ui: UI/chat form follow ups (#26743 )

* ui: split the markdown rendering setting per surface User content and thinking get their own toggle again, so turning off markdown for a message leaves reasoning blocks formatted. Both default to markdown. A stored renderContentAsRawText unfolds onto the user key and is dropped from the config. File mentions render as badges in the raw text path too, through a narrow pass over [name](file://path) that leaves everything else untouched. * ui: let the rich chat input scroll past its max height The contenteditable renderer caps its height with max-height but had no overflow rule, so a long buffer overflowed into the input area wrapper and got clipped by its overflow-hidden, leaving no way to reach the bottom of the message. The textarea renderer scrolls natively and was never affected. * ui: apply the new lint and format config * ui: move the render keys unfolding into the migration service Address review from @allozaur: the settings store no longer rewrites persisted config on load, the raw text toggle now unfolds onto the per-surface render keys in migration.service.ts, next to the other config migrations. The mention scanner flag and the directory path suffix become named constants.
ci : don't specify python version in server-sanitize for broader runner compatibility (#26840 )
2026-08-10 11:58:46 +00:00 · 2026-08-10 13:32:51 +02:00 · 2026-08-10 13:32:22 +02:00 · 2026-08-10 13:31:09 +02:00 · 2026-08-10 13:07:27 +02:00 · 2026-08-10 05:20:59 -05:00
691 changed files with 45052 additions and 9139 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 $env:GITHUB_ENV -Append -Encoding utf8
@@ -99,7 +99,6 @@ jobs:
        run: |
          cmake -B build -S . \
            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
            -DGPU_TARGETS="gfx1030" \
            -DGGML_HIP=ON
          cmake --build build --config Release -j $(nproc)
@@ -150,7 +150,6 @@ jobs:
            -DLLAMA_BUILD_BORINGSSL=ON `
            -DROCM_DIR="${env:HIP_PATH}" `
            -DGGML_HIP=ON `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
            -DGPU_TARGETS="gfx1100"  `
            -DGGML_RPC=ON
          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
@@ -15,6 +15,12 @@ on:
      '**/*.cpp'
    ]

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

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

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

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

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

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

@@ -0,0 +1,23 @@
+name: Convert PR to draft
+
+on:
+  pull_request_target:
+    types: [labeled]
+
+permissions:
+  pull-requests: write
+  issues: write
+  contents: write # required for "gh pr ready" command, see https://github.com/cli/cli/issues/8910
+
+jobs:
+  convert-to-draft:
+    if: github.event.label.name == 'draft' && github.event.pull_request.draft == false
+    runs-on: ubuntu-slim
+    steps:
+      - name: Convert PR to draft
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_URL: ${{ github.event.pull_request.html_url }}
+        run: |
+          gh pr ready --undo "$PR_URL"
+          gh pr edit "$PR_URL" --remove-label draft
@@ -848,6 +848,7 @@ jobs:
          name: llama-bin-win-${{ 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 +859,16 @@ 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 +886,7 @@ jobs:
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -885,54 +896,62 @@ 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]
@@ -1229,7 +1248,6 @@ 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 $(nproc)
@@ -1353,7 +1371,6 @@ jobs:
            -DGGML_NATIVE=OFF `
            -DGGML_CPU=OFF `
            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
            -DGGML_HIP=ON `
            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} `
            -DLLAMA_BUILD_BORINGSSL=ON
@@ -1681,6 +1698,7 @@ jobs:
            - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ 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)
@@ -25,6 +25,12 @@ on:
      'tools/server/**.*'
    ]

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

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

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

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

 </div>

@@ -10,6 +10,9 @@
 # # with CUDA support
 # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with ROCm support
+# GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 # # with SYCL support
 # GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
@@ -89,7 +92,7 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
 fi

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

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

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

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

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

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

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

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

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

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

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

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

+    // Muse Glimmer format using " to=<recipient>" recipients and <|eom|>/<|eot|> message terminators.
+    if (src.find("<atem:function_calls>") != std::string::npos && src.find("<|eom|>") != std::string::npos) {
+        LOG_DBG("Using specialized template: Muse Glimmer\n");
+        return common_chat_params_init_muse_glimmer(tmpl, params);
+    }
+
    // Functionary v3.2 - uses recipient-based format with >>>recipient\n{content}
    // Detection: template has ">>>all" for content and ">>>" prefix for tool calls
    if (src.find(">>>all") != std::string::npos && src.find(">>>${recipient}") != std::string::npos) {
@@ -655,6 +655,7 @@ 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
@@ -1032,7 +1032,14 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            return true;
        }

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

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


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

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


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

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


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

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


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

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

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

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

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

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

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

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

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

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

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

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

 ## Environment Variable

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

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

@@ -15,7 +15,7 @@ Legend:
 | Operation | BLAS | CANN | CPU | CUDA | ET | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|------|------|------|
 |                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -41,9 +41,9 @@ Legend:
 |                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -59,7 +59,7 @@ Legend:
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -68,7 +68,7 @@ Legend:
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -47,6 +47,7 @@ CMD_ARGS+=("../../convert_hf_to_gguf.py" "--verbose")
 CMD_ARGS+=("${MODEL_PATH}")
 CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
 CMD_ARGS+=("--outtype" "${TYPE}")
+CMD_ARGS+=("--model-name" "${MODEL_NAME}")
 [[ -n "$METADATA_OVERRIDE" ]] && CMD_ARGS+=("--metadata" "${METADATA_OVERRIDE}")
 [[ -n "$MMPROJ" ]] && CMD_ARGS+=("${MMPROJ}")

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

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

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

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

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

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

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


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

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

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

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

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

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

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

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

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

 endlocal

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

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

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

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

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

 endlocal

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

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

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

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

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

        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-        has_fp16_va = !!(hwcap & HWCAP_FPHP);
+        has_fp16    = !!(hwcap & HWCAP_FPHP) && !!(hwcap & HWCAP_ASIMDHP);
        has_sve     = !!(hwcap & HWCAP_SVE);
        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
@@ -75,7 +91,7 @@ static int ggml_backend_cpu_aarch64_score() {
    score += 1<<1;
 #endif
 #ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
-    if (!af.has_fp16_va) { return 0; }
+    if (!af.has_fp16) { return 0; }
    score += 1<<2;
 #endif
 #ifdef GGML_USE_SVE
@@ -195,6 +195,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -214,6 +215,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -253,9 +253,9 @@ static void ggml_cpy_f32_q8_0_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
+
+void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rms_norm, ggml_tensor * mul, ggml_tensor * rope, ggml_tensor * set_rows);
@@ -3816,7 +3816,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
    }

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

    const size_t smem = pipeline.smem;

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

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

    for (short i = tiitg; i < NK*DK16; i += NTG) {
        const short ik  = i/DK16;
@@ -1022,9 +1022,20 @@ static T block_reduce(T val, T * shared_vals, int block_size_template) {
 }

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

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

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

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

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

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

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

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

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

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

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

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


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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 #define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256

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

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

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

    bool pipeline_executable_properties_support {};

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

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

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

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

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

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

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

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

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

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

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

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

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

        device->idx = idx;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

        // Reset and move submitted events
        for (auto& event : vkev->events_submitted) {
@@ -3221,17 +3221,17 @@ 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 +3263,18 @@ 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 +3305,16 @@ 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],
@@ -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));
 }
@@ -7,32 +7,18 @@ enable chromium_experimental_subgroup_matrix;
 #define BYTE_HELPERS
 #include "common_decls.tmpl"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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


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

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

@@ -382,7 +382,7 @@ class TensorNameMap:
        ),

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

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

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

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

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

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

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

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

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

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

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

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

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

        MODEL_TENSOR.V_MM_POST_NORM: (
@@ -1,8 +1,9 @@
 {#- Iteration on laguna_glm_thinking_v8/chat_template.jinja -#}
 {#- No formatting instructions -#}
 {{- "〈|EOS|〉" -}}
-{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set enable_thinking = enable_thinking | default(true) -%}
 {%- set add_generation_prompt = add_generation_prompt | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}

 {#- ───── header (system message) ───── -#}
 {#- A caller-supplied system message with empty content opts out of the default below, producing no <system> block — used to train without a system message. -#}
@@ -51,7 +52,7 @@
        {%- set reasoning_content = message.reasoning_content -%}
      {%- endif -%}
      {#- Display reasoning content for all messages if enable_thinking -#}
-      {%- if enable_thinking -%}
+      {%- if enable_thinking or preserve_thinking -%}
        {{- '<think>' + reasoning_content + '</think>' -}}
      {%- else -%}
        {{- '</think>' -}}
@@ -1 +1 @@
-90951f99af1fbebef3fbdd58ff5b8715b0bb9c43
+30bf8685ed4eb0a47f2b06229543327749904150
@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess

