@ -1646,6 +1646,7 @@ static void ggml_backend_et_device_get_props(ggml_backend_dev_t dev, struct ggml
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -3930,6 +3930,7 @@ static void ggml_backend_hexagon_device_get_props(ggml_backend_dev_t dev, struct
        /* .host_buffer           = */ (bool) opt_hostbuf,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ false,
    };
 }

@@ -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;

@@ -681,6 +681,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
        /* .host_buffer          = */ false,
        /* .buffer_from_host_ptr = */ true,
        /* .events               = */ true,
+        /* .mmap_support         = */ true,
    };
 }

@@ -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;
 }

@@ -7058,6 +7076,19 @@ inline bool enable_adreno_trans_weight(const ggml_backend_opencl_context *backen
    return ((elem_num < 128 * 1024 * 1024) && adreno_kernel && shape_ok);  // max element num: 2**27
 }

+inline bool enable_adreno_trans_weight_q5_K(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+    if (!use_adreno_kernels(backend_ctx, tensor)) {
+        return false;
+    }
+
+    const size_t elem_num = ggml_nelements(tensor);
+    const size_t q_img_width = elem_num / 8;
+    const size_t qh_img_width = elem_num / 16;
+
+    return q_img_width <= backend_ctx->image_max_buffer_size &&
+           qh_img_width <= backend_ctx->image_max_buffer_size;
+}
+
 static inline bool use_flat_gemv_for_large_m_q4_K(const ggml_tensor *tensor) {
    // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
    // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
@@ -9237,7 +9268,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
        cl_kernel kernel = backend_ctx->kernel_convert_block_q5_K;
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (enable_adreno_trans_weight_q5_K(backend_ctx, tensor)) {
            kernel = backend_ctx->kernel_convert_block_q5_K_noshuffle;
        }
 #else
@@ -9272,7 +9303,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,

        tensor->extra = extra;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (enable_adreno_trans_weight_q5_K(backend_ctx, tensor)) {

            int M = tensor->ne[1];
            int K = tensor->ne[0];
@@ -10370,7 +10401,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
            CL_CHECK(clReleaseMemObject(data_device));
            return;
        }
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (enable_adreno_trans_weight_q5_K(backend_ctx, tensor)) {
            int M = tensor->ne[1];
            int K = tensor->ne[0];

@@ -10777,6 +10808,7 @@ static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ false,
    };
 }

@@ -18909,7 +18941,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
        }

        // q5_K x fp32
-        if (src0t == GGML_TYPE_Q5_K && src1t == GGML_TYPE_F32) {
+        if (src0t == GGML_TYPE_Q5_K && src1t == GGML_TYPE_F32 &&
+            enable_adreno_trans_weight_q5_K(backend_ctx, src0)) {
            ggml_cl_mul_mat_q5_K_f32_adreno(backend, src0, src1, dst);
            return;
        }
@@ -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);
@@ -763,6 +763,7 @@ static void ggml_backend_openvino_device_get_props(ggml_backend_dev_t dev, ggml_
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -1881,6 +1881,7 @@ static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggm
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -5649,6 +5649,7 @@ static void ggml_backend_sycl_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ host_buffer,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ events,
+        /* .mmap_support          = */ true,
    };
 }

@@ -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)
@@ -111,6 +111,7 @@ uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_
    apir_encode_bool_t(enc, &props.caps.host_buffer);
    apir_encode_bool_t(enc, &props.caps.buffer_from_host_ptr);
    apir_encode_bool_t(enc, &props.caps.events);
+    apir_encode_bool_t(enc, &props.caps.mmap_support);

    return 0;
 }
@@ -7,7 +7,7 @@
 #include <cstdint>

 #define APIR_PROTOCOL_MAJOR 0
-#define APIR_PROTOCOL_MINOR 1
+#define APIR_PROTOCOL_MINOR 2

 #define APIR_HANDSHAKE_MAGIC 0xab1e

@@ -11,9 +11,9 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml

    context->gpu = gpu;

