common/peg : suppress incomplete escape sequences (#26780 )

ggml-webgpu: fix CI errors from #25025 and #25262 (#26566 )
* test new flash_attn test * rebase and fix to disable subgrou matrices when max_kv_tile == 0 * delete log output * Add i32 support to cpy and enables the all ops test * restore the non target ci tests * comment out of TODO of build-cpu.yml * fix format
2026-08-14 12:05:09 +02:00 · 2026-08-11 07:10:31 +03:00 · 2026-08-11 07:10:00 +03:00 · 2026-08-11 00:02:25 +05:30 · 2026-08-10 11:09:19 -07:00 · 2026-08-10 19:53:12 +02:00
808 changed files with 62753 additions and 13855 deletions
@@ -57,7 +57,6 @@ COPY --from=web /app/tools/ui/dist tools/ui/dist
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build \
        -DGGML_HIP=ON \
-        -DGGML_HIP_ROCWMMA_FATTN=ON \
        -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
        -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
        -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
@@ -4,6 +4,10 @@ inputs:
  cuda_version:
    description: "CUDA toolkit version"
    required: true
+  cuda_arch:
+    description: "CUDA target architecture"
+    required: false
+    default: "x64"

 runs:
  using: "composite"
@@ -127,3 +131,26 @@ runs:
          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
          echo "CUDA_PATH_V13_3=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+    - name: Install Cuda Toolkit 13.4 for ARM64
+      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'arm64' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          choco install unzip -y
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cccl-windows-x86_64-13.3.4.1.2-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_crt-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_nvcc-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/libnvvm-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_cudart-windows-arm64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/libcublas-windows-arm64-13.7.0.10-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.1.2-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-arm64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-arm64-13.7.0.10-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
@@ -8,8 +8,26 @@ inputs:
 runs:
  using: "composite"
  steps:
-    - name: Setup ROCm
-      uses: ./.github/actions/install-exe
-      with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
-        args: -install
+    - name: Install ROCm with Wheels
+      shell: pwsh
+      run: |
+        $ErrorActionPreference = "Stop"
+        write-host "Setting up Python virtual environment"
+
+        # Create the venv directly at the cache location to avoid relocation issues
+        New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
+        python -m venv C:\TheRock\build\.venv
+        & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+        write-host "Upgrading pip"
+        python -m pip install --upgrade pip
+
+        write-host "Installing ROCm wheels for multi-arch support"
+        # Install ROCm wheels for multi-arch support (this may take several minutes)
+        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
+
+        # Pre-expand the devel tree so it is included in the cache
+        write-host "Initializing ROCm devel tree"
+        rocm-sdk init
+        if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+        write-host "Completed ROCm wheel installation to C:\TheRock\build"
@@ -60,10 +60,10 @@ jobs:
            -DCMAKE_BUILD_RPATH="@loader_path" \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=OFF \
            -DGGML_METAL_SHADER_DEBUG=ON \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1

@@ -126,7 +126,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -177,7 +176,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -211,7 +209,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -256,7 +253,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -123,8 +123,8 @@ jobs:
    runs-on: windows-2022

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -119,6 +119,7 @@ jobs:
        run: |
          source ./vulkan_sdk/setup-env.sh
          cmake -B build \
+            -DGGML_NATIVE=OFF \
            -DGGML_VULKAN=ON
          cmake --build build --config Release -j $(nproc)

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 </div>

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -0,0 +1,26 @@
+# Used to cross-compile ggml-cuda for Windows ARM64 on an x64 Windows host.
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+if ( DEFINED CUDAToolkit_ROOT )
+    file( TO_CMAKE_PATH "${CUDAToolkit_ROOT}" CUDA_ROOT )
+elseif ( DEFINED ENV{CUDA_PATH} )
+    file( TO_CMAKE_PATH "$ENV{CUDA_PATH}" CUDA_ROOT )
+else()
+    message( FATAL_ERROR "Set CUDAToolkit_ROOT or CUDA_PATH to a Windows CUDA Toolkit with ARM64 target libraries" )
+endif()
+
+if ( DEFINED ENV{VCToolsInstallDir} )
+    file( TO_CMAKE_PATH "$ENV{VCToolsInstallDir}" MSVC_TOOLS_ROOT )
+    set( CMAKE_CUDA_HOST_COMPILER "${MSVC_TOOLS_ROOT}/bin/Hostx64/arm64/cl.exe" CACHE FILEPATH "" )
+endif()
+
+set( CMAKE_CUDA_COMPILER "${CUDA_ROOT}/bin/nvcc.exe" CACHE FILEPATH "" )
+set( CMAKE_CUDA_FLAGS_INIT "-target-dir=arm64" )
+
+# FindCUDAToolkit selects lib/x64 from the host architecture on Windows.
+set( CUDA_CUDART              "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
+set( CUDA_cudart_LIBRARY      "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
+set( CUDA_cublas_LIBRARY      "${CUDA_ROOT}/lib/arm64/cublas.lib"   CACHE FILEPATH "" )
+set( CUDA_cublasLt_LIBRARY    "${CUDA_ROOT}/lib/arm64/cublasLt.lib" CACHE FILEPATH "" )
+set( CUDA_cuda_driver_LIBRARY "${CUDA_ROOT}/lib/arm64/cuda.lib"     CACHE FILEPATH "" )
@@ -27,6 +27,7 @@
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
+#include <cmath>
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
@@ -60,6 +61,7 @@ static std::initializer_list<enum llama_example> mmproj_examples = {
    LLAMA_EXAMPLE_MTMD,
    LLAMA_EXAMPLE_SERVER,
    LLAMA_EXAMPLE_CLI,
+    LLAMA_EXAMPLE_TTS,
 };

 static std::string read_file(const std::string & fname) {
@@ -359,7 +361,6 @@ static bool spec_types_is_default(const common_params & params) {
 common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex) {
    common_download_hf_plan plan;
    common_download_hf_plan plan_spec;
-    common_download_hf_plan plan_voc;
    common_download_opts opts;

    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
@@ -412,11 +413,7 @@ common_models_handler common_models_handler_init(const common_params & params, l
        plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts_spec);
    }

-    if (!params.vocoder.model.hf_repo.empty()) {
-        plan_voc = common_download_get_hf_plan(params.vocoder.model, opts);
-    }
-
-    return common_models_handler{plan, plan_spec, plan_voc, opts};
+    return common_models_handler{plan, plan_spec, opts};
 }

 bool common_models_handler_is_preset_repo(const common_models_handler & handler) {
@@ -466,7 +463,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params

    auto & plan      = handler.plan;
    auto & plan_spec = handler.plan_spec;
-    auto & plan_voc  = handler.plan_voc;

    auto opts = handler.opts; // copy
    opts.callback = callback;
@@ -481,7 +477,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
    };
    handle_url(params.model);
    handle_url(params.mmproj);
-    handle_url(params.vocoder.model);
    handle_url(params.speculative.draft.mparams);

    // optionally, if docker repo is set, resolve it
@@ -509,14 +504,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        task.opts       = opts;
        tasks.push_back(task);
    }
-    if (!params.vocoder.model.url.empty()) {
-        common_download_task task;
-        task.url        = params.vocoder.model.url;
-        task.local_path = params.vocoder.model.path;
-        task.opts       = opts;
-        tasks.push_back(task);
-    }
-
    bool had_spec_url = false;
    if (!params.speculative.draft.mparams.url.empty()) {
        common_download_task task;
@@ -630,11 +617,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        had_spec_url = true;
    }

-    // handle vocoder plan (e.g. --hf-repo-v)
-    if (!plan_voc.model_files.empty()) {
-        add_tasks(plan_voc.model_files, plan_voc.primary, params.vocoder.model);
-    }
-
    if (!plan.model_files.empty()) {
        add_tasks(plan.model_files, plan.primary, params.model);
    }
@@ -1360,6 +1342,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        params.n_parallel = -1;     // auto by default
    } else if (ex == LLAMA_EXAMPLE_TOKENIZE) {
        params.parse_special = true; // parse special tokens by default, like the old tokenize tool
+    } else if (ex == LLAMA_EXAMPLE_TTS) {
+        params.out_file = "output.wav";
+        params.sampling.penalty_repeat = 1.05f;
+        params.sampling.penalty_last_n = -1;
    }

    params.use_color = tty_can_use_colors();
@@ -2022,9 +2008,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_sampling());
    add_opt(common_arg(
        {"--repeat-last-n"}, "N",
-        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n),
+        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled)", params.sampling.penalty_last_n),
        [](common_params & params, int value) {
-            if (value < -1) {
+            if (value < 0) {
                throw std::runtime_error(string_format("error: invalid repeat-last-n = %d\n", value));
            }
            params.sampling.penalty_last_n = value;
@@ -2036,7 +2022,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--repeat-penalty"}, "N",
        string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
        [](common_params & params, const std::string & value) {
-            params.sampling.penalty_repeat = std::stof(value);
+            const float penalty_repeat = std::stof(value);
+            if (!std::isfinite(penalty_repeat) ||
+                penalty_repeat <= 0.0f ||
+                !std::isfinite(1.0f/penalty_repeat)) {
+                throw std::runtime_error("error: repeat-penalty must be finite and greater than 0\n");
+            }
+            params.sampling.penalty_repeat = penalty_repeat;
            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
        }
    ).set_sampling());
@@ -2044,14 +2036,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--presence-penalty"}, "N",
        string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
        [](common_params & params, const std::string & value) {
-            params.sampling.penalty_present = std::stof(value);
+            const float penalty_present = std::stof(value);
+            if (!std::isfinite(penalty_present)) {
+                throw std::runtime_error("error: presence-penalty must be finite\n");
+            }
+            params.sampling.penalty_present = penalty_present;
        }
    ).set_sampling());
    add_opt(common_arg(
        {"--frequency-penalty"}, "N",
        string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
        [](common_params & params, const std::string & value) {
-            params.sampling.penalty_freq = std::stof(value);
+            const float penalty_freq = std::stof(value);
+            if (!std::isfinite(penalty_freq)) {
+                throw std::runtime_error("error: frequency-penalty must be finite\n");
+            }
+            params.sampling.penalty_freq = penalty_freq;
        }
    ).set_sampling());
    add_opt(common_arg(
@@ -2081,9 +2081,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_sampling());
    add_opt(common_arg(
        {"--dry-penalty-last-n"}, "N",
-        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n),
+        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable)", params.sampling.dry_penalty_last_n),
        [](common_params & params, int value) {
-            if (value < -1) {
+            if (value < 0) {
                throw std::runtime_error(string_format("error: invalid dry-penalty-last-n = %d\n", value));
            }
            params.sampling.dry_penalty_last_n = value;
@@ -2567,7 +2567,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.mtmd_batch_max_tokens = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
-    if (llama_supports_rpc()) {
+    if (params.is_gen_docs || llama_supports_rpc()) {
        add_opt(common_arg(
            {"--rpc"}, "SERVERS",
            "comma-separated list of RPC servers (host:port)",
@@ -2968,20 +2968,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.model.hf_file = value;
        }
    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_HF_FILE"));
-    add_opt(common_arg(
-        {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
-        "Hugging Face model repository for the vocoder model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.model.hf_repo = value;
-        }
-    ).set_env("LLAMA_ARG_HF_REPO_V"));
-    add_opt(common_arg(
-        {"-hffv", "--hf-file-v"}, "FILE",
-        "Hugging Face model file for the vocoder model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.model.hf_file = value;
-        }
-    ).set_env("LLAMA_ARG_HF_FILE_V"));
    add_opt(common_arg(
        {"-hft", "--hf-token"}, "TOKEN",
        "Hugging Face access token (default: value from HF_TOKEN environment variable)",
@@ -3316,12 +3302,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--tools"}, "TOOL1,TOOL2,...",
        "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
        "specify \"all\" to enable all tools\n"
-        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime, get_info\n"
        "note: for security reasons, this will limit --cors-origins to localhost by default",
        [](common_params & params, const std::string & value) {
            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"
@@ -4257,24 +4254,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    //

    add_opt(common_arg(
-        {"-mv", "--model-vocoder"}, "FNAME",
-        "vocoder model for audio generation (default: unused)",
+        {"--tts-lang"}, "FNAME",
+        "language (ISO 639-1) for audio generation\n"
+        "see tts/README.md for per-model usage notes",
        [](common_params & params, const std::string & value) {
-            params.vocoder.model.path = value;
+            params.tts_lang = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
-     add_opt(common_arg(
-        {"--tts-use-guide-tokens"},
-        "Use guide tokens to improve TTS word recall",
-        [](common_params & params) {
-            params.vocoder.use_guide_tokens = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_TTS}));
    add_opt(common_arg(
        {"--tts-speaker-file"}, "FNAME",
        "speaker file path for audio generation",
        [](common_params & params, const std::string & value) {
-            params.vocoder.speaker_file = value;
+            params.tts_speaker_file = value;
        }
    ).set_examples({LLAMA_EXAMPLE_TTS}));