-HTTPLIB_VERSION = "refs/tags/v0.52.0"
+HTTPLIB_VERSION = "refs/tags/v0.53.0"

 vendor = {
    "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
@@ -21,34 +21,13 @@ vendor = {
    f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/split.py":  "split.py",
    f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/LICENSE":   "vendor/cpp-httplib/LICENSE",

-    "https://raw.githubusercontent.com/sheredom/subprocess.h/8671cee1fc09f11a70ce3782a0ee13177c3aa387/subprocess.h": "vendor/sheredom/subprocess.h",
+    "https://raw.githubusercontent.com/sheredom/subprocess.h/9ce0d701b6fb10f8f8c4445edd31e7c60a1237e3/subprocess.h": "vendor/sheredom/subprocess.h",
 }

-# TODO @ngxson : this is temporary, to be removed in the future
-patches = [
-    # https://github.com/sheredom/subprocess.h/pull/102
-    "vendor/sheredom/patch-bsd.patch",
-    # https://github.com/sheredom/subprocess.h/pull/101
-    "vendor/sheredom/patch-windows-quote-backslash.patch",
-    # https://github.com/sheredom/subprocess.h/pull/104
-    # note: must be applied after patch-bsd.patch, they touch adjacent lines
-    "vendor/sheredom/patch-glibc-older-than-2.29.patch",
-]
-
 for url, filename in vendor.items():
    print(f"downloading {url} to {filename}") # noqa: NP100
    urllib.request.urlretrieve(url, filename)

-for patch in patches:
-    print(f"applying {patch}") # noqa: NP100
-    try:
-        subprocess.check_call([
-            "git", "apply", "--directory", os.path.dirname(patch), patch
-        ])
-    except Exception as e:
-        print(f"Error: {e}") # noqa: NP100
-        sys.exit(1)
-
 print("Splitting httplib.h...") # noqa: NP100
 try:
    subprocess.check_call([
@@ -71,6 +71,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_OLMO,             "olmo"             },
    { LLM_ARCH_OLMO2,            "olmo2"            },
    { LLM_ARCH_OLMOE,            "olmoe"            },
+    { LLM_ARCH_MUSE_GLIMMER,     "muse-glimmer"     },
    { LLM_ARCH_OPENELM,          "openelm"          },
    { LLM_ARCH_ARCTIC,           "arctic"           },
    { LLM_ARCH_DEEPSEEK,         "deepseek"         },
@@ -100,6 +101,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_GRANITE,          "granite"          },
    { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
    { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
+    { LLM_ARCH_GRANITE_SWITCH,   "graniteswitch"    },
    { LLM_ARCH_CHAMELEON,        "chameleon"        },
    { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
    { LLM_ARCH_PLM,              "plm"              },
@@ -220,6 +222,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
    { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
    { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
+    { LLM_KV_ADAPTER_COUNT,                     "%s.adapters.count"                    },
+    { LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE,        "%s.adapters.token_ids_activate"       },
+    { LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE,      "%s.adapters.token_ids_substitute"     },
+    { LLM_KV_ADAPTER_LORA_RANK,                 "%s.adapters.lora_rank"                },
+    { LLM_KV_ADAPTER_ROUTER_GAIN,               "%s.adapters.router_gain"              },
    { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
    { LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         "%s.interleave_moe_layer_step"         },
    { LLM_KV_FULL_ATTENTION_INTERVAL,           "%s.full_attention_interval"           },
@@ -76,6 +76,7 @@ enum llm_arch {
    LLM_ARCH_OLMO,
    LLM_ARCH_OLMO2,
    LLM_ARCH_OLMOE,
+    LLM_ARCH_MUSE_GLIMMER,
    LLM_ARCH_OPENELM,
    LLM_ARCH_ARCTIC,
    LLM_ARCH_DEEPSEEK,
@@ -105,6 +106,7 @@ enum llm_arch {
    LLM_ARCH_GRANITE,
    LLM_ARCH_GRANITE_MOE,
    LLM_ARCH_GRANITE_HYBRID,
+    LLM_ARCH_GRANITE_SWITCH,
    LLM_ARCH_CHAMELEON,
    LLM_ARCH_WAVTOKENIZER_DEC,
    LLM_ARCH_PLM,
@@ -225,6 +227,11 @@ enum llm_kv {
    LLM_KV_TIME_DECAY_EXTRA_DIM,
    LLM_KV_RESIDUAL_SCALE,
    LLM_KV_EMBEDDING_SCALE,
+    LLM_KV_ADAPTER_COUNT,
+    LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE,
+    LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE,
+    LLM_KV_ADAPTER_LORA_RANK,
+    LLM_KV_ADAPTER_ROUTER_GAIN,
    LLM_KV_TOKEN_SHIFT_COUNT,
    LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
    LLM_KV_FULL_ATTENTION_INTERVAL,
@@ -3602,8 +3602,9 @@ llama_context * llama_init_from_model(
                       model->hparams.pooling_type, params.pooling_type);
    }

+    // router_layer >= 0 means n_layer_nextn is repurposed for a router layer, not real MTP
    if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP &&
-        model->hparams.n_layer_nextn == 0) {
+        (model->hparams.n_layer_nextn == 0 || model->hparams.router_layer >= 0)) {
        LLAMA_LOG_WARN("%s: context type MTP requested but model doesn't contain MTP layers\n", __func__);
        return nullptr;
    }
@@ -277,6 +277,16 @@ bool llama_hparams::has_kv(uint32_t il) const {
    return true;
 }

+bool llama_hparams::has_rope(uint32_t il) const {
+    // the router layer stores adapter routing signal, not positional info,
+    // so it must not be RoPE-shifted
+    if (router_layer >= 0 && (int32_t) il == router_layer) {
+        return false;
+    }
+
+    return true;
+}
+
 uint32_t llama_hparams::n_layer() const {
    return n_layer_all - n_layer_nextn;
 }
@@ -53,6 +53,10 @@ struct llama_hparams {
    uint32_t n_embd;
    uint32_t n_layer_all;
    uint32_t n_layer_nextn = 0;
+
+    // granite-switch: index of the single-head "router" KV layer that encodes
+    // per-token adapter selection. -1 when the model has no such layer.
+    int32_t  router_layer = -1;
    uint32_t n_expert = 0;
    uint32_t n_expert_used = 0;
    uint32_t n_rel_attn_bkts = 0;
@@ -371,6 +375,8 @@ struct llama_hparams {

    bool has_kv(uint32_t il) const;

+    bool has_rope(uint32_t il) const;
+
    // number of effective layers (excludes nextn layers)
    uint32_t n_layer() const;

@@ -1931,6 +1931,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
    for (const auto & layer : layers) {
        const uint32_t il = layer.il;

+        if (!hparams.has_rope(il)) {
+            continue;
+        }
+
        const int64_t n_head_kv    = hparams.n_head_kv(il);
        const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);

@@ -937,10 +937,11 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
            } break;
        case GGML_OP_MUL_MAT_ID:
            {
-                const int n_expert_used = hparams.n_expert_used;
-                GGML_ASSERT(n_expert_used > 0);
-                ggml_tensor * b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_expert_used, 512);
-                ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_expert_used, 512);
+                // Used for either MoE expert routing or embedded adapter routing
+                const int n_ids_used = hparams.router_layer >= 0 ? 1 : hparams.n_expert_used;
+                GGML_ASSERT(n_ids_used > 0);
+                ggml_tensor * b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_ids_used, 512);
+                ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_ids_used, 512);
                op_tensor = ggml_mul_mat_id(ctx, w, b, ids);
            } break;
        case GGML_OP_ADD:
@@ -1123,15 +1124,14 @@ struct ggml_tensor * llama_model_loader::create_tensor(
            return nullptr;
        }