-    bool async__unused, host_buffer__unused, events__unused;
+    bool async__unused, host_buffer__unused, events__unused, mmap_support__unused;
    bool buffer_from_host_ptr;
-    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused);
+    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused, &mmap_support__unused);

    if (buffer_from_host_ptr) {
        context->apir_context = apir_device_buffer_from_ptr(gpu, size, size);
@@ -65,7 +65,7 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_

    virtgpu * gpu = DEV_TO_GPU(dev);
    apir_device_get_props(gpu, &props->caps.async, &props->caps.host_buffer, &props->caps.buffer_from_host_ptr,
-                          &props->caps.events);
+                          &props->caps.events, &props->caps.mmap_support);

    props->caps.buffer_from_host_ptr = false;
    props->caps.async                = false;
@@ -144,7 +144,8 @@ void apir_device_get_props(virtgpu * gpu,
                           bool *    async,
                           bool *    host_buffer,
                           bool *    buffer_from_host_ptr,
-                           bool *    events) {
+                           bool *    events,
+                           bool *    mmap_support) {
    apir_encoder *        encoder;
    apir_decoder *        decoder;
    ApirForwardReturnCode ret;
@@ -157,6 +158,7 @@ void apir_device_get_props(virtgpu * gpu,
    apir_decode_bool_t(decoder, host_buffer);
    apir_decode_bool_t(decoder, buffer_from_host_ptr);
    apir_decode_bool_t(decoder, events);
+    apir_decode_bool_t(decoder, mmap_support);

    remote_call_finish(gpu, encoder, decoder);

@@ -13,7 +13,8 @@ void                           apir_device_get_props(struct virtgpu * gpu,
                                                     bool *           async,
                                                     bool *           host_buffer,
                                                     bool *           buffer_from_host_ptr,
-                                                     bool *           events);
+                                                     bool *           events,
+                                                     bool *           mmap_support);
 apir_buffer_context_t          apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);

 /* buffer-type */
@@ -4627,6 +4627,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_1], matmul_q5_1_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q8_0], matmul_q8_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q2_K], matmul_q2_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_TQ2_0], matmul_tq2_0_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q3_K], matmul_q3_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K], matmul_q4_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K], matmul_q5_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
@@ -4667,6 +4668,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_subgroup_tq2_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
@@ -4739,6 +4741,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );

        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ2_0], matmul_tq2_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -4783,6 +4786,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
+        CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_subgroup_tq2_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
@@ -4873,6 +4877,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1], matmul_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
+        CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ2_0], matmul_tq2_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
@@ -4921,6 +4926,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
+            CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_subgroup_tq2_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -4968,6 +4974,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_tq2_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -5047,6 +5054,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);

        CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
+        CREATE_MM(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ2_0].f32acc, matmul_tq2_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
        CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
@@ -5094,6 +5102,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_subgroup_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_subgroup_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_subgroup_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
+            CREATE_MM(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0].f32acc, matmul_id_subgroup_tq2_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_subgroup_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_subgroup_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_subgroup_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -5123,6 +5132,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0].f32acc, matmul_id_tq2_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -5226,6 +5236,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[reduc], arr_dmmv_q5_1_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[reduc], arr_dmmv_q8_0_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[reduc16], arr_dmmv_q2_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_TQ2_0][i], "mul_mat_vec_tq2_0_f32_f32", arr_dmmv_tq2_0_f32_f32_len[reduc16], arr_dmmv_tq2_0_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[reduc16], arr_dmmv_q3_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[reduc16], arr_dmmv_q4_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[reduc16], arr_dmmv_q5_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
@@ -5253,6 +5264,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[reduc], arr_dmmv_q5_1_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[reduc], arr_dmmv_q8_0_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[reduc16], arr_dmmv_q2_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_TQ2_0][i], "mul_mat_vec_tq2_0_f16_f32", arr_dmmv_tq2_0_f16_f32_len[reduc16], arr_dmmv_tq2_0_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[reduc16], arr_dmmv_q3_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[reduc16], arr_dmmv_q4_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[reduc16], arr_dmmv_q5_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
@@ -5307,6 +5319,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32",       arr_dmmv_id_q5_1_f32_f32_len[reduc],    arr_dmmv_id_q5_1_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32",       arr_dmmv_id_q8_0_f32_f32_len[reduc],    arr_dmmv_id_q8_0_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32",       arr_dmmv_id_q2_k_f32_f32_len[reduc16],    arr_dmmv_id_q2_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_TQ2_0], "mul_mat_vec_id_tq2_0_f32",     arr_dmmv_id_tq2_0_f32_f32_len[reduc16],   arr_dmmv_id_tq2_0_f32_f32_data[reduc16],   "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32",       arr_dmmv_id_q3_k_f32_f32_len[reduc16],    arr_dmmv_id_q3_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32",       arr_dmmv_id_q4_k_f32_f32_len[reduc16],    arr_dmmv_id_q4_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32",       arr_dmmv_id_q5_k_f32_f32_len[reduc16],    arr_dmmv_id_q5_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
@@ -5368,6 +5381,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_1], "dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_k", dequant_q2_k_len, dequant_q2_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_TQ2_0], "dequant_tq2_0", dequant_tq2_0_len, dequant_tq2_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_k", dequant_q3_k_len, dequant_q3_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_k", dequant_q4_k_len, dequant_q4_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_k", dequant_q5_k_len, dequant_q5_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
@@ -5396,6 +5410,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q2_K], "get_rows_q2_k", get_rows_q2_k_len, get_rows_q2_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_TQ2_0], "get_rows_tq2_0", get_rows_tq2_0_len, get_rows_tq2_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q3_K], "get_rows_q3_k", get_rows_q3_k_len, get_rows_q3_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_K], "get_rows_q4_k", get_rows_q4_k_len, get_rows_q4_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_K], "get_rows_q5_k", get_rows_q5_k_len, get_rows_q5_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -5424,6 +5439,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q2_K], "get_rows_q2_k_f32", get_rows_q2_k_f32_len, get_rows_q2_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_TQ2_0], "get_rows_tq2_0_f32", get_rows_tq2_0_f32_len, get_rows_tq2_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q3_K], "get_rows_q3_k_f32", get_rows_q3_k_f32_len, get_rows_q3_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_K], "get_rows_q4_k_f32", get_rows_q4_k_f32_len, get_rows_q4_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_K], "get_rows_q5_k_f32", get_rows_q5_k_f32_len, get_rows_q5_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -7638,6 +7654,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
        case GGML_TYPE_IQ4_NL:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_TQ2_0:
            break;
        default:
            return nullptr;
@@ -7712,6 +7729,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
        case GGML_TYPE_IQ4_NL:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_TQ2_0:
            break;
        default:
            return nullptr;
@@ -7781,6 +7799,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
        case GGML_TYPE_IQ4_NL:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_TQ2_0:
            break;
        default:
            return nullptr;
@@ -7874,6 +7893,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
        case GGML_TYPE_IQ4_NL:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_TQ2_0:
            break;
        default:
            return nullptr;
@@ -7946,6 +7966,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
        case GGML_TYPE_IQ4_NL:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_TQ2_0:
            break;
        default:
            return nullptr;
@@ -17891,6 +17912,7 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
        /* .host_buffer           = */ true,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ true,
+        /* .mmap_support          = */ !ctx->is_integrated_gpu,
    };
 }

@@ -18013,6 +18035,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                    case GGML_TYPE_IQ4_NL:
                    case GGML_TYPE_MXFP4:
                    case GGML_TYPE_NVFP4:
+                    case GGML_TYPE_TQ2_0:
                        break;
                    default:
                        return false;
@@ -18118,6 +18141,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                    case GGML_TYPE_IQ4_NL:
                    case GGML_TYPE_MXFP4:
                    case GGML_TYPE_NVFP4:
+                    case GGML_TYPE_TQ2_0:
                    case GGML_TYPE_I32:
                        return true;
                    default:
@@ -608,6 +608,20 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif

+#if defined(DATA_A_TQ2_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    // elem e -> byte qs[(e/128)*32 + e%32], bits 2*((e%128)/32); w = q - 1 (d applied via get_dm)
+    const uint qsi   = (iqs / 128) * 32 + (iqs % 32);  // iqs even -> qsi, qsi+1 in same group/level
+    const uint shift = 2 * ((iqs % 128) / 32);
+
+    const uvec2 qs = uvec2(data_a[a_offset + ib].qs[qsi], data_a[a_offset + ib].qs[qsi + 1]);
+    return vec2((qs >> shift) & 3) - 1.0;
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
+#endif
+
 #if defined(DATA_A_Q3_K)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
    iqs /= 2;
@@ -247,6 +247,44 @@ f16vec4 dequantFuncQ8_0_v(const in decodeBufQ8_0 bl, const in uint blockCoords[2
    return f16vec4(vec4(qi) * vec4(float(d)));
 }