@@ -4394,16 +4385,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_DEBUG}));

    // presets
-    add_opt(common_arg(
-        {"--tts-oute-default"},
-        string_format("use default OuteTTS models (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
-            params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
-            params.vocoder.model.hf_repo = "ggml-org/WavTokenizer";
-            params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf";
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS}));

    add_opt(common_arg(
        {"--embd-gemma-default"},
@@ -137,7 +137,6 @@ void common_params_add_preset_options(std::vector<common_arg> & args);
 struct common_models_handler {
    common_download_hf_plan plan;
    common_download_hf_plan plan_spec;
-    common_download_hf_plan plan_voc;
    common_download_opts opts;
 };

@@ -2114,6 +2114,11 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);

+    std::optional<json> additional_context;
+    if (is_v4 && has_response_format) {
+        additional_context = json{ { "response_format", inputs.json_schema } };
+    }
+
    const std::string DSML         = "｜DSML｜";
    const std::string THINK_START  = "<think>";
    const std::string THINK_END    = "</think>";
@@ -2125,9 +2130,12 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
    const std::string PARAM_START  = "<" + DSML + "parameter";
    const std::string PARAM_END    = "</" + DSML + "parameter>";
    const std::string GEN_PROMPT   = "<｜Assistant｜>";
+    const std::string TC_SEPARATOR = "\n\n";

-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
+    data.prompt = common_chat_template_direct_apply_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.supports_thinking  = true;
    data.thinking_start_tag = THINK_START;
@@ -2141,9 +2149,16 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
    if (inputs.has_continuation()) {
        const auto & msg = inputs.continue_msg;

-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
+        if (is_v4 && msg.reasoning_content.empty()) {
+            data.generation_prompt = GEN_PROMPT + THINK_END;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += msg.render_content();
+            }
+        } else {
+            data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += THINK_END + msg.render_content();
+            }
        }

        data.prompt += data.generation_prompt;
@@ -2242,7 +2257,9 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha

        if (extract_reasoning && inputs.enable_thinking) {
            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
-            reasoning_with_tc = THINK_START + p.reasoning(p.until_one_of({ FC_START, THINK_END })) + obligatory_tool_calls;
+            reasoning_with_tc = THINK_START +
+                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
+                p.space() + obligatory_tool_calls;
            allow_reasoning_with_tc = true;
        } else if (extract_reasoning) {
            // Thinking disabled but reasoning extraction requested: the generation prompt
@@ -2265,7 +2282,9 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
            return generation_prompt + reasoning + p.content(p.rest()) + end;
        }

-        auto content_before_tools = p.negate(p.literal(THINK_START)) + p.content(p.until(FC_START));
+        auto content_before_tools = p.negate(p.literal(THINK_START)) +
+            p.content(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START })) +
+            p.space();
        return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
            generation_prompt + reasoning + content_before_tools + tool_calls + end;
    });
@@ -3067,6 +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();
@@ -3095,6 +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) {
@@ -998,6 +998,23 @@ bool fs_is_directory(const std::string & path) {
    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
 }

+std::string common_get_env(const std::string & name) {
+    const char * value = std::getenv(name.c_str());
+    return value == nullptr ? "" : value;
+}
+
+void common_set_env(const std::string & name, const std::string & value) {
+#if defined(_WIN32)
+    _putenv_s(name.c_str(), value.c_str());
+#else
+    if (value.empty()) {
+        unsetenv(name.c_str());
+    } else {
+        setenv(name.c_str(), value.c_str(), 1);
+    }
+#endif
+}
+
 std::string fs_get_cache_directory() {
    std::string cache_directory = "";
    auto ensure_trailing_slash = [](std::string p) {
@@ -1285,16 +1302,6 @@ common_init_result::common_init_result(common_params & params, bool model_only)
                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
    }

-    //if (params.sampling.penalty_last_n == -1) {
-    //    LOG_TRC("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    //}
-
-    //if (params.sampling.dry_penalty_last_n == -1) {
-    //    LOG_TRC("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    //}
-
    // init the backend samplers as part of the context creation
    pimpl->samplers.resize(cparams.n_seq_max);
    pimpl->samplers_seq_config.resize(cparams.n_seq_max);
@@ -1462,18 +1469,18 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 common_init_result::~common_init_result() = default;

 std::string common_get_model_endpoint() {
-    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
-    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
-    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
-    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
-    std::string model_endpoint = "https://huggingface.co/";
-    if (endpoint_env) {
-        model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') {
-            model_endpoint += '/';
-        }
+    std::string endpoint = common_get_env("MODEL_ENDPOINT");
+    if (endpoint.empty()) {
+        // the HF_ENDPOINT variable is respected for backward compatibility
+        endpoint = common_get_env("HF_ENDPOINT");
    }
-    return model_endpoint;
+    if (endpoint.empty()) {
+        return "https://huggingface.co/";
+    }
+    if (endpoint.back() != '/') {
+        endpoint += '/';
+    }
+    return endpoint;
 }

 char * common_get_model_or_exit(int argc, char * argv[]) {
@@ -1632,6 +1639,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.n_seq_max         = params.n_parallel;
    cparams.n_rs_seq          = params.speculative.need_n_rs_seq();
    cparams.n_outputs_max     = std::max(params.n_outputs_max, 0);
+    cparams.n_outputs_max_per_seq = std::max(params.n_outputs_max_per_seq, 0);
    cparams.n_batch           = params.n_batch;
    cparams.n_ubatch          = params.n_ubatch;
    cparams.n_threads         = params.cpuparams.n_threads;
@@ -235,14 +235,14 @@ struct common_params_sampling {
    float   temp               = 0.80f;  // <= 0.0 to sample greedily, 0.0 to not output probabilities
    float   dynatemp_range     = 0.00f;  // 0.0 = disabled
    float   dynatemp_exponent  = 1.00f;  // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty)
    float   penalty_repeat     = 1.00f;  // 1.0 = disabled
    float   penalty_freq       = 0.00f;  // 0.0 = disabled
    float   penalty_present    = 0.00f;  // 0.0 = disabled
    float   dry_multiplier     = 0.0f;   // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
    float   dry_base           = 1.75f;  // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
    int32_t dry_allowed_length = 2;      // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = -1;     // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    int32_t dry_penalty_last_n = 64;     // how many tokens to scan for repetitions (0 = disable penalty)
    float   adaptive_target    = -1.0f;  // select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
    float   adaptive_decay     = 0.90f;  // EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
    int32_t mirostat           = 0;      // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
@@ -392,14 +392,6 @@ struct common_params_speculative {
    }
 };

-struct common_params_vocoder {
-    struct common_params_model model;
-
-    std::string speaker_file; // speaker file path
-
-    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy
-};
-
 struct common_params_diffusion {
    int32_t steps         = 128;
    bool    visual_mode   = false;
@@ -455,6 +447,7 @@ struct common_params {
    int32_t n_parallel            =     1; // number of parallel sequences to decode
    int32_t n_sequences           =     1; // number of sequences to decode
    int32_t n_outputs_max         =     0; // max outputs in a batch (0 = n_batch)
+    int32_t n_outputs_max_per_seq =     1; // max outputs per sequence
    int32_t grp_attn_n            =     1; // group-attention factor
    int32_t grp_attn_w            =   512; // group-attention width
    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -497,7 +490,6 @@ struct common_params {

    struct common_params_sampling    sampling;
    struct common_params_speculative speculative;
-    struct common_params_vocoder     vocoder;
    struct common_params_diffusion   diffusion;

    struct common_params_model model;
@@ -664,6 +656,7 @@ struct common_params {

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

    // MCP server configs (Cursor-compatible JSON)
    std::string mcp_servers_config;   // path to JSON file with MCP server definitions
@@ -739,6 +732,12 @@ struct common_params {
    llama_progress_callback load_progress_callback = NULL;
    void *                  load_progress_callback_user_data = NULL;
    bool no_alloc = false; // Don't allocate model buffers
+
+    // TTS params
+    std::string tts_lang = "";
+    std::string tts_speaker_file = "";
+
+    bool is_gen_docs = false; // whether we are running inside llama-gen-docs
 };

 // call once at the start of a program if it uses libcommon
@@ -863,6 +862,15 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat

 bool glob_match(const std::string & pattern, const std::string & str);

+//
+// Environment utils
+//
+
+// portable environment access, an unset variable reads as an empty string
+// and setting an empty value unsets the variable
+std::string common_get_env(const std::string & name);
+void        common_set_env(const std::string & name, const std::string & value);
+
 //
 // Filesystem utils
 //
@@ -136,7 +136,10 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
        devs.push_back(llama_model_get_device(model, i));
    }

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

@@ -482,6 +482,7 @@ caps caps_get(jinja::program & prog) {
            });
        },
        [&](context & ctx) {
+            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
            caps_apply_preserve_reasoning(ctx, true);
        },
        nullptr, // tools_fn
@@ -116,6 +116,8 @@ static llama_sampler_i llama_sampler_llg_i = {
    /* .backend_accept    = */ NULL,
    /* .backend_apply     = */ NULL,
    /* .backend_set_input = */ NULL,
+    /* .backend_reset     = */ NULL,
+    /* .copy_state        = */ NULL,
 };

 static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,
@@ -570,23 +570,34 @@ struct parser_executor {
    }

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

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

    static common_peg_parse_result handle_unicode_escape(common_peg_parse_context & ctx, size_t start, size_t & pos) {
@@ -217,6 +217,8 @@ static struct llama_sampler_i common_reasoning_budget_i = {
    /* .backend_accept    = */ nullptr,
    /* .backend_apply     = */ nullptr,
    /* .backend_set_input = */ nullptr,
+    /* .backend_reset     = */ nullptr,
+    /* .copy_state        = */ nullptr,
 };

 static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
@@ -184,9 +184,21 @@ std::string common_params_sampling::print() const {
    return std::string(result);
 }

-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params) {
+    if (!std::isfinite(params.penalty_repeat) ||
+        params.penalty_repeat <= 0.0f ||
+        !std::isfinite(1.0f/params.penalty_repeat)) {
+        throw std::invalid_argument("penalty_repeat must be finite and greater than 0");
+    }
+    if (!std::isfinite(params.penalty_freq)) {
+        throw std::invalid_argument("penalty_freq must be finite");
+    }
+    if (!std::isfinite(params.penalty_present)) {
+        throw std::invalid_argument("penalty_present must be finite");
+    }
    const llama_vocab * vocab = llama_model_get_vocab(model);
-
    llama_sampler_chain_params lparams = llama_sampler_chain_default_params();

    lparams.no_perf = params.no_perf;
@@ -338,7 +350,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
                        for (const auto & str : params.dry_sequence_breakers) {
                            c_breakers.push_back(str.c_str());
                        }
-                        samplers.push_back(llama_sampler_init_dry(vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        samplers.push_back(llama_sampler_init_dry(vocab, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                    }
                    break;
                case COMMON_SAMPLER_TYPE_TOP_K:
@@ -366,7 +378,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
                    samplers.push_back(llama_sampler_init_infill(vocab));
                    break;
                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    samplers.push_back(llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    samplers.push_back(llama_sampler_init_penalties(llama_vocab_n_tokens(vocab), params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
                    break;
                case COMMON_SAMPLER_TYPE_ADAPTIVE_P:
                    // the `adaptive-p` sampler is like `dist` and `mirostat` in that it selects
@@ -506,6 +518,26 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
    };
 }

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

@@ -37,7 +37,9 @@ struct common_sampler;
 // llama_sampler API overloads

 // note: can mutate params in some cases
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params);

 void common_sampler_free(struct common_sampler * gsmpl);

@@ -45,6 +47,7 @@ void common_sampler_free(struct common_sampler * gsmpl);
 void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool is_generated);
 void                    common_sampler_reset (struct common_sampler * gsmpl);
 struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
+void                    common_sampler_copy  (const struct common_sampler * src, struct common_sampler * dst);