-        // tensors with "bias" suffix are always used with GGML_OP_ADD or GGML_OP_ADD_ID
+        // tensors with "bias" suffix are always used with GGML_OP_ADD or GGML_OP_ADD_ID;
+        // embedded-adapter ".lora_a"/".lora_b" tensors are always used with GGML_OP_MUL_MAT_ID
        ggml_op op;
-        bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0;
-        if (bias) {
-            if (info.op == GGML_OP_MUL_MAT_ID) {
-                op = GGML_OP_ADD_ID;
-            } else {
-                op = GGML_OP_ADD;
-            }
+        if (tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0) {
+            op = info.op == GGML_OP_MUL_MAT_ID ? GGML_OP_ADD_ID : GGML_OP_ADD;
+        } else if (hparams.router_layer >= 0 && tn.suffix != nullptr &&
+                (strcmp(tn.suffix, "lora_a") == 0 || strcmp(tn.suffix, "lora_b") == 0)) {
+            op = GGML_OP_MUL_MAT_ID;
        } else {
            op = info.op;
        }
@@ -1249,7 +1249,13 @@ struct ggml_tensor * llama_model_loader::create_tensor(
        for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) {
            t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1;
            GGML_ASSERT(t_meta.ne[dim] >= 1);
-            t_meta.nb[dim] = dim == 0 ? ggml_type_size(type) : t_meta.ne[dim-1]*t_meta.nb[dim-1];
+            if (dim == 0) {
+                t_meta.nb[dim] = ggml_type_size(type);
+            } else if (dim == 1) {
+                t_meta.nb[dim] = ggml_row_size(type, t_meta.ne[dim-1]);
+            } else {
+                t_meta.nb[dim] = t_meta.nb[dim-1]*t_meta.ne[dim-1];
+            }
            GGML_ASSERT(t_meta.nb[dim] >= 1);
        }
        ggml_set_name(&t_meta, tn.str().c_str());
@@ -1272,10 +1278,18 @@ struct ggml_tensor * llama_model_loader::create_tensor(
    if (flags & TENSOR_ALLOW_RESHAPE) {
        for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) {
            t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1;
-            t_meta.nb[dim] = dim == 0 ? ggml_type_size(t_meta.type) : t_meta.ne[dim-1]*t_meta.nb[dim-1];
+            if (dim == 0) {
+                t_meta.nb[dim] = ggml_type_size(t_meta.type);
+            } else if (dim == 1) {
+                t_meta.nb[dim] = ggml_row_size(t_meta.type, t_meta.ne[dim-1]);
+            } else {
+                t_meta.nb[dim] = t_meta.ne[dim-1]*t_meta.nb[dim-1];
+            }
        }
    }

+    GGML_ASSERT(ggml_nbytes(&t_meta) == ggml_nbytes(cur));
+
    ggml_backend_buffer_type_t buft = buft_for_tensor(&t_meta);
    if (buft == nullptr) {
        return nullptr;
@@ -27,6 +27,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
        case LLM_ARCH_APERTUS:
        case LLM_ARCH_MIMO2:
        case LLM_ARCH_STEP35:
+        case LLM_ARCH_MUSE_GLIMMER:
        case LLM_ARCH_MELLUM:
        case LLM_ARCH_LAGUNA:
            return false;
@@ -213,7 +214,7 @@ void llama_model_saver::add_kv_from_model() {
    add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);
    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp);
-    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_chexp);
+    add_kv(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  hparams.n_ff_chexp);
    add_kv(LLM_KV_SWIGLU_CLAMP_EXP,                  hparams.swiglu_clamp_exp);
    add_kv(LLM_KV_SWIGLU_CLAMP_SHEXP,                hparams.swiglu_clamp_shexp);
    add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);
@@ -40,6 +40,8 @@

 static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params & params) {
    switch (arch) {
+        case LLM_ARCH_CLIP:
+            return new llama_model_clip(params);
        case LLM_ARCH_LLAMA:
            return new llama_model_llama(params);
        case LLM_ARCH_LLAMA4:
@@ -174,6 +176,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_olmo2(params);
        case LLM_ARCH_OLMOE:
            return new llama_model_olmoe(params);
+        case LLM_ARCH_MUSE_GLIMMER:
+            return new llama_model_muse_glimmer(params);
        case LLM_ARCH_OPENELM:
            return new llama_model_openelm(params);
        case LLM_ARCH_GPTNEOX:
@@ -234,6 +238,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_granite(params);
        case LLM_ARCH_GRANITE_MOE:
            return new llama_model_granite_moe(params);
+        case LLM_ARCH_GRANITE_SWITCH:
+            return new llama_model_granite_switch(params);
        case LLM_ARCH_MINICPM:
            return new llama_model_minicpm(params);
        case LLM_ARCH_GRANITE_HYBRID:
@@ -1912,6 +1918,7 @@ void llama_model::print_info() const {
                arch == LLM_ARCH_GRANITE ||
                arch == LLM_ARCH_GRANITE_MOE ||
                arch == LLM_ARCH_GRANITE_HYBRID ||
+                arch == LLM_ARCH_GRANITE_SWITCH ||
                arch == LLM_ARCH_NEMOTRON_H_MOE) {
            LLAMA_LOG_INFO("%s: f_embedding_scale     = %f\n", __func__, hparams.f_embedding_scale);
            LLAMA_LOG_INFO("%s: f_residual_scale      = %f\n", __func__, hparams.f_residual_scale);
@@ -2228,6 +2235,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                    params.ctx_type == LLAMA_CONTEXT_TYPE_MTP &&
                    (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE);

+                const bool mtp_on_hybrid_nemotron =
+                    params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && arch == LLM_ARCH_NEMOTRON_H_MOE;
+
                if (llm_arch_is_recurrent(arch)) {
                    res = new llama_memory_recurrent(
                            *this,
@@ -2238,7 +2248,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                            cparams.n_seq_max,
                            cparams.n_rs_seq,
                            nullptr);
-                } else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen) {
+                } else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen && !mtp_on_hybrid_nemotron) {
                    // The main difference between hybrid architectures is the
                    // layer filters, so pick the right one here
                    llama_memory_hybrid::layer_filter_cb filter_attn = nullptr;
@@ -2319,7 +2329,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                        };
                    }

-                    if (mtp_on_hybrid_qwen) {
+                    if (mtp_on_hybrid_qwen || mtp_on_hybrid_nemotron) {
                        filter = [&](uint32_t il) { return il >= hparams.n_layer(); };
                    }

@@ -2591,11 +2601,13 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_DEEPSEEK2OCR:
        case LLM_ARCH_DEEPSEEK32:
        case LLM_ARCH_DEEPSEEK4:
+        case LLM_ARCH_MUSE_GLIMMER:
        case LLM_ARCH_PLM:
        case LLM_ARCH_CHATGLM:
        case LLM_ARCH_GRANITE:
        case LLM_ARCH_GRANITE_MOE:
        case LLM_ARCH_GRANITE_HYBRID:
+        case LLM_ARCH_GRANITE_SWITCH:
        case LLM_ARCH_CHAMELEON:
        case LLM_ARCH_BAILINGMOE:
        case LLM_ARCH_NEO_BERT:
@@ -223,6 +223,24 @@ struct llama_layer_nextn {
    struct ggml_tensor * shared_head_norm      = nullptr;
 };

+struct llama_layer_switch_lora {
+    struct ggml_tensor * a_q    = nullptr;
+    struct ggml_tensor * b_q    = nullptr;
+    struct ggml_tensor * a_k    = nullptr;
+    struct ggml_tensor * b_k    = nullptr;
+    struct ggml_tensor * a_v    = nullptr;
+    struct ggml_tensor * b_v    = nullptr;
+    struct ggml_tensor * a_o    = nullptr;
+    struct ggml_tensor * b_o    = nullptr;
+
+    struct ggml_tensor * a_gate = nullptr;
+    struct ggml_tensor * b_gate = nullptr;
+    struct ggml_tensor * a_up   = nullptr;
+    struct ggml_tensor * b_up   = nullptr;
+    struct ggml_tensor * a_down = nullptr;
+    struct ggml_tensor * b_down = nullptr;
+};
+
 struct llama_layer {
    // normalization
    struct ggml_tensor * attn_norm       = nullptr;
@@ -533,6 +551,8 @@ struct llama_layer {
    struct llama_layer_shortconv shortconv;

    struct llama_layer_nextn nextn;
+
+    struct llama_layer_switch_lora switch_lora;
 };