+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufTQ2_0 {
+   block_tq2_0 block;
+};
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufTQ2_0_packed16 {
+   block_tq2_0_packed16 block;
+};
+
+float16_t dequantFuncTQ2_0(const in decodeBufTQ2_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    decodeBufTQ2_0_packed16 bl16 = decodeBufTQ2_0_packed16(bl);
+    const uint idx = coordInBlock[1];
+
+    const uint qsshift = (idx & 0x60) >> 4;             // 0,2,4,6
+
+    uint qs = uint32_t(bl16.block.qs[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
+    qs = (qs >> qsshift) & 0x0303;
+    qs = unpack8(qs)[idx & 1];
+
+    return bl.block.d * (float16_t(int(qs)) - float16_t(1.0));
+}
+
+f16vec4 dequantFuncTQ2_0_v(const in decodeBufTQ2_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint idx = coordInBlock[1];
+
+    const uint qsshift = (idx & 0x60) >> 4;             // 0,2,4,6
+    const uint qsi = ((idx & 0x80) >> 2) + (idx & 0x1C); // byte index of 4-aligned group
+
+    const uint qsw = (uint(bl.block.qs[qsi]))
+                   | (uint(bl.block.qs[qsi + 1]) << 8)
+                   | (uint(bl.block.qs[qsi + 2]) << 16)
+                   | (uint(bl.block.qs[qsi + 3]) << 24);
+    const u8vec4 q = unpack8((qsw >> qsshift) & 0x03030303);
+
+    return bl.block.d * (f16vec4(q) - f16vec4(1.0));
+}
+
 layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_K {
   block_q2_K block;
 };
@@ -1368,6 +1406,9 @@ f16vec4 dequantFuncNVFP4_v(const in decodeBufNVFP4 bl, const in uint blockCoords
 #elif defined(DATA_A_Q8_0)
 #define dequantFuncA dequantFuncQ8_0
 #define dequantFuncA_v dequantFuncQ8_0_v
+#elif defined(DATA_A_TQ2_0)
+#define dequantFuncA dequantFuncTQ2_0
+#define dequantFuncA_v dequantFuncTQ2_0_v
 #elif defined(DATA_A_Q2_K)
 #define dequantFuncA dequantFuncQ2_K
 #define dequantFuncA_v dequantFuncQ2_K_v
@@ -0,0 +1,31 @@
+#version 450
+
+#include "dequant_head.glsl"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.nel / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint ip = tid / 32;         // group 0,1 (128 elems each)
+        const uint il = tid - 32 * ip;    // byte in group 0..31
+
+        const uint y_idx = i * QUANT_K + 128 * ip + il;
+
+        const uint8_t qs = data_a[i].qs[32 * ip + il];
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[i].d);
+        data_b[y_idx +  0] = D_TYPE(d * FLOAT_TYPE(int((qs >> 0) & 3) - 1));
+        data_b[y_idx + 32] = D_TYPE(d * FLOAT_TYPE(int((qs >> 2) & 3) - 1));
+        data_b[y_idx + 64] = D_TYPE(d * FLOAT_TYPE(int((qs >> 4) & 3) - 1));
+        data_b[y_idx + 96] = D_TYPE(d * FLOAT_TYPE(int((qs >> 6) & 3) - 1));
+    }
+}
@@ -0,0 +1,102 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.glsl"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+// ternary TQ2_0: w = (q - 1) * d. Same qs group/level layout as q2_K, but a
+// single f16 scale per 256-block and no mins:
+//   sum_e b_e * (q_e - 1) * d  =  d * (sum_e b_e * q_e  -  sum_e b_e)
+void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + y_offset;
+
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
+        if (i >= num_blocks_per_row) {
+            continue;
+        }
+
+        const uint32_t qs_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+        const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303));
+        const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303));
+        const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
+        const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
+            vec2 b16 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  8]);
+            vec2 b32 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
+            vec2 b48 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
+            vec2 b64 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
+            vec2 b80 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
+            vec2 b96 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
+            vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
+
+            FLOAT_TYPE sumq = FLOAT_TYPE(0.0);
+            FLOAT_TYPE sumb = FLOAT_TYPE(0.0);
+            [[unroll]] for (int l = 0; l < 2; ++l) {
+                sumq = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(qs_u32_0[l  ]),
+                       fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(qs_u32_0[l+2]),
+                       fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(qs_u32_2[l  ]),
+                       fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(qs_u32_2[l+2]),
+                       fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(qs_u32_4[l  ]),
+                       fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(qs_u32_4[l+2]),
+                       fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(qs_u32_6[l  ]),
+                       fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(qs_u32_6[l+2]), sumq))))))));
+                sumb += FLOAT_TYPE(b0[l]) + FLOAT_TYPE(b16[l]) + FLOAT_TYPE(b32[l]) + FLOAT_TYPE(b48[l])
+                      + FLOAT_TYPE(b64[l]) + FLOAT_TYPE(b80[l]) + FLOAT_TYPE(b96[l]) + FLOAT_TYPE(b112[l]);
+            }
+            temp[j][n] = fma(d, sumq - sumb, temp[j][n]);
+        }
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint it_size = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid%16;  // 0...15
+    const uint ix = tid/16;
+
+    const uint v_im = itid/8;                                // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - 8*v_im;                         // 0...7
+
+    const uint l0 = 2*v_in;                                  // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 128*v_im + l0;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    for (uint i0 = 0; i0 < num_blocks_per_row; i0 += it_size)
+        calc_superblock(a_offset, b_offset, v_im, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -182,6 +182,22 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin

            buf_a[buf_idx    ] = FLOAT_TYPEV2(v.xy);
            buf_a[buf_idx + 1] = FLOAT_TYPEV2(v.zw);
+#elif defined(DATA_A_TQ2_0)
+            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+
+            const uint ib = idx / 128;                         // 2 values per idx
+            const uint iqs = (idx % 128) * 2;                  // elem 0,2,4..254
+
+            const uint qsi   = (iqs / 128) * 32 + (iqs % 32);  // byte pair start
+            const uint shift = 2 * ((iqs % 128) / 32);         // 0,2,4,6
+
+            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
+            const float d = float(data_a[ib].d);
+
+            const vec2 v = d * (vec2((qs >> shift) & 3) - 1.0);
+
+            buf_a[buf_idx] = FLOAT_TYPEV2(v.xy);
 #elif defined(DATA_A_Q3_K)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -303,6 +303,30 @@ struct block_q2_K_packed32
 #define DATA_A_QUANT_K
 #endif

+#define QUANT_K_TQ2_0 256
+
+// ternary (BitNet): 2-bit codes, w = (q - 1) * d; qs layout matches q2_K's
+// two 32-byte groups with four bit-levels per byte
+struct block_tq2_0
+{
+    uint8_t qs[QUANT_K_TQ2_0/4];
+    float16_t d;
+};
+
+struct block_tq2_0_packed16
+{
+    uint16_t qs[QUANT_K_TQ2_0/4/2];
+    float16_t d;
+};
+
+#if defined(DATA_A_TQ2_0)
+#define QUANT_K QUANT_K_TQ2_0
+#define QUANT_R 1
+#define A_TYPE block_tq2_0
+#define A_TYPE_PACKED16 block_tq2_0_packed16
+#define DATA_A_QUANT_K
+#endif
+
 #define QUANT_K_Q3_K 256

 struct block_q3_K
@@ -72,6 +72,7 @@ const std::vector<std::string> type_names = {
    "iq4_nl",
    "mxfp4",
    "nvfp4",
+    "tq2_0",
    "bf16",
 };

@@ -733,7 +734,7 @@ void process_shaders() {
    for (const auto& tname : type_names) {
        // mul mat vec
        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
+        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "tq2_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";

        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}));
        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}));
@@ -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],
@@ -3954,6 +3942,7 @@ static void ggml_backend_webgpu_device_get_props(ggml_backend_dev_t dev, struct
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -4295,9 +4284,8 @@ 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,
 };
@@ -487,7 +487,8 @@ static void ggml_backend_zdnn_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .async                = */ false,
        /* .host_buffer          = */ false,
        /* .buffer_from_host_ptr = */ false,
-        /* .events               = */ false
+        /* .events               = */ false,
+        /* .mmap_support         = */ true,
    };
 }

--- a/Show More
+++ b/Show More