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

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

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

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

+common_speculative_output_limits common_speculative_get_output_limits(
+        int32_t n_batch, int32_t n_parallel, int32_t n_draft) {
+    const int64_t per_seq = 1 + (int64_t) std::max(0, n_draft);
+    const int64_t total   = (int64_t) n_parallel * per_seq;
+
+    return {
+        /* .total   = */ (int32_t) std::min<int64_t>(n_batch, total),
+        /* .per_seq = */ (int32_t) std::min<int64_t>(n_batch, per_seq),
+    };
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -2385,57 +2404,28 @@ common_speculative * common_speculative_init(common_params_speculative & params,
    {
        uint32_t enabled_configs = common_get_enabled_speculative_configs(params.types);

-        bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
-        bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
-        bool has_draft_mtp    = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP))    && params.draft.ctx_dft != nullptr;
-        bool has_draft_dflash = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)) && params.draft.ctx_dft != nullptr;
-        bool has_draft_dspark = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)) && params.draft.ctx_dft != nullptr;
-
-
-
-        bool has_ngram_cache   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_CACHE));
-        bool has_ngram_simple  = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE));
-        bool has_ngram_map_k   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K));
-        bool has_ngram_map_k4v = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V));
-        bool has_ngram_mod     = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MOD));
+        auto add_config_if_enabled = [&](common_speculative_type type, bool available = true) {
+            if (available && (enabled_configs & (1u << type))) {
+                configs.emplace_back(type, params);
+            }
+        };

        // when adding a new type - update here the logic above
        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 11);

        // this list here defines the priority of the speculators
        // the one with highest priority are listed first
-        if (has_ngram_simple) {
-            // This implementation can guess a lot of tokens without any draft model.
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, params));
-        }
-        if (has_ngram_map_k) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, params));
-        }
-        if (has_ngram_map_k4v) {
-            // This implementation can guess tokens with high acceptance rate but is more expensive.
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
-        }
-        if (has_ngram_mod) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, params));
-        }
-        if (has_ngram_cache) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
-        }
-        if (has_draft_simple) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, params));
-        }
-        if (has_draft_eagle3) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
-        }
-        if (has_draft_mtp) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
-        }
-        if (has_draft_dflash) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params));
-        }
-        if (has_draft_dspark) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK, params));
-        }
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MOD);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
+
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_MTP,    params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK, params.draft.ctx_dft != nullptr);
    }

    std::vector<std::unique_ptr<common_speculative_impl>> impls = {};
@@ -25,6 +25,15 @@ int32_t common_speculative_n_max(const common_params_speculative * spec);

 common_params common_base_params_to_speculative(const common_params & params);

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

 void common_speculative_free(common_speculative * spec);
@@ -70,6 +70,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Exaone4ForCausalLM": "exaone",
    "ExaoneForCausalLM": "exaone",
    "ExaoneMoEForCausalLM": "exaone",
+    "ExaoneMoeForCausalLM": "exaone",
    "FalconForCausalLM": "falcon",
    "FalconH1ForCausalLM": "falcon_h1",
    "FalconMambaForCausalLM": "mamba",
@@ -102,6 +103,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteMoeForCausalLM": "granite",
    "GraniteMoeHybridForCausalLM": "granite",
    "GraniteMoeSharedForCausalLM": "granite",
+    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
    "Grok1ForCausalLM": "grok",
@@ -181,6 +183,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Olmo3ForCausalLM": "olmo",
    "OlmoForCausalLM": "olmo",
    "OlmoeForCausalLM": "olmo",
+    "MuseGlimmerAssistantModel": "muse_glimmer",
+    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
    "OpenELMForCausalLM": "openelm",
    "OrionForCausalLM": "orion",
    "PLMForCausalLM": "plm",
@@ -210,6 +214,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3MoeForCausalLM": "qwen",
    "Qwen3NextForCausalLM": "qwen",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
    "Qwen3_5ForCausalLM": "qwen",
@@ -295,6 +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",
@@ -304,6 +310,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "Qwen2_5_VLForConditionalGeneration": "qwenvl",
    "Qwen3ASRForConditionalGeneration": "qwen3vl",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
    "Qwen3_5ForConditionalGeneration": "qwen3vl",
@@ -81,7 +81,7 @@ class ChatGLMModel(TextModel):

    @staticmethod
    def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
+        from transformers.convert_slow_tokenizer import bytes_to_unicode
        byte_encoder = bytes_to_unicode()
        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])

@@ -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
@@ -447,12 +460,43 @@ class DeepseekV2Model(TextModel):
 class DeepseekV32Model(DeepseekV2Model):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK32
    skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)

+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        # DeepSeek V3.2 appends the NextN/MTP block past num_hidden_layers
+        # (model.layers.61 -> blk.61 in the 62-block file).
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
    def set_vocab(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
@@ -463,7 +507,7 @@ class DeepseekV32Model(DeepseekV2Model):
        super().set_gguf_parameters()

        # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)

        # DSA indexer parameters
@@ -489,6 +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)
@@ -504,7 +555,10 @@ class DeepseekV4Model(TextModel):
            logger.info("Skipping %d DeepSeek-V4 MTP tensor(s) for conversion v0", type(self)._skipped_mtp_tensors)

        # add a default chat template; if the model has a built-in template, it will be overridden later
-        template_path = Path(__file__).parent.parent / "models" / "templates" / "deepseek-ai-DeepSeek-V4.jinja"
+        model_id_hint = self.remote_hf_model_id or self.dir_model.name
+        is_0731 = "0731" in model_id_hint
+        template_name = "deepseek-ai-DeepSeek-V4-Flash-0731.jinja" if is_0731 else "deepseek-ai-DeepSeek-V4.jinja"
+        template_path = Path(__file__).parent.parent / "models" / "templates" / template_name
        if template_path.is_file():
            with open(template_path, "r", encoding="utf-8") as f:
                self.gguf_writer.add_chat_template(f.read())
@@ -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

@@ -206,10 +206,70 @@ class Glm4MoeModel(TextModel):
@ModelBase.register("Glm4MoeLiteForCausalLM")
 class Glm4MoeLiteModel(DeepseekV2Model):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+    skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None

    def set_vocab(self):
        return self._set_vocab_glm()

+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        num_hidden_layers = self.hparams["num_hidden_layers"]
+        self.num_nextn_predict_layers = self.hparams.get("num_nextn_predict_layers", 0)
+        self.skip_mtp = self.no_mtp or self.num_nextn_predict_layers == 0
+
+        if self.skip_mtp:
+            self.block_count = num_hidden_layers
+        else:
+            self.block_count = num_hidden_layers + self.num_nextn_predict_layers
+
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        if self.skip_mtp:
+            return
+
+        self.gguf_writer.add_nextn_predict_layers(self.num_nextn_predict_layers)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item):
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        if cls._n_main_layers is not None:
+            match = re.match(r"model\.layers\.(\d+)\.", name)
+            is_mtp = match is not None and int(match.group(1)) >= cls._n_main_layers
+            if is_mtp and cls.no_mtp:
+                return None
+            if cls.mtp_only and not is_mtp and name not in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+            ):
+                return None
+
+        return 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"
+

@ModelBase.register("GlmMoeDsaForCausalLM")
 class GlmMoeDsaModel(DeepseekV2Model):
@@ -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
@@ -119,7 +119,7 @@ class LlamaModel(TextModel):
        path_tekken_json = self.dir_model / "tekken.json"
        path_tokenizer_json = self.dir_model / "tokenizer.json"
        if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
-            self._set_vocab_mistral()
+            return self._set_vocab_mistral()

        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
        if tokenizer_config_file.is_file():
@@ -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
@@ -18,7 +18,7 @@ class QwenModel(TextModel):

    @staticmethod
    def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
+        from transformers.convert_slow_tokenizer import bytes_to_unicode
        byte_encoder = bytes_to_unicode()
        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])

@@ -268,8 +268,101 @@ class Qwen3MoeModel(Qwen2MoeModel):
        super().set_vocab()


+class _QwenMtpMixin:
+    """Shared MTP wiring for Qwen3-Next and Qwen3.5/3.6 text variants. The HF
+    config carries the MTP block under `mtp_num_hidden_layers` (computed from
+    the checkpoint when absent, e.g. Qwen3-Next) and the tensors under
+    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
+    `mtp.*` to the standard layer-indexed nextn naming so the existing
+    tensor_map handles them."""
+
+    supports_mtp_export = True
+    hparams: dict[str, Any]
+    model_arch: gguf.MODEL_ARCH
+    gguf_writer: gguf.GGUFWriter
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+    no_mtp: bool
+    mtp_only: bool
+    _original_block_count: int | None = None
+    opt_num_mtp_layers: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
+            # Qwen-3-Next doesn't include `mtp_num_hidden_layers` in config.
+            if n_mtp == 0:
+                assert self.opt_num_mtp_layers != 0
+                n_mtp = self.opt_num_mtp_layers
+            self.block_count += n_mtp
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._original_block_count = hparams.get(key)
+        type(self).opt_num_mtp_layers = 0
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
+
+    @classmethod
+    def filter_tensors(cls, item):
+        assert cls._original_block_count is not None
+        # TODO: change TextModel to super()
+        if (titem := TextModel.filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith("model.mtp."):
+            name = name.replace("model.", "", 1)
+        if name.startswith("mtp."):
+            if cls.no_mtp:
+                return None
+            remapper = {
+                "fc":                    "eh_proj",
+                "pre_fc_norm_embedding": "enorm",
+                "pre_fc_norm_hidden":    "hnorm",
+                "norm":                  "shared_head.norm",
+            }
+            parts = name.split(".", 3)
+            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
+                mtp_idx = int(parts[2])
+                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
+                cls.opt_num_mtp_layers = max(cls.opt_num_mtp_layers, mtp_idx + 1)
+            elif len(parts) == 3 and parts[1] in remapper:
+                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
+        elif cls.mtp_only:
+            keep = name in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+                "embed_tokens.weight", "norm.weight",
+            )
+            if not keep:
+                return None
+        return name, gen
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
+        if self.no_mtp:
+            return
+        if (n := self.block_count - self.hparams["num_hidden_layers"]) > 0:
+            self.gguf_writer.add_nextn_predict_layers(n)
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+
@ModelBase.register("Qwen3NextForCausalLM")
-class Qwen3NextModel(Qwen2MoeModel):
+class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
    model_arch = gguf.MODEL_ARCH.QWEN3NEXT

    def set_gguf_parameters(self):
@@ -284,16 +377,6 @@ class Qwen3NextModel(Qwen2MoeModel):
            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.25)))

-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-
-        if name.startswith("mtp"):
-            # ignore MTP layers for now
-            return None
-
-        return super().filter_tensors(item)
-
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if name.endswith(".A_log"):
            data_torch = -torch.exp(data_torch)
@@ -536,97 +619,13 @@ class _Qwen35MRopeMixin:
            self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION)


-class _Qwen35MtpMixin:
-    """Shared MTP wiring for Qwen3.5/3.6 text variants. The HF config carries
-    the MTP block under `mtp_num_hidden_layers` and the tensors under
-    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
-    `mtp.*` to the standard layer-indexed nextn naming so the existing
-    tensor_map handles them."""
-
-    supports_mtp_export = True
-    hparams: dict[str, Any]
-    model_arch: gguf.MODEL_ARCH
-    gguf_writer: gguf.GGUFWriter
-    block_count: int
-    tensor_map: gguf.TensorNameMap
-    no_mtp: bool
-    mtp_only: bool
-    _original_block_count: int | None = None
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"]
-        if not self.no_mtp:
-            self.block_count += self.hparams.get("mtp_num_hidden_layers", 0)
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
-        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
-        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
-        type(self)._original_block_count = hparams.get(key)
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
-
-    @classmethod
-    def filter_tensors(cls, item):
-        assert cls._original_block_count is not None
-        # TODO: change TextModel to super()
-        if (titem := TextModel.filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-        if name.startswith("model.mtp."):
-            name = name.replace("model.", "", 1)
-        if name.startswith("mtp."):
-            if cls.no_mtp:
-                return None
-            remapper = {
-                "fc":                    "eh_proj",
-                "pre_fc_norm_embedding": "enorm",
-                "pre_fc_norm_hidden":    "hnorm",
-                "norm":                  "shared_head.norm",
-            }
-            parts = name.split(".", 3)
-            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
-                mtp_idx = int(parts[2])
-                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
-            elif len(parts) == 3 and parts[1] in remapper:
-                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
-        elif cls.mtp_only:
-            keep = name in (
-                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-                "embed_tokens.weight", "norm.weight",
-            )
-            if not keep:
-                return None
-        return name, gen
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
-        if self.no_mtp:
-            return
-        if (n := self.hparams.get("mtp_num_hidden_layers", 0)) > 0:
-            self.gguf_writer.add_nextn_predict_layers(n)
-
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
-
@ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
-class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
    model_arch = gguf.MODEL_ARCH.QWEN35


@ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")
-class Qwen3_5MoeTextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
    model_arch = gguf.MODEL_ARCH.QWEN35MOE


@@ -0,0 +1,471 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, Callable, Iterable, TYPE_CHECKING
+
+import torch
+import torch.nn.functional as F
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, MmprojModel, TextModel, gguf
+
+# Tricks being used to support this model via existing llama.cpp code paths:
+# - Text projection MLP is folded into the embedding table
+# - codec_embedding is concat to the text embedding table, vocab is extended
+#   example: codec_bos_id(2149) --> "<|codec_bos|>"
+#            codec_eos_token_id(2150) --> "<|codec_eos_token|>"
+#            codec_language_id.chinese(2055) --> "<|codec_language_chinese|>"
+#            other rows --> "<|codec_0|>", "<|codec_1|>", ..., "<|codec_1023|>"
+# - output tensor codec_head is smaller than vocab, so logits will be padded at inference time
+# - suppress_tokens is used to limit the backbone to only sample either semantic or EOS (stop) token
+
+# pipeline stage mapping:
+#   speaker reference encoder --> mapped to normal mtmd audio encoder
+#   backbone --> mapped to normal libllama text model (autoregressive)
+#   code_predictor --> MTMD_GEN_PROCESS_TYPE_GEN_CODE
+#   code2wav --> MTMD_GEN_PROCESS_TYPE_GEN_WAV
+
+# torch activation functions used by Qwen3TTSTalkerResizeMLP (config's hidden_act)
+_ACT2FN = {
+    "silu": F.silu,
+    "gelu": F.gelu,
+    "relu": F.relu,
+}
+
+
+@ModelBase.register("Qwen3TTSForConditionalGeneration")
+class Qwen3TTSTalkerModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3TTS
+
+    _TEXT_PROJ_KEYS = (
+        "model.text_embedding.weight",
+        "text_projection.linear_fc1.weight",
+        "text_projection.linear_fc1.bias",
+        "text_projection.linear_fc2.weight",
+        "text_projection.linear_fc2.bias",
+    )
+
+    _text_proj_buffer: dict[str, Tensor]
+    _folded_text_embed: Tensor | None
+    _codec_embed: Tensor | None
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        raw_talker_config = dict(hparams["talker_config"])
+        self._talker_config = raw_talker_config
+        self.n_codec_vocab = raw_talker_config["vocab_size"]
+        talker_config = dict(raw_talker_config)
+        talker_config["vocab_size"] = talker_config["text_vocab_size"]
+        hparams["text_config"] = talker_config
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+        self._text_proj_buffer = {}
+        self._folded_text_embed = None
+        self._codec_embed = None
+
+    def _codec_token_names(self) -> list[str]:
+        # start every row with a generic name, then override the ones with a
+        # known meaning (bos/eos/language/etc, derived from the *_id fields
+        # of talker_config) with a more descriptive one
+        names = [f"<|codec_{i}|>" for i in range(self.n_codec_vocab)]
+        for key, val in self._talker_config.items():
+            if not key.endswith("_id"):
+                continue
+            prefix = key[:-len("_id")]
+            if isinstance(val, int):
+                names[val] = f"<|{prefix}|>"
+            elif isinstance(val, dict):
+                for subkey, subval in val.items():
+                    names[subval] = f"<|{prefix}_{subkey}|>"
+        return names
+
+    def set_vocab(self):
+        codec_tokens = self._codec_token_names()
+        codec_toktypes = [gguf.TokenType.CONTROL] * len(codec_tokens)
+
+        try:
+            tokens, scores, toktypes = self._create_vocab_sentencepiece()
+            self.gguf_writer.add_tokenizer_model("llama")
+            self.gguf_writer.add_tokenizer_pre("default")
+            tokens += [t.encode("utf-8") for t in codec_tokens]
+            scores += [0.0] * len(codec_tokens)
+            toktypes += codec_toktypes
+            self.gguf_writer.add_token_list(tokens)
+            self.gguf_writer.add_token_scores(scores)
+            self.gguf_writer.add_token_types(toktypes)
+            special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+            special_vocab.add_to_gguf(self.gguf_writer)
+            return
+        except FileNotFoundError:
+            pass
+
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        tokens += codec_tokens
+        toktypes += codec_toktypes
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        # make sure that the model has no chat template, so chat will be disabled
+        self.gguf_writer.add_chat_template(None)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # note: final vocab layout is [text_vocab | codec_vocab], with text_vocab is actually padded with -inf in cgraph
+        # for codec_vocab, only first 2048 rows can be sampled for semantic code
+        # plus codec_eos_token_id that used for signaling end of generation
+        # ref: https://github.com/QwenLM/Qwen3-TTS/blob/022e286b98fbec7e1e916cb940cdf532cd9f488e/qwen_tts/core/models/modeling_qwen3_tts.py#L2059-L2063
+
+        vocab_size = self.hparams["vocab_size"] + self.n_codec_vocab
+        codec_eos_token_id = self.hparams["vocab_size"] + self._talker_config["codec_eos_token_id"]
+        self.gguf_writer.add_suppress_tokens([
+            i for i in range(vocab_size - 1024, vocab_size)
+            if i != codec_eos_token_id
+        ])
+        self.gguf_writer.add_eos_token_id(codec_eos_token_id)
+        self.gguf_writer.add_add_eos_token(False)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if not name.startswith("talker.") or name.startswith("talker.code_predictor."):
+            return None
+
+        name = name[len("talker."):]
+        return super().filter_tensors((name, gen))
+
+    def _maybe_emit_token_embd(self) -> Iterable[tuple[str, Tensor]]:
+        if self._folded_text_embed is None or self._codec_embed is None:
+            return
+        combined = torch.cat([self._folded_text_embed, self._codec_embed], dim=0)
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD), combined)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # codec_embedding rows are appended after the text vocab, extending the embedding table
+        if name == "model.codec_embedding.weight":
+            self._codec_embed = data_torch
+            yield from self._maybe_emit_token_embd()
+            return
+
+        # codec_head is the output head for the (smaller) codec vocab; logits get padded to
+        # the extended vocab size at inference time
+        if name == "codec_head.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch)
+            return
+
+        if name in self._TEXT_PROJ_KEYS:
+            self._text_proj_buffer[name] = data_torch
+            if len(self._text_proj_buffer) < len(self._TEXT_PROJ_KEYS):
+                return
+
+            # fold MLP into the embedding table at conversion time, MLP won't be used at inference time anyway
+            act_fn = _ACT2FN[self.hparams["hidden_act"]]
+            embed = self._text_proj_buffer["model.text_embedding.weight"]
+            hidden = act_fn(F.linear(embed,
+                                     self._text_proj_buffer["text_projection.linear_fc1.weight"],
+                                     self._text_proj_buffer["text_projection.linear_fc1.bias"]))
+            folded = F.linear(hidden,
+                              self._text_proj_buffer["text_projection.linear_fc2.weight"],
+                              self._text_proj_buffer["text_projection.linear_fc2.bias"])
+            self._folded_text_embed = folded
+            yield from self._maybe_emit_token_embd()
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3TTSForConditionalGeneration")
+class Qwen3TTSSpeakerEncoderModel(MmprojModel):
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    # talker.code_predictor.model.layers.{bid}.<key> -> A_GEN_CODE_*
+    # bypass tensor_mapping.py for now to make it simple
+    _CODE_LAYER_TENSOR_MAP = {
+        "input_layernorm": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_NORM,
+        "self_attn.q_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_Q,
+        "self_attn.q_norm": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_Q_NORM,
+        "self_attn.k_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_K,
+        "self_attn.k_norm": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_K_NORM,
+        "self_attn.v_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_V,
+        "self_attn.o_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_OUT,
+        "post_attention_layernorm": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_NORM,
+        "mlp.gate_proj": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_GATE,
+        "mlp.up_proj": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_UP,
+        "mlp.down_proj": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_DOWN,
+    }
+
+    # note: codebook pages will be stacked to 3D
+    _CODE_GEN_N_CODEBOOKS = 15
+    _code_embed_buffer: dict[int, Tensor] = {}
+    _code_head_buffer: dict[int, Tensor] = {}
+    _wav_config_cache: dict[str, Any] | None = None
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        hparams["text_config"] = {"hidden_size": hparams["talker_config"]["hidden_size"]}
+        # ECAPA-TDNN has a fixed 4-stage backbone, but MmprojModel.__init__ needs a n_block_keys
+        hparams["speaker_encoder_config"]["n_layers"] = 4
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+        self._wav_config_cache = None
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("speaker_encoder_config")
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_clip_has_audio_encoder(True)
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.QWEN3TTS_SPKENC)
+
+        # handle speaker encoder config
+        self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+        # mel_spectrogram() front-end: sr=24000, n_fft=1024, hop=256, n_mels=128, fmin=0, fmax=12000 (=sr/2, the clip.cpp default)
+        self.gguf_writer.add_audio_num_mel_bins(128)
+        # 3 SE-Res2Net stages; the stem conv, mfa, asp and fc are not counted here
+        self.gguf_writer.add_audio_block_count(3)
+        # ECAPA-TDNN has no attention/FFN, these are dummy to allow clip.cpp to load it
+        self.gguf_writer.add_audio_embedding_length(1536)
+        self.gguf_writer.add_audio_head_count(1)
+        self.gguf_writer.add_audio_feed_forward_length(1536)
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+
+        # handle code predictor config
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.QWEN3TTS_GEN)
+        code_predictor_config = self.global_config["talker_config"]["code_predictor_config"]
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(code_predictor_config["hidden_size"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(code_predictor_config["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(code_predictor_config["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(code_predictor_config["num_attention_heads"])
+        self.gguf_writer.add_gen_audio_head_count_kv(code_predictor_config["num_key_value_heads"])
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(code_predictor_config["rms_norm_eps"])
+        # note: code2wav hparams are hardcoded on the mtmd/clip.cpp side for now, not written here
+
+    def _wav_decoder_config(self) -> dict[str, Any] | None:
+        # code2wav has its own config.json, inside the speech_tokenizer dir
+        if self._wav_config_cache is None:
+            path = self.dir_model / "speech_tokenizer" / "config.json"
+            with open(path, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+            self._wav_config_cache = cfg["decoder_config"]
+        return self._wav_config_cache