 struct llama_device {
@@ -0,0 +1,18 @@
+#include "models.h"
+
+// Stub to allow llama-quantize to open mmproj GGUFs
+
+[[noreturn]]
+void llama_model_clip::load_arch_hparams(llama_model_loader &) {
+    GGML_ABORT("CLIP is a quant-only stub; load_arch_hparams should not be called");
+}
+
+[[noreturn]]
+void llama_model_clip::load_arch_tensors(llama_model_loader &) {
+    GGML_ABORT("CLIP is a quant-only stub; load_arch_tensors should not be called");
+}
+
+[[noreturn]]
+std::unique_ptr<llm_graph_context> llama_model_clip::build_arch_graph(const llm_graph_params &) const {
+    GGML_ABORT("CLIP has no inference graph via llama_model dispatch; runtime lives in tools/mtmd/clip.cpp");
+}
@@ -0,0 +1,426 @@
+#include "models.h"
+
+#include <cmath>
+
+void llama_model_granite_switch::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
+    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, false);
+    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, false);
+    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);
+
+    bool rope_finetuned = true;
+    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+    hparams.rope_finetuned = rope_finetuned;
+
+    switch (hparams.n_layer()) {
+        case 40: type = hparams.n_embd == 4096 ? LLM_TYPE_8B : LLM_TYPE_3B; break;
+        case 64: type = LLM_TYPE_30B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+
+    ml.get_key(LLM_KV_ADAPTER_COUNT,     n_adapters);
+    ml.get_key(LLM_KV_ADAPTER_LORA_RANK, max_lora_rank);
+    ml.get_key(LLM_KV_ADAPTER_ROUTER_GAIN, router_gain, /* required */ false);
+
+    // bound counts that size tensors
+    if (n_adapters > 4096) {
+        throw std::runtime_error(format("graniteswitch: invalid adapter count %u", n_adapters));
+    }
+    if (max_lora_rank > 4096) {
+        throw std::runtime_error(format("graniteswitch: invalid lora rank %u", max_lora_rank));
+    }
+
+    std::vector<llama_token> token_ids;
+    std::vector<llama_token> substitute_ids;
+    ml.get_arr(LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE,   token_ids);
+    ml.get_arr(LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE, substitute_ids);
+
+    if (token_ids.size() != n_adapters || substitute_ids.size() != n_adapters) {
+        throw std::runtime_error(format(
+            "graniteswitch: adapter token id arrays (%zu activate, %zu substitute) do not match adapter count %u",
+            token_ids.size(), substitute_ids.size(), n_adapters));
+    }
+
+    adapter_token_to_slot.clear();
+    adapter_token_to_substitute.clear();
+    for (uint32_t i = 0; i < n_adapters; ++i) {
+        // adapter i -> stacked slot i+1 (slot 0 is the base/zero delta)
+        adapter_token_to_slot[token_ids[i]]       = (int32_t) (i + 1);
+        adapter_token_to_substitute[token_ids[i]] = substitute_ids[i];
+    }
+
+    // extra single-head attention layer at the END (index n_real) holds the router
+    // K/V. reusing n_layer_nextn keeps n_layer() == n_real, so the regular layers
+    // keep their indices and the KV cache shift/defrag skips the router layer.
+    // n_layer_nextn is repurposed here (no MTP): it leaks as 1 into the
+    // llama_model_n_layer_nextn() getter and a re-saved nextn_predict_layers
+    const uint32_t n_real = hparams.n_layer();
+    if (n_real >= LLAMA_MAX_LAYERS) {
+        throw std::runtime_error(format("graniteswitch: block count %u exceeds LLAMA_MAX_LAYERS", n_real));
+    }
+    hparams.router_layer  = (int32_t) n_real;
+    hparams.n_layer_all   = n_real + 1;
+    hparams.n_layer_nextn = 1;
+
+    hparams.n_head_arr[n_real]    = 1;
+    hparams.n_head_kv_arr[n_real] = 1;
+    hparams.n_ff_arr[n_real]      = 0;
+}
+
+void llama_model_granite_switch::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_slots     = (int64_t) n_adapters + 1; // slot 0 = base/zero delta
+    const int64_t n_rank      = (int64_t) max_lora_rank;
+    const int64_t n_embd_q    = n_embd_head_k * n_head;
+    const int64_t n_embd_kv   = n_embd_k_gqa;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // substitute ids index tok_embd rows directly; range-check against n_vocab
+    for (const auto & kv : adapter_token_to_substitute) {
+        const llama_token sub = kv.second;
+        if (sub < 0 || (int64_t) sub >= n_vocab) {
+            throw std::runtime_error(format(
+                "graniteswitch: substitute token id %d out of range [0, %d)", sub, (int) n_vocab));
+        }
+    }
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd_q + 2*n_embd_kv}, 0);
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+        auto & sl = layer.switch_lora;
+
+        sl.a_q = create_tensor(tn(LLM_TENSOR_ATTN_Q, "lora_a", i), {n_embd,  n_rank, n_slots}, 0);
+        sl.b_q = create_tensor(tn(LLM_TENSOR_ATTN_Q, "lora_b", i), {n_rank, n_embd_q, n_slots}, 0);
+        sl.a_k = create_tensor(tn(LLM_TENSOR_ATTN_K, "lora_a", i), {n_embd,  n_rank, n_slots}, 0);
+        sl.b_k = create_tensor(tn(LLM_TENSOR_ATTN_K, "lora_b", i), {n_rank, n_embd_kv, n_slots}, 0);
+        sl.a_v = create_tensor(tn(LLM_TENSOR_ATTN_V, "lora_a", i), {n_embd,  n_rank, n_slots}, 0);
+        sl.b_v = create_tensor(tn(LLM_TENSOR_ATTN_V, "lora_b", i), {n_rank, n_embd_kv, n_slots}, 0);
+
+        sl.a_o = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "lora_a", i), {n_embd_q, n_rank, n_slots}, 0);
+        sl.b_o = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "lora_b", i), {n_rank,   n_embd, n_slots}, 0);
+
+        sl.a_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+        sl.b_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "lora_b", i), {n_rank,  n_ff,  n_slots}, 0);
+        sl.a_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+        sl.b_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "lora_b", i), {n_rank,  n_ff,  n_slots}, 0);
+        sl.a_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "lora_a", i), {  n_ff, n_rank, n_slots}, 0);
+        sl.b_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "lora_b", i), {n_rank, n_embd, n_slots}, 0);
+    }
+}
+
+class llm_graph_input_switch : public llm_graph_input_i {
+public:
+    llm_graph_input_switch(const llama_model_granite_switch & smodel) : smodel(smodel) {}
+    virtual ~llm_graph_input_switch() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * sub_tokens  = nullptr; // I32 [n_tokens] adapter-substituted token ids
+    ggml_tensor * router_ksig = nullptr; // F32 [n_tokens] router K signal (+/-gain)
+    ggml_tensor * router_vval = nullptr; // F32 [n_tokens] router V value (adapter slot / 0)
+    ggml_tensor * router_q    = nullptr; // F32 [n_tokens] router Q value (constant 1.0)
+
+    const llama_model_granite_switch & smodel;
+};
+
+// K dim-0 is +gain for an adapter token, -gain otherwise; the causal softmax then
+// lets a single visible adapter token dominate so the readback recovers its slot.
+void llm_graph_input_switch::set_input(const llama_ubatch * ubatch) {
+    if (!ubatch->token) {
+        return;
+    }
+
+    const int64_t n_tokens = ubatch->n_tokens;
+
+    std::vector<int32_t> sub (n_tokens);
+    std::vector<float>   ksig(n_tokens);
+    std::vector<float>   vval(n_tokens);
+    std::vector<float>   q   (n_tokens, 1.0f);
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        const llama_token tok = ubatch->token[i];
+
+        const auto it = smodel.adapter_token_to_slot.find(tok);
+        if (it != smodel.adapter_token_to_slot.end()) {
+            ksig[i] = +smodel.router_gain;
+            vval[i] = (float) it->second;
+        } else {
+            ksig[i] = -smodel.router_gain;
+            vval[i] = 0.0f;
+        }
+
+        const auto sit = smodel.adapter_token_to_substitute.find(tok);
+        sub[i] = (sit != smodel.adapter_token_to_substitute.end())
+            ? (int32_t) sit->second
+            : (int32_t) tok;
+    }
+
+    ggml_backend_tensor_set(sub_tokens,  sub.data(),  0, n_tokens*ggml_element_size(sub_tokens));
+    ggml_backend_tensor_set(router_ksig, ksig.data(), 0, n_tokens*ggml_element_size(router_ksig));
+    ggml_backend_tensor_set(router_vval, vval.data(), 0, n_tokens*ggml_element_size(router_vval));
+    ggml_backend_tensor_set(router_q,    q.data(),    0, n_tokens*ggml_element_size(router_q));
+}
+
+std::unique_ptr<llm_graph_context> llama_model_granite_switch::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+// per-token switched LoRA delta: B_a*(A_a*x), adapter selected per token via ids.
+// cur: {n_in, n_tokens}, ids: {n_tokens} -> {n_out, n_tokens}
+ggml_tensor * llama_model_granite_switch::graph::build_switched_lora_delta(
+          ggml_tensor * lora_a,
+          ggml_tensor * lora_b,
+          ggml_tensor * cur,
+          ggml_tensor * ids) {
+    const int64_t n_in     = cur->ne[0];
+    const int64_t n_tokens = cur->ne[1];
+
+    ggml_tensor * x    = ggml_reshape_3d(ctx0, cur, n_in, 1, n_tokens);
+    ggml_tensor * ids2 = ggml_reshape_2d(ctx0, ids, 1, n_tokens);
+
+    ggml_tensor * a = ggml_mul_mat_id(ctx0, lora_a, x, ids2); // {max_rank, 1, n_tokens}
+    ggml_tensor * d = ggml_mul_mat_id(ctx0, lora_b, a, ids2); // {n_out,    1, n_tokens}
+
+    return ggml_reshape_2d(ctx0, d, d->ne[0], n_tokens);
+}
+
+ggml_tensor * llama_model_granite_switch::graph::build_switched_lora_mm(
+          ggml_tensor * w,
+          ggml_tensor * lora_a,
+          ggml_tensor * lora_b,
+          ggml_tensor * cur,
+          ggml_tensor * ids) {
+    ggml_tensor * base  = ggml_mul_mat(ctx0, w, cur);
+    ggml_tensor * delta = build_switched_lora_delta(lora_a, lora_b, cur, ids);
+    return ggml_add(ctx0, base, delta);
+}
+
+llama_model_granite_switch::graph::graph(
+    const llama_model & model,
+    const llm_graph_params & params)
+    : llm_graph_context(params) {
+
+    const auto & smodel = static_cast<const llama_model_granite_switch &>(model);
+
+    // TODO: support raw embedding input (multimodal / pre-embedded tokens) when needed
+    GGML_ASSERT(ubatch.token && "granite-switch requires token input");
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
+
+    auto inp_switch = std::make_unique<llm_graph_input_switch>(smodel);
+    inp_switch->sub_tokens  = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    inp_switch->router_ksig = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
+    inp_switch->router_vval = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
+    inp_switch->router_q    = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
+    ggml_set_input(inp_switch->sub_tokens);
+    ggml_set_input(inp_switch->router_ksig);
+    ggml_set_input(inp_switch->router_vval);
+    ggml_set_input(inp_switch->router_q);
+    ggml_tensor * sub_tokens  = inp_switch->sub_tokens;
+    ggml_tensor * router_ksig = inp_switch->router_ksig;
+    ggml_tensor * router_vval = inp_switch->router_vval;
+    ggml_tensor * router_q    = inp_switch->router_q;
+    res->add_input(std::move(inp_switch));
+
+    // embed the substituted ids directly; build_inp_embd would embed the raw tokens
+    ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embd, sub_tokens);
+    if (hparams.f_embedding_scale != 0.0f) {
+        inpL = ggml_scale(ctx0, inpL, hparams.f_embedding_scale);
+    }
+    cb(inpL, "inp_embd", -1);
+
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        inp_pos = build_inp_pos();
+    }
+    auto * inp_attn = build_attn_inp_kv();
+
+    // single causal head at layer R recovers the adapter index in-graph: only dim 0
+    // carries signal (Q[0]=1, K[0]=+/-gain, V[0]=slot/0), the rest is zero-padded.
+    const int R = hparams.router_layer;
+    GGML_ASSERT(R >= 0);
+    auto router_lane = [&](ggml_tensor * sig1d) {
+        ggml_tensor * t = ggml_reshape_3d(ctx0, sig1d, 1, 1, n_tokens);
+        return ggml_pad(ctx0, t, (int) n_embd_head - 1, 0, 0, 0);
+    };
+    ggml_tensor * Qr = router_lane(router_q);
+    ggml_tensor * Kr = router_lane(router_ksig);
+    ggml_tensor * Vr = router_lane(router_vval);
+
+    ggml_tensor * router_out = build_attn(inp_attn,
+            nullptr, nullptr, nullptr,
+            Qr, Kr, Vr, nullptr, nullptr, nullptr, /*kq_scale=*/1.0f, /*il=*/R);
+    cb(router_out, "router_out", R);
+
+    // row 0 of router_out is the attended slot; clamp+round to an I32 index
+    ggml_tensor * slot_f = ggml_cont(ctx0,
+        ggml_view_2d(ctx0, router_out, 1, n_tokens, router_out->nb[1], 0));
+    slot_f = ggml_reshape_1d(ctx0, slot_f, n_tokens);
+    slot_f = ggml_clamp(ctx0, slot_f, 0.0f, (float) smodel.n_adapters);
+    slot_f = ggml_round(ctx0, slot_f);
+    ggml_tensor * adapter_ids = ggml_cast(ctx0, slot_f, GGML_TYPE_I32);
+    cb(adapter_ids, "adapter_ids", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    ggml_tensor * cur;
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        cur = build_attention_layer(cur, inp_pos, adapter_ids, inp_attn, model, n_embd_head, il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur,   inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            // keep adapter_ids aligned to the kept rows (2D round-trip for get_rows)
+            const int64_t n_out = inp_out_ids->ne[0];
+            adapter_ids = ggml_get_rows(ctx0,
+                ggml_reshape_2d(ctx0, adapter_ids, 1, adapter_ids->ne[0]), inp_out_ids);
+            adapter_ids = ggml_reshape_1d(ctx0, adapter_ids, n_out);
+        }
+
+        cur = build_layer_ffn(cur, inpSA, adapter_ids, model, il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur, model.output_s);
+
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llama_model_granite_switch::graph::build_attention_layer(
+          ggml_tensor             * cur,
+          ggml_tensor             * inp_pos,
+          ggml_tensor             * adapter_ids,
+          llm_graph_input_attn_kv * inp_attn,
+    const llama_model             & model,
+    const int64_t                 n_embd_head,
+    const int                     il) {
+
+    const auto & layer = model.layers[il];
+    const auto & sl    = layer.switch_lora;
+
+    const int64_t n_head    = hparams.n_head(il);
+    const int64_t n_head_kv = hparams.n_head_kv(il);
+
+    ggml_tensor * qkv = ggml_mul_mat(ctx0, layer.wqkv, cur);
+    cb(qkv, "wqkv", il);
+
+    const int64_t n_embd_q  = n_embd_head * n_head;
+    const int64_t n_embd_kv = n_embd_head * n_head_kv;
+
+    // slice fused qkv into Q/K/V, made contiguous so LoRA deltas can be added
+    ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_q,  qkv->ne[1], qkv->nb[1], 0));
+    ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_kv, qkv->ne[1], qkv->nb[1], n_embd_q*ggml_element_size(qkv)));
+    ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_kv, qkv->ne[1], qkv->nb[1], (n_embd_q + n_embd_kv)*ggml_element_size(qkv)));
+
+    Qcur = ggml_add(ctx0, Qcur, build_switched_lora_delta(sl.a_q, sl.b_q, cur, adapter_ids));
+    Kcur = ggml_add(ctx0, Kcur, build_switched_lora_delta(sl.a_k, sl.b_k, cur, adapter_ids));
+    Vcur = ggml_add(ctx0, Vcur, build_switched_lora_delta(sl.a_v, sl.b_v, cur, adapter_ids));
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+    if (hparams.rope_finetuned) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow);
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow);
+    }
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f
+        ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    // wo = nullptr so build_attn returns concatenated heads; o-proj is switched below
+    ggml_tensor * attn = build_attn(inp_attn,
+            nullptr, nullptr, nullptr,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(attn, "attn_pre_o", il);
+
+    cur = build_switched_lora_mm(layer.wo, sl.a_o, sl.b_o, attn, adapter_ids);
+    cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llama_model_granite_switch::graph::build_layer_ffn(
+          ggml_tensor       * cur,
+          ggml_tensor       * inpSA,
+          ggml_tensor       * adapter_ids,
+    const llama_model       & model,
+    const int                 il) {
+
+    const auto & layer = model.layers[il];
+    const auto & sl    = layer.switch_lora;
+
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il);
+    cb(cur, "ffn_norm", il);
+
+    ggml_tensor * g = build_switched_lora_mm(layer.ffn_gate, sl.a_gate, sl.b_gate, cur, adapter_ids);
+    ggml_tensor * u = build_switched_lora_mm(layer.ffn_up,   sl.a_up,   sl.b_up,   cur, adapter_ids);
+    g = ggml_silu(ctx0, g);
+    ggml_tensor * gu = ggml_mul(ctx0, g, u);
+    cur = build_switched_lora_mm(layer.ffn_down, sl.a_down, sl.b_down, gu, adapter_ids);
+    cb(cur, "ffn_out", il);
+
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}
@@ -386,6 +386,22 @@ struct llama_model_bloom : public llama_model_base {
 };