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
+        if new_name.endswith(".weight") and (
+            new_name in ("a.gen.wav.pre_conv.weight", "a.gen.wav.dac.entry.weight", "a.gen.wav.dac.post_conv.weight")
+            or (".up.blk." in new_name and new_name.endswith(".dwconv.weight"))
+            or (".dac.blk." in new_name and (new_name.endswith(".conv1.weight") or new_name.endswith(".conv2.weight")))
+        ):
+            return gguf.GGMLQuantizationType.F16
+        # ConvTranspose1d kernels: only F16/F32 are implemented, no BF16
+        if new_name.endswith(".conv.weight") and (".up.blk." in new_name or ".dac.blk." in new_name):
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if not (
+            name.startswith("speaker_encoder.")
+            or name.startswith("talker.code_predictor.")
+            or name == "talker.model.codec_embedding.weight"
+        ):
+            return None
+
+        return super().filter_tensors((name, gen))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # code2wav tensors are already named by generate_extra_tensors(), pass them through
+        if name.startswith("a.gen.wav."):
+            yield (name, data_torch)
+            return
+
+        # codebook-0 embedding, fed back to the talker backbone (codebooks 1-15 live in code_predictor)
+        if name == "talker.model.codec_embedding.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_OUT_EMBD), data_torch)
+            return
+
+        if name == "talker.code_predictor.model.norm.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_OUTPUT_NORM), data_torch)
+            return
+
+        if name.startswith("talker.code_predictor.small_to_mtp_projection."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_PROJ_IN, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("talker.code_predictor.model.codec_embedding."):
+            idx = int(name.split("codec_embedding.")[1].split(".")[0])
+            self._code_embed_buffer[idx] = data_torch
+            if len(self._code_embed_buffer) < self._CODE_GEN_N_CODEBOOKS:
+                return
+            stacked = torch.stack([self._code_embed_buffer.pop(i) for i in range(self._CODE_GEN_N_CODEBOOKS)], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_EMBD), stacked)
+            return
+
+        if name.startswith("talker.code_predictor.lm_head."):
+            idx = int(name.split("lm_head.")[1].split(".")[0])
+            self._code_head_buffer[idx] = data_torch
+            if len(self._code_head_buffer) < self._CODE_GEN_N_CODEBOOKS:
+                return
+            stacked = torch.stack([self._code_head_buffer.pop(i) for i in range(self._CODE_GEN_N_CODEBOOKS)], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_HEAD), stacked)
+            return
+
+        if name.startswith("talker.code_predictor.model.layers."):
+            rest = name.split("model.layers.")[1]        # "{bid}.<key>.weight"
+            _, key_with_suffix = rest.split(".", 1)       # "<key>.weight"
+            key = key_with_suffix.rsplit(".", 1)[0]        # "<key>"
+            tensor = self._CODE_LAYER_TENSOR_MAP.get(key)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid), data_torch)
+                return
+
+        if "res2net_block.blocks." in name:
+            assert bid is not None  # the outer stage index, picked up from the tensor name automatically
+            xid = int(name.split("res2net_block.blocks.")[1].split(".")[0])
+            suffix = "." + name.rsplit(".", 1)[1]
+            new_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_ENC_CONV_RES2].format(bid=bid, xid=xid) + suffix
+            yield (new_name, data_torch)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        yield from self._generate_code2wav_tensors()
+
+    def _generate_code2wav_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # code2wav weights live in speech_tokenizer/model.safetensors, not the main safetensors
+        from safetensors.torch import load_file
+
+        wav_config = self._wav_decoder_config()
+        state_dict = load_file(self.dir_model / "speech_tokenizer" / "model.safetensors")
+
+        def get(name: str) -> Tensor:
+            return state_dict[name]
+
+        def snake_fold(alpha: Tensor, beta: Tensor) -> tuple[Tensor, Tensor]:
+            # fold SnakeBeta's exp()/reciprocal here, so the graph is only mul/sin/sqr/mul/add
+            return torch.exp(alpha), 1.0 / (torch.exp(beta) + 1e-9)
+
+        def rvq_codebook(prefix: str, n_layers: int) -> Tensor:
+            # checkpoint has EMA accumulators, so codebook[i] = embedding_sum[i] / cluster_usage[i]
+            books = []
+            for i in range(n_layers):
+                embedding_sum = get(f"{prefix}.vq.layers.{i}._codebook.embedding_sum")
+                cluster_usage = get(f"{prefix}.vq.layers.{i}._codebook.cluster_usage")
+                books.append(embedding_sum / cluster_usage.clamp_min(1e-5).unsqueeze(-1))
+            return torch.stack(books, dim=0) if n_layers > 1 else books[0]
+
+        T = gguf.MODEL_TENSOR
+
+        # --- quantizer: RVQ codebook decode ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_IN), get("decoder.quantizer.rvq_first.input_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_OUT), get("decoder.quantizer.rvq_first.output_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_CB), rvq_codebook("decoder.quantizer.rvq_first", 1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_IN), get("decoder.quantizer.rvq_rest.input_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_OUT), get("decoder.quantizer.rvq_rest.output_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_CB), rvq_codebook("decoder.quantizer.rvq_rest", self._CODE_GEN_N_CODEBOOKS))
+
+        # --- pre_conv ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_PRE_CONV, suffix=".weight"), get("decoder.pre_conv.conv.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_PRE_CONV, suffix=".bias"), get("decoder.pre_conv.conv.bias"))
+
+        # --- pre_transformer ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_IN_PROJ, suffix=".weight"), get("decoder.pre_transformer.input_proj.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_IN_PROJ, suffix=".bias"), get("decoder.pre_transformer.input_proj.bias"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUT_PROJ, suffix=".weight"), get("decoder.pre_transformer.output_proj.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUT_PROJ, suffix=".bias"), get("decoder.pre_transformer.output_proj.bias"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUTPUT_NORM), get("decoder.pre_transformer.norm.weight"))
+
+        tfm_layer_map = {
+            "input_layernorm.weight":         T.A_GEN_WAV_TFM_ATTN_NORM,
+            "self_attn.q_proj.weight":        T.A_GEN_WAV_TFM_ATTN_Q,
+            "self_attn.k_proj.weight":        T.A_GEN_WAV_TFM_ATTN_K,
+            "self_attn.v_proj.weight":        T.A_GEN_WAV_TFM_ATTN_V,
+            "self_attn.o_proj.weight":        T.A_GEN_WAV_TFM_ATTN_OUT,
+            "self_attn_layer_scale.scale":    T.A_GEN_WAV_TFM_ATTN_SCALE,
+            "post_attention_layernorm.weight": T.A_GEN_WAV_TFM_FFN_NORM,
+            "mlp.gate_proj.weight":           T.A_GEN_WAV_TFM_FFN_GATE,
+            "mlp.up_proj.weight":             T.A_GEN_WAV_TFM_FFN_UP,
+            "mlp.down_proj.weight":           T.A_GEN_WAV_TFM_FFN_DOWN,
+            "mlp_layer_scale.scale":          T.A_GEN_WAV_TFM_FFN_SCALE,
+        }
+        assert wav_config is not None
+        for bid in range(wav_config["num_hidden_layers"]):
+            for key, tensor_id in tfm_layer_map.items():
+                yield (self.format_tensor_name(tensor_id, bid), get(f"decoder.pre_transformer.layers.{bid}.{key}"))
+
+        # --- upsample: 2x (causal ConvTranspose1d + ConvNeXt block) ---
+        up_map = {
+            "0.conv.weight":     (T.A_GEN_WAV_UP_CONV, ".weight"),
+            "0.conv.bias":       (T.A_GEN_WAV_UP_CONV, ".bias"),
+            "1.dwconv.conv.weight": (T.A_GEN_WAV_UP_DWCONV, ".weight"),
+            "1.dwconv.conv.bias":   (T.A_GEN_WAV_UP_DWCONV, ".bias"),
+            "1.norm.weight":     (T.A_GEN_WAV_UP_NORM, ".weight"),
+            "1.norm.bias":       (T.A_GEN_WAV_UP_NORM, ".bias"),
+            "1.pwconv1.weight":  (T.A_GEN_WAV_UP_PW1, ".weight"),
+            "1.pwconv1.bias":    (T.A_GEN_WAV_UP_PW1, ".bias"),
+            "1.pwconv2.weight":  (T.A_GEN_WAV_UP_PW2, ".weight"),
+            "1.pwconv2.bias":    (T.A_GEN_WAV_UP_PW2, ".bias"),
+            "1.gamma":           (T.A_GEN_WAV_UP_GAMMA, ""),
+        }
+        for bid in range(len(wav_config["upsampling_ratios"])):
+            for key, (tensor_id, suffix) in up_map.items():
+                yield (self.format_tensor_name(tensor_id, bid, suffix=suffix), get(f"decoder.upsample.{bid}.{key}"))
+
+        # --- DAC decoder ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_ENTRY, suffix=".weight"), get("decoder.decoder.0.conv.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_ENTRY, suffix=".bias"), get("decoder.decoder.0.conv.bias"))
+
+        n_dac_blocks = len(wav_config["upsample_rates"])
+        for bid in range(n_dac_blocks):
+            py = bid + 1  # decoder.decoder.0 is the entry conv, blocks start at 1
+
+            a, b = snake_fold(get(f"decoder.decoder.{py}.block.0.alpha"), get(f"decoder.decoder.{py}.block.0.beta"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_SNAKE, bid, suffix=".alpha"), a)
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_SNAKE, bid, suffix=".beta"), b)
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_CONV, bid, suffix=".weight"), get(f"decoder.decoder.{py}.block.1.conv.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_CONV, bid, suffix=".bias"), get(f"decoder.decoder.{py}.block.1.conv.bias"))
+
+            for xid in range(3):
+                ridx = xid + 2  # block.2/3/4 are the 3 residual units
+
+                a1, b1 = snake_fold(get(f"decoder.decoder.{py}.block.{ridx}.act1.alpha"), get(f"decoder.decoder.{py}.block.{ridx}.act1.beta"))
+                name1 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_ACT1].format(bid=bid, xid=xid)
+                yield (name1 + ".alpha", a1)
+                yield (name1 + ".beta", b1)
+
+                name_conv1 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_CONV1].format(bid=bid, xid=xid)
+                yield (name_conv1 + ".weight", get(f"decoder.decoder.{py}.block.{ridx}.conv1.conv.weight"))
+                yield (name_conv1 + ".bias", get(f"decoder.decoder.{py}.block.{ridx}.conv1.conv.bias"))
+
+                a2, b2 = snake_fold(get(f"decoder.decoder.{py}.block.{ridx}.act2.alpha"), get(f"decoder.decoder.{py}.block.{ridx}.act2.beta"))
+                name2 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_ACT2].format(bid=bid, xid=xid)
+                yield (name2 + ".alpha", a2)
+                yield (name2 + ".beta", b2)
+
+                name_conv2 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_CONV2].format(bid=bid, xid=xid)
+                yield (name_conv2 + ".weight", get(f"decoder.decoder.{py}.block.{ridx}.conv2.conv.weight"))
+                yield (name_conv2 + ".bias", get(f"decoder.decoder.{py}.block.{ridx}.conv2.conv.bias"))
+
+        a5, b5 = snake_fold(get("decoder.decoder.5.alpha"), get("decoder.decoder.5.beta"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_SNAKE, suffix=".alpha"), a5)
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_SNAKE, suffix=".beta"), b5)
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_CONV, suffix=".weight"), get("decoder.decoder.6.conv.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_CONV, suffix=".bias"), get("decoder.decoder.6.conv.bias"))
@@ -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.

@@ -133,6 +133,7 @@ Note:
 - To debug the multimodal preprocessor and encoder, you can use [llama-mtmd-debug](tools/mtmd/debug/mtmd-debug.cpp).
 - Adding a model-specific API or CLI is an anti-pattern in `libmtmd`. The goal of `libmtmd` is to provide an easy-to-use, model-agnostic library for multimodal pipeline.
 - In most cases, `llama-mtmd-cli` should not be modified. If a model requires a specific prompt, either let the user provide it or bake it into the Jinja chat template.
+- For audio generation models, see `tools/mtmd/README-dev.md`

 ## Tips and tricks

@@ -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 | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -23,16 +23,16 @@ Legend:
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -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 | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -51,24 +51,24 @@ Legend:
 |                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -76,13 +76,13 @@ Legend:
 |                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | 🟡 |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
 |                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -103,13 +103,13 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -202,6 +202,12 @@ Example Video:

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

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

 ```
@@ -70,6 +70,8 @@ static void write_table(std::ostringstream & ss, std::vector<common_arg *> & opt

 static void write_help(std::ostringstream & ss, const md_file & md) {
    common_params params;
+    params.is_gen_docs = true;
+
    auto ctx_arg = common_params_parser_init(params, md.ex);

    std::vector<common_arg *> common_options;
@@ -3,9 +3,11 @@
 #include "common.h"
 #include "ngram-cache.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "log.h"
 #include "llama.h"

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

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

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

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

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

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

        auto params_dft = params;

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

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

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

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

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

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

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

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

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


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

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

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

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

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

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

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

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

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

 endlocal

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

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

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

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

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

 endlocal

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)

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

@@ -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,
@@ -765,8 +765,9 @@ struct ggml_backend_sched_split {
    int backend_id;
    int i_start;
    int i_end;
-    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** inputs;
    int n_inputs;
+    int inputs_capacity;
    // graph view of this split
    struct ggml_cgraph graph;
 };
@@ -805,8 +806,9 @@ struct ggml_backend_sched {
    int cur_copy;
    int next_copy;
    ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
-    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** graph_inputs;
    int n_graph_inputs;
+    int graph_inputs_capacity;

    struct ggml_context * ctx;

@@ -832,6 +834,36 @@ struct ggml_backend_sched {
 #define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
 #define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)

+static void ggml_backend_sched_split_inputs_grow(struct ggml_backend_sched_split * split) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (split->inputs_capacity > 0) {
+        new_cap = 2*split->inputs_capacity;
+        GGML_LOG_WARN("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) split->inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow split inputs container");
+    }
+    split->inputs = pnew;
+    split->inputs_capacity = new_cap;
+}
+
+static void ggml_backend_sched_graph_inputs_grow(ggml_backend_sched_t sched) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (sched->graph_inputs_capacity > 0) {
+        new_cap = 2*sched->graph_inputs_capacity;
+        GGML_LOG_WARN("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) sched->graph_inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow graph inputs container");
+    }
+    sched->graph_inputs = pnew;
+    sched->graph_inputs_capacity = new_cap;
+}
+
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
    for (int i = 0; i < sched->n_backends; i++) {
@@ -1297,7 +1329,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                    }
                    // check if the split has too many inputs
                    // FIXME: count the number of inputs instead of only checking when full
-                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                    if (split->n_inputs >= split->inputs_capacity) {
                        const size_t id = hash_id(src);
                        int src_backend_id = sched->hv_tensor_backend_ids[id];
                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
@@ -1313,10 +1345,14 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                split->i_end = i;
                i_split++;
                if (i_split >= sched->splits_capacity) {
+                    int old_cap = sched->splits_capacity;
                    sched->splits_capacity *= 2;
                    sched->splits = (ggml_backend_sched_split *)
                        realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                    GGML_ASSERT(sched->splits != NULL);
+                    for (int k = old_cap; k < sched->splits_capacity; k++) {
+                        memset(&sched->splits[k], 0, sizeof(struct ggml_backend_sched_split));
+                    }
                }
                split = &sched->splits[i_split];
                split->backend_id = node_backend_id;
@@ -1353,7 +1389,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                            SET_CAUSE(tensor_copy, "4.cpy");
                        }
                        int n_graph_inputs = sched->n_graph_inputs++;
-                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_graph_inputs >= sched->graph_inputs_capacity) {
+                            ggml_backend_sched_graph_inputs_grow(sched);
+                        }
                        sched->graph_inputs[n_graph_inputs] = src;
                    }
                }
@@ -1373,7 +1411,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                            SET_CAUSE(tensor_copy, "4.cpy");
                        }
                        int n_inputs = split->n_inputs++;
-                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_inputs >= split->inputs_capacity) {
+                            ggml_backend_sched_split_inputs_grow(split);
+                        }
                        split->inputs[n_inputs] = src;
                    }
                    node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
@@ -1399,7 +1439,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
        sched->prev_leaf_backend_ids = tmp;
    }

-    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+    int total_inputs = sched->n_graph_inputs;
+    for (int i = 0; i < sched->n_splits; i++) {
+        total_inputs += sched->splits[i].n_inputs;
+    }
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + total_inputs * 2 * sched->n_copies;

    // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
    sched->debug_prev_graph_size = sched->debug_graph_size;
@@ -1782,6 +1826,9 @@ ggml_backend_sched_t ggml_backend_sched_new(
    sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
    sched->splits_capacity = initial_splits_capacity;

+    sched->graph_inputs_capacity = GGML_SCHED_MAX_SPLIT_INPUTS;
+    sched->graph_inputs = (struct ggml_tensor **) calloc(sched->graph_inputs_capacity, sizeof(struct ggml_tensor *));
+
    for (int b = 0; b < n_backends; b++) {
        sched->backends[b] = backends[b];
        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
@@ -1814,7 +1861,11 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
    ggml_gallocr_free(sched->galloc);
    ggml_free(sched->ctx);
    ggml_hash_set_free(&sched->hash_set);
+    for (int i = 0; i < sched->splits_capacity; i++) {
+        free(sched->splits[i].inputs);
+    }
    free(sched->splits);
+    free(sched->graph_inputs);
    free(sched->hv_tensor_backend_ids);
    free(sched->hv_tensor_copies);
    free(sched->node_backend_ids);
@@ -8,6 +8,22 @@
 #include <sys/sysctl.h>
 #endif

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

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

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

 static bool ggml_thread_apply_affinity(const bool * mask) {
@@ -195,6 +195,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -214,6 +215,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -627,7 +627,8 @@ template <typename T> struct block_reduce_policy<block_reduce_method::MAX, T> {
 };

 template <block_reduce_method reduce_method_t, const unsigned int block_size_template = 0, typename T>
-static __device__ T block_reduce(T val, T * shared_vals) {
+static __device__ T block_reduce(T val, [[maybe_unused]] T * shared_vals) {
+    // for multi-warp reductions, callers must not reuse shared_vals until all reads from this invocation have completed
    val                           = block_reduce_policy<reduce_method_t, T>::reduce(val);
    const unsigned int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
    if (block_size > WARP_SIZE) {
@@ -253,9 +253,9 @@ static void ggml_cpy_f32_q8_0_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
 #ifdef GGML_CUDA_NO_PEER_COPY
+    GGML_UNUSED(dev);
    return nullptr;
 #else
    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
@@ -64,7 +64,7 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
        tmp += xi * xi;
    }

-    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum + 32);

    const float variance = tmp / group_size;
    const float scale = rsqrtf(variance + eps);
@@ -297,7 +297,7 @@ static void group_norm_f32_cuda(
        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
    } else {
        const dim3 block_dims(1024, 1, 1);
-        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
+        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 2 * 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
    }
 }

@@ -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);
@@ -116,6 +116,11 @@ static __global__ void soft_max_f32(
        vals[col] = val;
    }

+    if (block_size > WARP_SIZE) {
+        // sync is needed as we reuse buf_iw across block_reduce invocations, see #26385
+        // for block_size <= WARP_SIZE, block_reduce does not access buf_iw
+        __syncthreads();
+    }
    // find the sum of exps in the block
    tmp = block_reduce<block_reduce_method::SUM, block_size_template>(tmp, buf_iw);

@@ -142,6 +147,8 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
                                                                float * __restrict__ dst,
                                                                float * __restrict__ tmp_maxs,
                                                                float * __restrict__ tmp_sums,
+                                                                float * shared_vals_max,
+                                                                float * shared_vals_sum,
                                                                const soft_max_params p) {
    namespace cg = cooperative_groups;

@@ -154,7 +161,6 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
    float     local_vals[n_elem_per_thread] = { -INFINITY, -INFINITY, -INFINITY, -INFINITY };
    float     local_max                     = -INFINITY;
    const int step_size                     = gridDim.x * blockDim.x;
-    __shared__ float shared_vals[32];

    // Compute thread-local max
    for (int col = col_start; col < p.ncols;) {
@@ -171,7 +177,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
    }

    // Compute CTA-level max
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);

    // Store CTA-level max to GMEM
    if (tid == 0) {
@@ -186,7 +192,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
    } else {
        local_max = -INFINITY;
    }
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);

    // Compute softmax dividends, accumulate divisor
    float tmp_expf = 0.0f;
@@ -209,7 +215,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
    }

    // Reduce divisor within CTA
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);

    // Store CTA-level sum to GMEM
    if (tid == 0) {
@@ -223,7 +229,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
    } else {
        tmp_expf = 0.0f;
    }
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);

    // Divide dividend by global sum + store data
    for (int col = col_start; col < p.ncols;) {
@@ -310,9 +316,11 @@ __launch_bounds__(8*WARP_SIZE, 1) static __global__ void soft_max_f32_paralleliz
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#grid-synchronization
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#class-cluster-group
 {
+    __shared__ float shared_vals[2][32];
+
    for (int rowx = 0; rowx < p.ne01 * p.ne02 * p.ne03; rowx++) {
        soft_max_f32_parallelize_cols_single_row(x + int64_t(rowx) * p.ncols, dst + int64_t(rowx) * p.ncols, tmp_maxs,
-                                                 tmp_sums, p);
+                                                 tmp_sums, shared_vals[0], shared_vals[1], p);
    }
 }

@@ -477,6 +477,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max(ggml_me
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    char name[256];
+
+    snprintf(name, 256, "kernel_lightning_indexer_%s", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc(ggml_metal_library_t lib, ggml_op op) {
    const char * name = nullptr;

@@ -2135,6 +2152,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_sgd(ggm
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_SILU_BACK);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_silu_back_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_memset(ggml_metal_library_t lib, const ggml_tensor *  op) {