+// Quant-only stub for mmproj GGUFs
+// none of these are ever called, they only exist to satisfy the llama_model_base interface
+struct llama_model_clip : public llama_model_base {
+    llama_model_clip(const struct llama_model_params & params) : llama_model_base(params) {}
+
+    [[noreturn]]
+    void load_arch_hparams(llama_model_loader & ml) override;
+
+    [[noreturn]]
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    [[noreturn]]
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_mpt : public llama_model_base {
    llama_model_mpt(const struct llama_model_params & params) : llama_model_base(params) {}
    void load_arch_hparams(llama_model_loader & ml) override;
@@ -1028,6 +1044,19 @@ struct llama_model_olmoe : public llama_model_base {
 };


+struct llama_model_muse_glimmer : public llama_model_base {
+    llama_model_muse_glimmer(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_openelm : public llama_model_base {
    llama_model_openelm(const struct llama_model_params & params) : llama_model_base(params) {}
    void load_arch_hparams(llama_model_loader & ml) override;
@@ -1461,6 +1490,10 @@ struct llama_model_nemotron_h_moe : public llama_model_nemotron_h {

    using graph = llama_model_nemotron_h::graph;

+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };

@@ -1596,6 +1629,56 @@ struct llama_model_granite_moe : public llama_model_base {
 };