    GGML_ASSERT(op->type == GGML_TYPE_I64);

@@ -117,6 +117,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_concat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -124,6 +125,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_bl
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_tri               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
@@ -2,6 +2,7 @@

 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
+#import "ggml-metal-impl.h"

 #include <Foundation/Foundation.h>

@@ -1137,6 +1138,14 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                default:
                    return false;
            }
+        case GGML_OP_SILU_BACK:
+            return (op->src[0]->type == GGML_TYPE_F32) &&
+                (op->src[1]->type == GGML_TYPE_F32) &&
+                (op->type == GGML_TYPE_F32) &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                ggml_is_contiguous(op) &&
+                ggml_are_same_shape(op->src[0], op->src[1]);
        case GGML_OP_GLU:
            switch (ggml_get_glu_op(op)) {
                case GGML_GLU_OP_REGLU:
@@ -1259,8 +1268,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
        case GGML_OP_ARGSORT:
        case GGML_OP_TOP_K:
        case GGML_OP_ARANGE:
-        case GGML_OP_ROLL:
            return true;
+        case GGML_OP_ROLL:
+            return ggml_is_contiguous(op->src[0]);
        case GGML_OP_FLASH_ATTN_EXT:
            // for new head sizes, add checks here
            if (op->src[0]->ne[0] != 32 &&
@@ -1300,6 +1310,36 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                    return false;
            }
            return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case GGML_OP_LIGHTNING_INDEXER:
+            if (op->src[0]->ne[0] != OP_LIGHTNING_INDEXER_DK ||
+                op->src[0]->ne[1] != OP_LIGHTNING_INDEXER_NH) {
+                return false;
+            }
+            if (!has_simdgroup_mm ||
+                op->src[0]->type != GGML_TYPE_F32 ||
+                op->src[2]->type != GGML_TYPE_F32 ||
+                op->src[3]->type != GGML_TYPE_F16 ||
+                op->type         != GGML_TYPE_F32 ||
+                !ggml_is_contiguous_rows(op->src[0]) ||
+                !ggml_is_contiguous_rows(op->src[1]) ||
+                !ggml_is_contiguous_rows(op->src[2]) ||
+                !ggml_is_contiguous_rows(op->src[3])) {
+                return false;
+            }
+            switch (op->src[1]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                case GGML_TYPE_BF16:
+                    return has_bfloat;
+                default:
+                    return false;
+            }
        case GGML_OP_DSV4_HC_COMB:
            return has_simdgroup_reduction &&
                op->src[0]->type == GGML_TYPE_F32 &&
@@ -112,6 +112,13 @@
 #define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32

+#define OP_LIGHTNING_INDEXER_DK    128
+#define OP_LIGHTNING_INDEXER_NH     64
+#define OP_LIGHTNING_INDEXER_NHPTG   8
+#define OP_LIGHTNING_INDEXER_NKPSG   8
+#define OP_LIGHTNING_INDEXER_NSG     8
+#define OP_LIGHTNING_INDEXER_NBPTG   8
+
 #define OP_UNARY_NUM_SCALE      10
 #define OP_UNARY_NUM_FILL       11
 #define OP_UNARY_NUM_CLAMP      12
@@ -1171,6 +1178,23 @@ typedef struct {
    int64_t val;
 } ggml_metal_kargs_memset;

+typedef struct {
+    int32_t  n_kv;
+    int32_t  n_batch;
+    int32_t  mask_ne3;
+    uint64_t nb1;
+    uint64_t nb3;
+    uint64_t nbq1;
+    uint64_t nbq2;
+    uint64_t nbq3;
+    uint64_t nbk2;
+    uint64_t nbk3;
+    uint64_t nbw1;
+    uint64_t nbw3;
+    uint64_t nbm1;
+    uint64_t nbm3;
+} ggml_metal_kargs_lightning_indexer;
+
 typedef struct {
    int32_t  n_tokens;
    int32_t  n_iter;
@@ -1265,4 +1289,8 @@ typedef struct {
    int64_t  np;
 } ggml_metal_kargs_opt_step_sgd;

+typedef struct {
+    int64_t ne;
+} ggml_metal_kargs_silu_back;
+
 #endif // GGML_METAL_IMPL
@@ -299,6 +299,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
            {
                n_fuse = ggml_metal_op_unary(ctx, idx);
            } break;
+        case GGML_OP_SILU_BACK:
+            {
+                n_fuse = ggml_metal_op_silu_back(ctx, idx);
+            } break;
        case GGML_OP_GLU:
            {
                n_fuse = ggml_metal_op_glu(ctx, idx);
@@ -316,6 +320,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
            {
                n_fuse = ggml_metal_op_cumsum(ctx, idx);
            } break;
+        case GGML_OP_LIGHTNING_INDEXER:
+            {
+                n_fuse = ggml_metal_op_lightning_indexer(ctx, idx);
+            } break;
        case GGML_OP_DSV4_HC_COMB:
        case GGML_OP_DSV4_HC_PRE:
        case GGML_OP_DSV4_HC_POST:
@@ -1303,6 +1311,72 @@ int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
    return 1;
 }

+int ggml_metal_op_lightning_indexer(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    const ggml_tensor * q = op->src[0];
+    const ggml_tensor * k = op->src[1];
+    const ggml_tensor * w = op->src[2];
+    const ggml_tensor * m = op->src[3];
+
+    GGML_ASSERT(q->type == GGML_TYPE_F32);
+    GGML_ASSERT(k->type == GGML_TYPE_F32  ||
+                k->type == GGML_TYPE_F16  ||
+                k->type == GGML_TYPE_BF16 ||
+                k->type == GGML_TYPE_Q4_0 ||
+                k->type == GGML_TYPE_Q4_1 ||
+                k->type == GGML_TYPE_Q5_0 ||
+                k->type == GGML_TYPE_Q5_1 ||
+                k->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(w->type == GGML_TYPE_F32);
+    GGML_ASSERT(m->type == GGML_TYPE_F16);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(q->ne[0] == OP_LIGHTNING_INDEXER_DK);
+    GGML_ASSERT(q->ne[1] == OP_LIGHTNING_INDEXER_NH);
+
+    ggml_metal_kargs_lightning_indexer args = {
+        /*.n_kv      =*/ (int32_t) k->ne[2],
+        /*.n_batch   =*/ (int32_t) q->ne[2],
+        /*.mask_ne3  =*/ (int32_t) m->ne[3],
+        /*.nb1       =*/ op->nb[1],
+        /*.nb3       =*/ op->nb[3],
+        /*.nbq1      =*/ q->nb[1],
+        /*.nbq2      =*/ q->nb[2],
+        /*.nbq3      =*/ q->nb[3],
+        /*.nbk2      =*/ k->nb[2],
+        /*.nbk3      =*/ k->nb[3],
+        /*.nbw1      =*/ w->nb[1],
+        /*.nbw3      =*/ w->nb[3],
+        /*.nbm1      =*/ m->nb[1],
+        /*.nbm3      =*/ m->nb[3],
+    };
+
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(q),  1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(k),  2);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(w),  3);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(m),  4);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 5);
+
+    const int nsg   = OP_LIGHTNING_INDEXER_NSG;
+    const int nkptg = OP_LIGHTNING_INDEXER_NKPSG*nsg;
+    const int nbptg = OP_LIGHTNING_INDEXER_NBPTG;
+
+    auto pipeline = ggml_metal_library_get_pipeline_lightning_indexer(ctx->lib, op);
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_dispatch_threadgroups(enc,
+            (k->ne[2] + nkptg - 1)/nkptg,
+            (q->ne[2] + nbptg - 1)/nbptg,
+            q->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_dsv4_hc(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -3473,6 +3547,36 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
    return n_fuse;
 }

+int ggml_metal_op_silu_back(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    auto pipeline = ggml_metal_library_get_pipeline_silu_back(lib, op);
+
+    const int64_t ne = ggml_nelements(op);
+
+    ggml_metal_kargs_silu_back args = {
+        /*.ne =*/ ne,
+    };
+
+    int arg_idx{0};
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op), arg_idx++);
+
+    const int nth = std::min<int64_t>(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne);
+    const int64_t n = (ne + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -3712,7 +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;

@@ -54,6 +54,7 @@ int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_diag              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_lightning_indexer (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_dsv4_hc           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
@@ -71,6 +72,7 @@ int ggml_metal_op_mul_mat_id        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_add_id            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_flash_attn_ext    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_bin               (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_silu_back         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_l2_norm           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);
@@ -1255,6 +1255,20 @@ template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_un
 template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
 template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;

+kernel void kernel_silu_back_f32(
+        constant ggml_metal_kargs_silu_back & args,
+        device const float * dy,
+        device const float * x,
+        device       float * dx,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= args.ne) {
+        return;
+    }
+
+    const float s = 1.0f / (1.0f + exp(-x[gid]));
+    dx[gid] = dy[gid] * s * (1.0f + x[gid] * (1.0f - s));
+}
+
 // OP: 0 - add, 1 - sub, 2 - mul, 3 - div
 constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
 constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
@@ -11281,6 +11295,155 @@ typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;

 template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;

+template<
+    typename kd4x4_t,
+    short nl_k,
+    void (*deq_k)(device const kd4x4_t *, short, thread half4x4 &)>
+kernel void kernel_lightning_indexer(
+        constant ggml_metal_kargs_lightning_indexer & args,
+        device const char * q,
+        device const char * k,
+        device const char * w,
+        device const char * m,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    constexpr short DK    = OP_LIGHTNING_INDEXER_DK;
+    constexpr short NH    = OP_LIGHTNING_INDEXER_NH;
+    constexpr short NHPTG = OP_LIGHTNING_INDEXER_NHPTG;
+    constexpr short NKPSG = OP_LIGHTNING_INDEXER_NKPSG;
+    constexpr short NSG   = OP_LIGHTNING_INDEXER_NSG;
+    constexpr short NBPTG = OP_LIGHTNING_INDEXER_NBPTG;
+
+    constexpr short DK4  = DK/4;
+    constexpr short DK8  = DK/8;
+    constexpr short DK16 = DK/16;
+
+    constexpr short NK  = NKPSG*NSG; // keys    per threadgroup
+    constexpr short NTG = 32*NSG;    // threads per threadgroup
+
+    const int i_stream = tgpig.z;
+    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 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;
+        const short i16 = i%DK16;
+
+        half4x4 tmp;
+
+        if (i_kv_0 + ik < args.n_kv) {
+            device const kd4x4_t * kr = (device const kd4x4_t *) (k + (i_kv_0 + ik)*args.nbk2 + i_stream*args.nbk3);
+
+            deq_k(kr + i16/nl_k, i16%nl_k, tmp);
+        } else {
+            FOR_UNROLL (short j = 0; j < 4; ++j) {
+                tmp[j] = half4(0.0h);
+            }
+        }
+
+        sk4x4[i] = tmp;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // K tile of this simdgroup, transposed to [DK, NKPSG]
+    simdgroup_half8x8 mk[DK8];
+
+    FOR_UNROLL (short i = 0; i < DK8; ++i) {
+        simdgroup_load(mk[i], sk + sgitg*NKPSG*DK + 8*i, DK, 0, true);
+    }
+
+    threadgroup half4   sq4[NHPTG*DK4];
+    threadgroup half  * sq = (threadgroup half *) sq4;
+
+    threadgroup float sw [NHPTG];
+    threadgroup float sqk[NSG*NHPTG*NKPSG];
+
+    const int i_batch_0 = tgpig.y*NBPTG;
+    const int n_batch   = min((int) NBPTG, args.n_batch - i_batch_0);
+
+    for (short ib = 0; ib < n_batch; ++ib) {
+        const int i_batch = i_batch_0 + ib;
+
+        device const char * pq = q + i_batch*args.nbq2 + i_stream*args.nbq3;
+        device const char * pw = w + i_batch*args.nbw1 + i_stream*args.nbw3;
+
+        float score = 0.0f;
+
+        FOR_UNROLL (short i_head = 0; i_head < NH; i_head += NHPTG) {
+            // stage the Q tile [DK, NHPTG] and the (prescaled) head weights
+            for (short i = tiitg; i < NHPTG*DK4; i += NTG) {
+                const short ih = i/DK4;
+                const short i4 = i%DK4;
+
+                device const float4 * q4 = (device const float4 *) (pq + (i_head + ih)*args.nbq1);
+
+                sq4[ih*DK4 + i4] = half4(q4[i4]);
+            }
+
+            if (tiitg < NHPTG) {
+                sw[tiitg] = ((device const float *) pw)[i_head + tiitg];
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            simdgroup_float8x8 mqk = make_filled_simdgroup_matrix<float, 8>(0.0f);
+
+            FOR_UNROLL (short i = 0; i < DK8; ++i) {
+                simdgroup_half8x8 mq;
+
+                simdgroup_load(mq, sq + 8*i, DK, 0, false);
+                simdgroup_multiply_accumulate(mqk, mq, mk[i], mqk);
+            }
+
+            threadgroup float * pqk = sqk + sgitg*NHPTG*NKPSG;
+
+            simdgroup_store(mqk, pqk, NKPSG, 0, false);
+            simdgroup_barrier(mem_flags::mem_threadgroup);
+
+            // one lane per key: ReLU, apply the head weight and accumulate over the head tile
+            if (tiisg < NKPSG) {
+                FOR_UNROLL (short ih = 0; ih < NHPTG; ++ih) {
+                    score += max(pqk[ih*NKPSG + tiisg], 0.0f)*sw[ih];
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+
+        if (tiisg < NKPSG) {
+            const int ik = i_kv + tiisg;
+            if (ik < args.n_kv) {
+                device const half  * pm = (device const half  *) (m   + i_batch*args.nbm1 + (i_stream % args.mask_ne3)*args.nbm3);
+                device       float * pd = (device       float *) (dst + i_batch*args.nb1  + i_stream*args.nb3);
+
+                pd[ik] = score + (float) pm[ik];
+            }
+        }
+    }
+}
+
+typedef decltype(kernel_lightning_indexer<half4x4, 1, dequantize_f16>) kernel_lightning_indexer_t;
+
+template [[host_name("kernel_lightning_indexer_f32")]]  kernel kernel_lightning_indexer_t kernel_lightning_indexer<float4x4, 1, dequantize_f32>;
+template [[host_name("kernel_lightning_indexer_f16")]]  kernel kernel_lightning_indexer_t kernel_lightning_indexer<half4x4,  1, dequantize_f16>;
+
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_lightning_indexer_bf16")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<bfloat4x4, 1, dequantize_bf16>;
+#endif
+
+template [[host_name("kernel_lightning_indexer_q4_0")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_lightning_indexer_q4_1")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_lightning_indexer_q5_0")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_lightning_indexer_q5_1")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_lightning_indexer_q8_0")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q8_0, 2, dequantize_q8_0>;
+
 kernel void kernel_dsv4_hc_comb_f32(
        constant ggml_metal_kargs_dsv4_hc_comb & args,
        device const char * mixes,
@@ -11438,4 +11601,3 @@ kernel void kernel_dsv4_hc_post_f32(
        *(device float *) (dst + i0*args.nb_d0 + idst*args.nb_d1 + it*args.nb_d2) = result[idst];