+struct llama_model_granite_switch : public llama_model_base {
+    llama_model_granite_switch(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    uint32_t n_adapters    = 0;
+    uint32_t max_lora_rank = 0;
+    float    router_gain   = 15.0f;
+
+    std::unordered_map<llama_token, int32_t>     adapter_token_to_slot;
+    std::unordered_map<llama_token, llama_token> adapter_token_to_substitute;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+
+    private:
+        ggml_tensor * build_switched_lora_delta(
+                  ggml_tensor * lora_a,
+                  ggml_tensor * lora_b,
+                  ggml_tensor * cur,
+                  ggml_tensor * ids);
+
+        ggml_tensor * build_switched_lora_mm(
+                  ggml_tensor * w,
+                  ggml_tensor * lora_a,
+                  ggml_tensor * lora_b,
+                  ggml_tensor * cur,
+                  ggml_tensor * ids);
+
+        ggml_tensor * build_attention_layer(
+                  ggml_tensor             * cur,
+                  ggml_tensor             * inp_pos,
+                  ggml_tensor             * adapter_ids,
+                  llm_graph_input_attn_kv * inp_attn,
+            const llama_model             & model,
+            const int64_t                 n_embd_head,
+            const int                     il);
+
+        ggml_tensor * build_layer_ffn(
+                  ggml_tensor       * cur,
+                  ggml_tensor       * inpSA,
+                  ggml_tensor       * adapter_ids,
+            const llama_model       & model,
+            const int                 il);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_minicpm : public llama_model_base {
    llama_model_minicpm(const struct llama_model_params & params) : llama_model_base(params) {}
    void load_arch_hparams(llama_model_loader & ml) override;
@@ -0,0 +1,208 @@
+#include "models.h"
+
+void llama_model_muse_glimmer::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
+    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
+
+    hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    uint32_t swa_period = 4;
+    if (ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false)) {
+        hparams.set_swa_pattern(swa_period);
+    } else {
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
+    }
+
+    switch (hparams.n_layer()) {
+        case 52: type = LLM_TYPE_30B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_muse_glimmer::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // Pre/post-attention norms (Muse Glimmer's `weight + 1` applied at conversion time).
+        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        // Q/K/V/O projections.
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // QK-norm. Weights are synthesized at conversion time to absorb `qk_scale_factor`.
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        // Attention output gate: sigmoid(gate) * attn_out before o_proj (same as afmoe).
+        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+
+        // Pre/post-FFN norms (FFN_PRE_NORM is aliased to LLM_TENSOR_FFN_NORM).
+        layer.ffn_norm      = create_tensor(tn(LLM_TENSOR_FFN_NORM,      "weight", i), {n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        // Dense FFN (unlike afmoe, no MoE branches).
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+    }
+}
+
+llama_model_muse_glimmer::graph::graph(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    // Different to f_norm_rms_eps for post-attn / post-FFN norms
+    const float post_norm_eps = 1e-8f;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    inpL = build_norm(inpL, nullptr, nullptr, LLM_NORM_RMS, -1);
+    cb(inpL, "embd_norm", -1);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+    for (int il = 0; il < n_layer; ++il) {
+        // expose per-layer residual for speculative drafts (see LLM_KV_TARGET_LAYERS).
+        res->t_layer_inp[il] = inpL;
+
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        ggml_tensor * inpSA = inpL;
+
+        // RoPE runs on the SWA layers, NoPE on full ones.
+        const bool use_rope = hparams.is_swa(il);
+
+        // pre-attention norm (weight+1 folded at conversion time)
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention: attention output gate around SDPA (afmoe.cpp:147-191)
+        {
+            ggml_tensor * attn_inp = cur;  // save input for gate computation
+
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            // gate = wqkv_gate @ attn_inp (from pre-attn hidden state)
+            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+            cb(gate, "attn_gate_proj", il);
+
+            // QK-norm. attn_q_norm weight was synthesized at conversion to broadcast
+            // qk_scale_factor across head_dim; attn_k_norm is identity (ones).
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur_rope", il);
+            }
+
+            // SDPA. wo is deferred; the gate goes between attn_out and o_proj.
+            cur = build_attn(inp_attn,
+                    NULL, NULL, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            gate = ggml_sigmoid(ctx0, gate);
+            cb(gate, "attn_gate_sig", il);
+            cur = ggml_mul(ctx0, cur, gate);
+            cb(cur, "attn_gated", il);
+
+            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
+            cb(cur, "attn_o_proj", il);
+        }
+
+        cur = ggml_rms_norm(ctx0, cur, post_norm_eps);
+        cur = ggml_mul(ctx0, cur, model.layers[il].attn_post_norm);
+        cb(cur, "attn_post_norm", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // pre-FFN norm
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // SwiGLU dense FFN
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_rms_norm(ctx0, cur, post_norm_eps);
+        cur = ggml_mul(ctx0, cur, model.layers[il].ffn_post_norm);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // final norm
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head, followed by output multiplier
+    cur = build_lora_mm(model.output, cur, model.output_s);
+    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
+
+    // Final logit tanh softcap (from gemma3.cpp).
+    if (hparams.f_final_logit_softcapping) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+std::unique_ptr<llm_graph_context> llama_model_muse_glimmer::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
@@ -1,6 +1,156 @@
 #include "models.h"

 std::unique_ptr<llm_graph_context> llama_model_nemotron_h_moe::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
    return std::make_unique<graph>(*this, params);
 }

+// MTP draft head for Nemotron-H MoE
+llama_model_nemotron_h_moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "NEMOTRON_H_MOE MTP currently supports a single MTP block");
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    const int il = hparams.n_layer();
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj && layer.nextn.enorm && layer.nextn.hnorm);
+    GGML_ASSERT(layer.ffn_gate_inp);
+
+    // token embedding weights
+    ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+    GGML_ASSERT(tok_embd_w != nullptr && "NEMOTRON_H_MOE MTP requires token embeddings");
+
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // attention fills KV over all tokens, but the MoE is position-wise: gather output rows before
+    // it to save FFN compute (unless unmasked embeddings_nextn needs the full-length hidden state)
+    const bool emit_h_nextn    = cparams.embeddings_nextn;
+    const bool crop_before_ffn = inp_out_ids && (!emit_h_nextn || cparams.embeddings_nextn_masked);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    // dense NoPE attention sub-layer (mtp.layers.0)
+    ggml_tensor * inpSA = cur;
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    {
+        auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);
+        const float kq_scale = hparams.f_attention_scale == 0.0f
+            ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        cur = build_attn(inp_attn, layer.wo, layer.wo_b, layer.wo_s,
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        cb(cur, "mtp_attn_out", il);
+    }
+
+    cur = ggml_add(ctx0, cur, inpSA);
+    cb(cur, "mtp_attn_residual", il);
+
+    // gather the output rows here so the MoE FFN below only runs on the positions we keep
+    if (crop_before_ffn) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
+    // MoE FFN sub-layer (mtp.layers.1)
+    ggml_tensor * ffn_residual = cur;
+    cur = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_post_norm", il);
+
+    {
+        ggml_tensor * router_logits = build_lora_mm(layer.ffn_gate_inp, cur);
+        cb(router_logits, "mtp_ffn_moe_logits", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                layer.ffn_gate_inp,
+                layer.ffn_up_exps,
+                nullptr, // no gate
+                layer.ffn_down_exps,
+                layer.ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
+                hparams.expert_weights_scale,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                il,
+                router_logits, nullptr,
+                layer.ffn_up_exps_s,
+                nullptr, // no gate
+                layer.ffn_down_exps_s);
+        cb(moe_out, "mtp_ffn_moe_out", il);
+
+        ggml_tensor * ffn_shexp = build_ffn(cur,
+                layer.ffn_up_shexp,   NULL, layer.ffn_up_shexp_s,
+                NULL,                 NULL, NULL,
+                layer.ffn_down_shexp, NULL, layer.ffn_down_shexp_s,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+        cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+        cur = ggml_add(ctx0, moe_out, ffn_shexp);
+        cb(cur, "mtp_ffn_out", il);
+    }
+
+    cur = ggml_add(ctx0, cur, ffn_residual);
+    cb(cur, "mtp_post_ffn", il);
+
+    // final head norm: the MTP head has its own LayerNorm
+    GGML_ASSERT(layer.nextn.shared_head_norm && "NEMOTRON_H_MOE MTP: missing final head norm");
+    cur = build_norm(cur, layer.nextn.shared_head_norm, nullptr, LLM_NORM, -1);
+
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (!crop_before_ffn && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
+    // LM head
+    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
+    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;
+    GGML_ASSERT(head_w != nullptr && "NEMOTRON_H_MOE MTP requires an output projection");
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
@@ -7,13 +7,18 @@ void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);

+    // NextN/MTP: optional draft head appended as extra trailing block(s)
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");
+
    // A layer is recurrent IFF the n_head_kv value is set to 0 and
-    // the n_ff value is set to 0
-    for (uint32_t i = 0; i < hparams.n_layer(); ++i) {
-        hparams.is_recr_impl[i] = (hparams.n_head_kv(i) == 0 && hparams.n_ff(i) == 0);
+    // the n_ff value is set to 0. Appended MTP blocks are dense (non-recurrent)
+    for (uint32_t i = 0; i < hparams.n_layer_all; ++i) {
+        hparams.is_recr_impl[i] = i < hparams.n_layer() && hparams.n_head_kv(i) == 0 && hparams.n_ff(i) == 0;
    }

    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,     hparams.f_norm_eps); // MTP head final_layernorm

    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
@@ -30,9 +35,13 @@ void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
    }
 }

-void llama_model_nemotron_h::load_arch_tensors(llama_model_loader &) {
+void llama_model_nemotron_h::load_arch_tensors(llama_model_loader & ml) {
    LLAMA_LOAD_LOCALS;

+    const bool mtp_only    = hparams.n_layer_nextn > 0 && ml.get_weight("blk.0.attn_norm.weight") == nullptr;
+    const int  trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
+    const int  mtp_flags   = !ml.load_mtp ? TENSOR_SKIP : 0;
+
    // mamba2 Mixer SSM params
    // NOTE: int64_t for tensor dimensions
    const int64_t d_conv     = hparams.ssm_d_conv;
@@ -60,61 +69,94 @@ void llama_model_nemotron_h::load_arch_tensors(llama_model_loader &) {
        auto & layer = layers[i];

        // all blocks use the attn norm
-        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, trunk_flags);

        if (hparams.is_recr(i)) {
            // ssm layers
-            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, trunk_flags);

-            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, trunk_flags);
            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);

-            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
+            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, trunk_flags);

            // no "weight" suffix for these
-            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
-            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
+            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, trunk_flags);
+            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, trunk_flags);

-            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, trunk_flags);

            // out_proj
-            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, trunk_flags);
        } else if (hparams.n_ff(i) == 0) {
            // attention layers (with optional bias)
            const int64_t n_head_i = hparams.n_head(i);
            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
-            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);
-            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
+            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, trunk_flags);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, trunk_flags);
            layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
        }  else {
            if (n_expert != 0) {
                const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
                const int64_t n_ff_shexp = hparams.n_ff_shexp;

-                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
-                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, trunk_flags);
+                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, trunk_flags);

                // MoE branch
                layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, TENSOR_NOT_REQUIRED);
                layer.ffn_latent_up   = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP,   "weight", i), {moe_n_embd, n_embd}, TENSOR_NOT_REQUIRED);

-                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, 0);
-                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp, n_expert}, 0);
+                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, trunk_flags);
+                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp, n_expert}, trunk_flags);

                // Shared expert branch
-                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
-                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, trunk_flags);
+                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, trunk_flags);

            } else {
                // mlp layers
-                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
-                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
+                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, trunk_flags);
+                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, trunk_flags);
                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
            }
        }
    }
+
+    // NextN/MTP draft head: each predict layer folds an attention sub-layer and a MoE
+    // sub-layer into a single trailing block
+    for (int i = n_layer; i < n_layer_all; ++i) {
+        auto & layer = layers[i];
+
+        const int64_t n_head_i       = hparams.n_head(i);
+        const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
+        const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
+        const int64_t n_ff_exp       = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_shexp     = hparams.n_ff_shexp;
+
+        // NextN input-fusion tensors
+        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), {n_embd}, mtp_flags);
+        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), {n_embd}, mtp_flags);
+        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), {2*n_embd, n_embd}, mtp_flags);
+        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, mtp_flags);
+
+        // attention sub-layer
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, mtp_flags);
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, mtp_flags);
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, mtp_flags);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd}, mtp_flags | TENSOR_NOT_REQUIRED);
+
+        // MoE sub-layer
+        layer.attn_post_norm  = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM,  "weight", i), {n_embd}, mtp_flags);
+        layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, mtp_flags);
+        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, mtp_flags);
+        layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,   "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, mtp_flags);
+        layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,     "weight", i), {moe_n_embd, n_ff_exp,   n_expert}, mtp_flags);
+        layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,  "weight", i), {n_ff_shexp, n_embd}, mtp_flags);
+        layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,    "weight", i), {n_embd, n_ff_shexp}, mtp_flags);
+    }
 }

 std::unique_ptr<llm_graph_context> llama_model_nemotron_h::build_arch_graph(const llm_graph_params & params) const {
@@ -153,7 +195,7 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
            cur = build_ffn_layer(cur, model, il);
        }

-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
        }
@@ -170,6 +212,14 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_

    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);

+    // seed for the MTP/NextN draft head
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (!cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
    cb(cur, "result_norm", -1);
    res->t_embd = cur;

@@ -2584,6 +2584,7 @@ struct test_rms_norm_mul_rope : public test_case {
    const float eps;
    const bool multi_add; // test a sequence of adds feeding into rms_norm
    const bool set_rows;
+    const bool broadcast; // multiply by a 1D [ne0] weight, as model norm weights are
    int mode;

    std::string op_desc(ggml_tensor * t) override {
@@ -2594,12 +2595,12 @@ struct test_rms_norm_mul_rope : public test_case {
    bool run_whole_graph() override { return true; }

    std::string vars() override {
-        return VARS_TO_STR5(ne, eps, multi_add, set_rows, mode);
+        return VARS_TO_STR6(ne, eps, multi_add, set_rows, broadcast, mode);
    }

    test_rms_norm_mul_rope(std::array<int64_t, 4> ne, float eps = 1e-6f, bool multi_add = false,
-                           bool set_rows = false, int mode = GGML_ROPE_TYPE_NORMAL)
-        : ne(ne), eps(eps), multi_add(multi_add), set_rows(set_rows), mode(mode) {}
+                           bool set_rows = false, bool broadcast = false, int mode = GGML_ROPE_TYPE_NORMAL)
+        : ne(ne), eps(eps), multi_add(multi_add), set_rows(set_rows), broadcast(broadcast), mode(mode) {}

    ggml_tensor * build_graph(ggml_context * ctx) override {
        ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], 1);
@@ -2610,7 +2611,9 @@ struct test_rms_norm_mul_rope : public test_case {
            a = ggml_add(ctx, ggml_add(ctx, a, b), c);
        }

-        a = ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b);
+        ggml_tensor * w = broadcast ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ne[0]) : b;
+
+        a = ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), w);

        ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);

@@ -8576,6 +8579,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
            test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {-1,-1,-1,-1}, {1, 0, 2, 3})); // cpy not-contiguous
        }
    }
+    // quant block count not a multiple of the kernel block size
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_Q4_0, {96, 1, 1, 1}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_Q4_0, GGML_TYPE_F32, {96, 1, 1, 1}));
    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_I32, {256, 2, 3, 4}));
    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_I32, {256, 2, 3, 4}, {-1,-1,-1,-1}, {1, 0, 2, 3}));
    test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_F32, {256, 2, 3, 4}));
@@ -8722,6 +8728,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, true));
            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false, true));
        }
+        // row lengths that are not a multiple of 32, for the scalar (33) and float4 (132, 260) paths
+        for (uint32_t n : { 33, 132, 260 }) {
+            for (bool v : { false, true }) {
+                test_cases.emplace_back(new test_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
+                test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
+            }
+        }
    }

    // in-place tests
@@ -8746,16 +8759,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {

    for (auto multi_add : {false, true}) {
        for (auto set_rows : {false, true}) {
-            for (auto rope : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX}) {
-                test_cases.emplace_back(new test_rms_norm_mul_rope({768, 1, 1, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({768, 3, 1, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({768, 3, 5, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 32, 2, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 4, 2, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 32, 50, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 4, 50, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({8192, 2, 2, 1}, 1e-6f, multi_add, set_rows, rope));
-                test_cases.emplace_back(new test_rms_norm_mul_rope({8192, 2, 2, 1}, 1e-6f, multi_add, set_rows, rope));
+            for (auto broadcast : {false, true}) {
+                for (auto rope : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX}) {
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({768, 1, 1, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({768, 3, 1, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({768, 3, 5, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({128, 32, 2, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({128, 4, 2, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({128, 32, 50, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({128, 4, 50, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({8192, 2, 2, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                    test_cases.emplace_back(new test_rms_norm_mul_rope({8192, 2, 2, 1}, 1e-6f, multi_add, set_rows, broadcast, rope));
+                }
            }
        }
    }
@@ -9747,6 +9762,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
    std::vector<std::unique_ptr<test_case>> test_cases;

+    // SWIGLU at a 27B-class FFN width, fused [gate|up] vs split operands
+    // note: same bytes either way, so a backend that indexes them differently shows it here
+    for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+        for (int64_t n_tokens : {512, 2048}) {
+            test_cases.emplace_back(new test_glu(GGML_GLU_OP_SWIGLU, type, { 2*17408, n_tokens, 1, 1 }, 0, false));
+            test_cases.emplace_back(new test_glu_split(GGML_GLU_OP_SWIGLU, type, { 17408, n_tokens, 1, 1 }, 0));
+        }
+    }
+
    // Conv2d: K=CRS=NPQ=4096 matmul performance
    uint32_t                        iwh_idx  = 0;
    uint32_t                        kwh_idx  = 1;
@@ -1668,9 +1668,18 @@ static std::vector<const backend_test_case *> collect_tests_to_run(const std::st
        }
    } else {
        for (const auto & test : BACKEND_TESTS) {
-            if (test.enabled_by_default) {
-                selected.push_back(&test);
+            if (!test.enabled_by_default) {
+                continue;
            }
+#ifdef GGML_USE_HIP
+            // TODO: remove this when https://github.com/ggml-org/llama.cpp/pull/26592 is merged
+            if (test.name == "penalties" || test.name == "set_sampler" ||
+                test.name == "mixed"     || test.name == "top_p") {
+                fprintf(stderr, "Skipping test '%s' on HIP backend (no backend TOP_K support)\n", test.name.c_str());
+                continue;
+            }
+#endif // GGML_USE_HIP
+            selected.push_back(&test);
        }
    }

@@ -63,6 +63,7 @@ static void test_laguna_tool_format(testing & t);
 static void test_laguna_s_analysis(testing & t);
 static void test_laguna_s_reasoning_detection(testing & t);
 static void test_laguna_s_tool_format(testing & t);
+static void test_laguna_s_preserve_reasoning(testing & t);
 static void test_laguna_xs2_analysis(testing & t);
 static void test_laguna_xs2_reasoning_detection(testing & t);
 static void test_laguna_xs2_tool_format(testing & t);
@@ -1451,9 +1452,14 @@ static void test_laguna_s_tool_format(testing & t) {
    analysis.analyze_template(tmpl);
    t.assert_equal("Laguna-S(v8) arg_value_suffix should be '</arg_value>'", "</arg_value>", analysis.tools.arguments.value_suffix);
 }
+static void test_laguna_s_preserve_reasoning(testing & t) {
+    common_chat_template tmpl = load_laguna_s_template(t);
+    t.assert_true("Laguna-S(v8) supports preserving reasoning", tmpl.original_caps().supports_preserve_reasoning);
+}
 static void test_laguna_s_analysis(testing & t) {
    t.test("Laguna-S(v8) reasoning detection", test_laguna_s_reasoning_detection);
    t.test("Laguna-S(v8) tool format", test_laguna_s_tool_format);
+    t.test("Laguna-S(v8) preserve reasoning", test_laguna_s_preserve_reasoning);
 }

 static common_chat_template load_laguna_xs2_template(testing & t) {
@@ -192,7 +192,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
        ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,              10000.0f);
        // SWA pattern: every 5th layer is full attention (matches E2B layer_types)
        ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, uint32_t(5));
-    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35) {
+    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_MUSE_GLIMMER) {
        std::vector<uint32_t> pattern;
        pattern.reserve(n_layer);
        for (uint32_t il = 0; il < n_layer; il++) {
@@ -217,6 +217,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {

    if (moe) {
        ms.add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, n_ff);
+        ms.add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, n_ff / 2);  // distinct from n_ff so a saver key-clobber surfaces on reload
        ms.add_kv(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,  uint32_t(2));
        ms.add_kv(LLM_KV_EXPERT_COUNT,               uint32_t(2));
        ms.add_kv(LLM_KV_EXPERT_USED_COUNT,          uint32_t(1));
@@ -410,6 +411,9 @@ static bool arch_supported(const llm_arch arch) {
    if (arch == LLM_ARCH_GEMMA4 || arch == LLM_ARCH_GEMMA4_ASSISTANT) {
        return false; // FIXME @ngxson
    }
+    if (arch == LLM_ARCH_GRANITE_SWITCH) {
+        return false; // FIXME adapter fixture
+    }
    if (arch == LLM_ARCH_LLAMA_EMBED || arch == LLM_ARCH_GEMMA_EMBEDDING || arch == LLM_ARCH_T5ENCODER) {
        return false; // FIXME Embedding (?) models produce inconsistent results.
    }
@@ -432,11 +436,19 @@ static bool arch_supported(const llm_arch arch) {

    // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MINIMAX_M3) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
        return false;
    }
 #endif // GGML_USE_WEBGPU

+    // FIXME: jamba produces incorrect output (~0.55 NMSE vs CPU) on the HIP
+    // backend on RDNA3.5 (gfx1151); the SSM kernels need investigation.
+#ifdef GGML_USE_HIP
+    if (arch == LLM_ARCH_JAMBA) {
+        return false;
+    }
+#endif // GGML_USE_HIP
+
    return true;
 }

@@ -195,7 +195,7 @@ static const std::vector<std::string> dspark_dflash = {

 struct plan_case {
    const char * name;
-    const std::vector<std::string> & files;
+    const std::vector<std::string> files;
    const char * hf_repo;
    const char * hf_file;
    bool sidecars;        // request mmproj + mtp + dflash + eagle3 + dspark
@@ -1,4 +1,6 @@
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 #include <assert.h>

 #include "mtmd.h"
@@ -62,6 +64,72 @@ int main(void) {
        }
    }

+    // test chunk save/load round-trip
+    for (size_t i = 0; i < n_chunks; i++) {
+        const mtmd_input_chunk * chunk = mtmd_input_chunks_get(chunks, i);
+        assert(chunk != NULL);
+        enum mtmd_input_chunk_type type = mtmd_input_chunk_get_type(chunk);
+
+        // query the required buffer size (out_buf == NULL)
+        size_t expected_len = 0;
+        int32_t rc = mtmd_input_chunk_save(chunk, NULL, 0, &expected_len);
+        printf("    Chunk %zu: save query rc = %d, expected_len = %zu\n", i, rc, expected_len);
+        assert(rc == 0);
+        assert(expected_len > 0);
+
+        // saving into a too-small buffer must fail, not crash
+        char tiny_buf[1];
+        rc = mtmd_input_chunk_save(chunk, tiny_buf, sizeof(tiny_buf), NULL);
+        printf("    Chunk %zu: save into too-small buffer rc = %d (expect non-zero)\n", i, rc);
+        assert(rc != 0);
+
+        // save into a properly-sized buffer
+        char * buf = (char *) malloc(expected_len);
+        assert(buf != NULL);
+        rc = mtmd_input_chunk_save(chunk, buf, expected_len, NULL);
+        assert(rc == 0);
+
+        // loading from a truncated buffer must fail gracefully, not crash
+        if (expected_len > 1) {
+            mtmd_input_chunk * bad = mtmd_input_chunk_load(buf, expected_len - 1);
+            printf("    Chunk %zu: load from truncated buffer = %p (expect NULL)\n", i, (void *) bad);
+            assert(bad == NULL);
+        }
+
+        // load it back
+        mtmd_input_chunk * loaded = mtmd_input_chunk_load(buf, expected_len);
+        assert(loaded != NULL);
+
+        // metadata must match the original chunk
+        assert(mtmd_input_chunk_get_type(loaded) == type);
+        assert(mtmd_input_chunk_get_n_tokens(loaded) == mtmd_input_chunk_get_n_tokens(chunk));
+        assert(mtmd_input_chunk_get_n_pos(loaded) == mtmd_input_chunk_get_n_pos(chunk));
+
+        if (type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            size_t n_tok_orig, n_tok_loaded;
+            const llama_token * tok_orig   = mtmd_input_chunk_get_tokens_text(chunk, &n_tok_orig);
+            const llama_token * tok_loaded = mtmd_input_chunk_get_tokens_text(loaded, &n_tok_loaded);
+            printf("    Chunk %zu: loaded %zu text tokens (orig %zu), first token %d (orig %d)\n",
+                i, n_tok_loaded, n_tok_orig,
+                n_tok_loaded > 0 ? tok_loaded[0] : -1,
+                n_tok_orig   > 0 ? tok_orig[0]   : -1);
+            assert(n_tok_orig == n_tok_loaded);
+            for (size_t j = 0; j < n_tok_orig; j++) {
+                assert(tok_orig[j] == tok_loaded[j]);
+            }
+        } else if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE || type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+            const char * id_orig   = mtmd_input_chunk_get_id(chunk);
+            const char * id_loaded = mtmd_input_chunk_get_id(loaded);
+            printf("    Chunk %zu: loaded id '%s' (orig '%s')\n", i, id_loaded, id_orig);
+            assert(id_orig != NULL && id_loaded != NULL);
+            assert(strcmp(id_orig, id_loaded) == 0);
+        }
+
+        mtmd_input_chunk_free(loaded);
+        free(buf);
+    }
+    printf("Chunk save/load round-trip OK\n");
+
    // Free the chunks
    mtmd_input_chunks_free(chunks);

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