    }
 }
-
@@ -73,6 +73,7 @@ typedef const void * (*get_adreno_bin_kernel_func_t)(
 //------------------------------------------------------------------------------

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

@@ -7065,7 +7083,7 @@ static inline bool use_flat_gemv_for_large_m_q4_K(const ggml_tensor *tensor) {
    return tensor->ne[1] >= 32768 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }

-static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
+static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
    // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
    // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
    // q6_K flat gemv is worse for smaller K; 2048 seems to be a reasonable threshold.
@@ -7083,7 +7101,15 @@ static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
    if ((tensor->ne[1] % 128 != 0) && tensor->ne[2] == 1 && tensor->ne[3] == 1) {
        return true;
    }
-    return tensor->ne[1] >= 32768 && tensor->ne[0] >= 2048 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
+
+    // The gemv_noshuffle slowdown tracks TOTAL weight size, not ne0 alone; ne0 >= 2048 is a
+    // proxy for "large weight" that misses a narrow-hidden vocab-scale lm_head.
+    // Add a direct size escape so such weights also take the flat path, without changing
+    // which weights ne0 >= 2048 already routes there.
+    // The size escape is not taken on the A7X since its compiler miscompiles the flat K-quant GEMV
+    return tensor->ne[1] >= 32768
+        && (tensor->ne[0] >= 2048 || (backend_ctx->adreno_gen != ADRENO_GPU_GEN::A7X && ggml_nbytes(tensor) >= (256ull << 20)))
+        && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }

 static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
@@ -9403,7 +9429,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        cl_kernel kernel;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
        kernel = backend_ctx->kernel_convert_block_q6_K;
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
            kernel = backend_ctx->kernel_convert_block_q6_K_noshuffle;
        }
 #else
@@ -9436,7 +9462,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        tensor->extra  = extra;

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
            cl_int M = tensor->ne[1];   // ne01
            cl_int K = tensor->ne[0];   // ne00

@@ -10473,7 +10499,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
            CL_CHECK(clReleaseMemObject(data_device));
            return;
        }
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
            static ggml_cl_buffer buf_trans_ql;
            static ggml_cl_buffer buf_trans_qh;
            static ggml_cl_buffer buf_trans_s;
@@ -18895,7 +18921,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
        }

        // q6_K x fp32
-        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(src0)) {
+        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(backend_ctx, src0)) {
            ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);
            return;
        }
@@ -211,7 +211,30 @@ __kernel void FA_TILE_NAME(

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

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

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

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

 #ifdef FA_HAVE_INT_DOT
+// Accessors so the staging code is layout-agnostic.
+#ifdef FA_K_LDS_T
+#define FA_K_PACKED(ROW, IDX) l_k_packed[IDX][ROW]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[BLK][ROW]
+#else
+#define FA_K_PACKED(ROW, IDX) l_k_packed[ROW][IDX]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[ROW][BLK]
+#endif
+
+#ifdef FA_K_LDS_T
+    // K tile transposed: [block*8 + g][kv row] instead of [kv row][block*8 + g].
+    //
+    // The QK loop walks 4 KV rows at a time against the same (b, g), so in the original
+    // layout those 4 values are BLOCK_N*8 uints apart and cost 4 separate 32-bit local
+    // reads. Transposed they are adjacent, so they are one 128-bit read -- 4x fewer LDS
+    // issues for the same bytes and no extra registers. That matters because the QK loop
+    // is LDS-read-issue-bound: a wrong-math probe that kept every dp4a but cut the LDS
+    // reads ran the whole kernel 41% faster (18.51 -> 10.91 ms/op), and deleting QK
+    // outright only reached 10.88 -- i.e. essentially ALL of QK's cost is these reads.
+    __local uint  l_k_packed[DK_Q8_BLOCKS_PREFILL * 8][BLOCK_N];
+    __local float l_k_scale [DK_Q8_BLOCKS_PREFILL][BLOCK_N];
+#else
    __local uint  l_k_packed[BLOCK_N][DK_Q8_BLOCKS_PREFILL * 8];
    __local float l_k_scale [BLOCK_N][DK_Q8_BLOCKS_PREFILL];
+#endif
 #else
    __local half4 l_k[BLOCK_N][DK_VEC];
 #endif
@@ -1427,7 +1450,7 @@ __kernel void flash_attn_f32_q8_0(
                    const global char * blk_ptr = k_base + k_row_off + blk * Q8_0_BLOCK_SIZE;
                    const float df = (float) vload_half(0, (const global half *) blk_ptr);
                    const global uchar * qs = (const global uchar *)(blk_ptr + 2);
-                    l_k_scale[row][blk] = df;
+                    FA_K_SCALE(row, blk) = df;
                    #pragma unroll
                    for (int j = 0; j < 8; ++j) {
                        uint k_packed =
@@ -1435,12 +1458,12 @@ __kernel void flash_attn_f32_q8_0(
                             ((uint) qs[j*4 + 1]) <<  8 |
                             ((uint) qs[j*4 + 2]) << 16 |
                             ((uint) qs[j*4 + 3]) << 24;
-                        l_k_packed[row][blk * 8 + j] = k_packed;
+                        FA_K_PACKED(row, blk * 8 + j) = k_packed;
                    }
                } else {
-                    l_k_scale[row][blk] = 0.0f;
+                    FA_K_SCALE(row, blk) = 0.0f;
                    #pragma unroll
-                    for (int j = 0; j < 8; ++j) l_k_packed[row][blk * 8 + j] = 0u;
+                    for (int j = 0; j < 8; ++j) FA_K_PACKED(row, blk * 8 + j) = 0u;
                }
            }
 #else
@@ -1556,6 +1579,19 @@ __kernel void flash_attn_f32_q8_0(
                for (int b_local = 0; b_local < SPLIT_DK_Q8_BLOCKS; ++b_local) {
                    const int b = k_blk_base + b_local;
                    int sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+#if defined(FA_K_LDS_T)
+                    // The 4 KV rows are adjacent in the transposed tile, so each (b, g)
+                    // step is ONE 128-bit local read instead of four 32-bit ones.
+                    #pragma unroll
+                    for (int g = 0; g < 8; ++g) {
+                        const uint qp = q_packed_pf[b_local * 8 + g];
+                        const uint4 kq4 = vload4(0, &l_k_packed[b * 8 + g][j]);
+                        sum0 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s0, sum0);
+                        sum1 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s1, sum1);
+                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s2, sum2);
+                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s3, sum3);
+                    }
+#else
                    #pragma unroll
                    for (int g = 0; g < 8; ++g) {
                        const uint qp = q_packed_pf[b_local * 8 + g];
@@ -1564,11 +1600,20 @@ __kernel void flash_attn_f32_q8_0(
                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+2][b * 8 + g], sum2);
                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+3][b * 8 + g], sum3);
                    }
+#endif
                    const float qd = q_d_pf[b_local];
+#ifdef FA_K_LDS_T
+                    const float4 ks4 = vload4(0, &l_k_scale[b][j]);
+                    s0 += (float)sum0 * qd * ks4.s0;
+                    s1 += (float)sum1 * qd * ks4.s1;
+                    s2 += (float)sum2 * qd * ks4.s2;
+                    s3 += (float)sum3 * qd * ks4.s3;
+#else
                    s0 += (float)sum0 * qd * l_k_scale[j  ][b];
                    s1 += (float)sum1 * qd * l_k_scale[j+1][b];
                    s2 += (float)sum2 * qd * l_k_scale[j+2][b];
                    s3 += (float)sum3 * qd * l_k_scale[j+3][b];
+#endif
                }
 #else
                ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
@@ -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
@@ -127,7 +127,15 @@ static void concat_T_sycl_non_cont(
    int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
    uint64_t nb3, int32_t dim) {
  sycl::range<3> gridDim(ne3, ne2, ne1);
-  stream->parallel_for(sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+
+  // Avoid oversubscribing device when there is not enough elements along the innermost dim to
+  // fill a full SYCL_CONCAT_BLOCK_SIZE. For larger # of elements, the full SYCL_CONCAT_BLOCK_SIZE
+  // is used.
+  const int64_t ne0_pad   = GGML_PAD(ne0, WARP_SIZE);
+  const int64_t block_ne0 = ne0_pad < SYCL_CONCAT_BLOCK_SIZE ? ne0_pad : (int64_t) SYCL_CONCAT_BLOCK_SIZE;
+  sycl::range<3> blockDim(1, 1, block_ne0);
+
+  stream->parallel_for(sycl::nd_range<3>(gridDim * blockDim, blockDim), [=](sycl::nd_item<3> item_ct1) {
      int64_t i3 = item_ct1.get_group(0);
      int64_t i2 = item_ct1.get_group(1);
      int64_t i1 = item_ct1.get_group(2);
@@ -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);
+    });
 }


@@ -2,11 +2,13 @@
 #include <cstdio>
 #include <cstring>
 #include <string>
+#include <optional>
 #include <unordered_map>
 #include <vector>

 #include "fattn-onednn.hpp"
 #include "fattn-tile.hpp"
+#include "convert.hpp"

 // set minimum query length to treat as prefill (32)
 #define GGML_SYCL_FA_ONEDNN_MIN_Q 32
@@ -33,10 +35,30 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
    const ggml_tensor * mask  = dst->src[3];
    const ggml_tensor * sinks = dst->src[4];

-    // gate for f16 KV only for now
-    // need to implement quantized KV
+    // F16 KV: native SDPA at any KV length.
+    // Non-F16: dequant to F16 then SDPA at prefill lengths. Only the
+    // standard quantized KV cache types (Q4_0-Q8_0) and F32 are accepted
+    // because their to_fp16_sycl conversion is verified. BF16 and IQ*
+    // are excluded: BF16 needs a strided conversion kernel that does not
+    // exist yet; IQ types are model-weight-only quants with no dequant
+    // registration and are never used as KV caches.
    if (K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16) {
-        return false;
+        auto kt = K->type, vt = V->type;
+        bool k_ok = kt == GGML_TYPE_F32 || kt == GGML_TYPE_Q4_0 || kt == GGML_TYPE_Q4_1 ||
+                    kt == GGML_TYPE_Q5_0 || kt == GGML_TYPE_Q5_1 || kt == GGML_TYPE_Q8_0;
+        bool v_ok = vt == GGML_TYPE_F32 || vt == GGML_TYPE_Q4_0 || vt == GGML_TYPE_Q4_1 ||
+                    vt == GGML_TYPE_Q5_0 || vt == GGML_TYPE_Q5_1 || vt == GGML_TYPE_Q8_0;
+        if (!k_ok || !v_ok) {
+            return false;
+        }
+        if (Q->ne[1] < 32 || K->ne[1] < 1024) {
+            return false;
+        }
+        for (const ggml_tensor * t : {K, V}) {
+            if (t->type == GGML_TYPE_F16 && t->nb[1] % (t->ne[0] * 2) != 0) {
+                return false;
+            }
+        }
    }
    // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
    // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
@@ -205,13 +227,101 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
    dnnl::engine    eng    = ctx.engine_dnnl(stream);
    dnnl::stream    strm   = ctx.stream_dnnl(stream);

-    // cont/cast inputs to contiguous f16 (head-major) -- the layout the fast systolic path wants.
-    ggml_sycl_pool_alloc<sycl::half> Qf(ctx.pool(), (size_t) H   * q   * d);
-    ggml_sycl_pool_alloc<sycl::half> Kf(ctx.pool(), (size_t) Hkv * seq * d);
-    ggml_sycl_pool_alloc<sycl::half> Vf(ctx.pool(), (size_t) Hkv * seq * d);
-    cont_to_f16_sycl<float>     ((const char *) Q->data, Qf.get(), d, q,   H,   mb, Q->nb[1], Q->nb[2], Q->nb[3], stream);
-    cont_to_f16_sycl<sycl::half>((const char *) K->data, Kf.get(), d, seq, Hkv, mb, K->nb[1], K->nb[2], K->nb[3], stream);
-    cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf.get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
+    // Q: always f32 -- copy to dense f16.
+    ggml_sycl_pool_alloc<sycl::half> Qf(ctx.pool(), (size_t) H * q * d);
+    cont_to_f16_sycl<float>((const char *) Q->data, Qf.get(), d, q, H, mb, Q->nb[1], Q->nb[2], Q->nb[3], stream);
+
+    // K/V: use pool-alloc for both F16 and dequant paths.
+    sycl::half * K_ptr = nullptr;
+    sycl::half * V_ptr = nullptr;
+    std::optional<ggml_sycl_pool_alloc<sycl::half>> Kf_pool;
+    std::optional<ggml_sycl_pool_alloc<sycl::half>> Vf_pool;
+
+    if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+        Kf_pool.emplace(ctx.pool(), (size_t) Hkv * seq * d);
+        Vf_pool.emplace(ctx.pool(), (size_t) Hkv * seq * d);
+        cont_to_f16_sycl<sycl::half>((const char *) K->data, Kf_pool->get(), d, seq, Hkv, mb, K->nb[1], K->nb[2], K->nb[3], stream);
+        cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf_pool->get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
+        K_ptr = Kf_pool->get();
+        V_ptr = Vf_pool->get();
+    } else if (ggml_is_quantized(K->type)) {
+        // Quantized K/V: dequant to dense F16 using pool, same lifetime as F16 path.
+        Kf_pool.emplace(ctx.pool(), ggml_nelements(K));
+        K_ptr = Kf_pool->get();
+        {
+            const char * K_data = (const char *)K->data;
+            const bool k_non_dense = ((int64_t)K->ne[1] * K->nb[1] != K->nb[2]) && K->ne[2] > 1;
+            const bool k_gemma = k_non_dense &&
+                ((int64_t)K->nb[2] < (int64_t)K->ne[1] * (int64_t)K->nb[1]);
+            if (ggml_is_contiguously_allocated(K) && !k_non_dense) {
+                to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(K->type, dst);
+                to_fp16(K_data, K_ptr, ggml_nelements(K), stream);
+            } else {
+                const size_t bs = ggml_blck_size(K->type);
+                const size_t ts = ggml_type_size(K->type);
+                to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(K->type);
+                int64_t s01, s02, s03;
+                if (k_gemma) {
+                    const int64_t blk_per_row = (int64_t)K->ne[0] / bs;
+                    s01 = (int64_t)Hkv * blk_per_row;
+                    s02 = blk_per_row;
+                    s03 = (int64_t)K->ne[1] * s01;
+                } else {
+                    s01 = (int64_t)K->nb[1] / ts;
+                    s02 = (int64_t)K->nb[2] / ts;
+                    s03 = (int64_t)K->nb[3] / ts;
+                }
+                to_fp16(K_data, K_ptr,
+                        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                        s01, s02, s03, stream);
+            }
+        }
+        // Quantized V: always dequant separately. Even when K and V share
+        // the same underlying allocation (V is a view of K with the same
+        // data pointer), their logical values differ because the quantized
+        // elements at different positions/offsets represent different K/V
+        // data. Master's F16 path also never aliases K and V.
+        Vf_pool.emplace(ctx.pool(), ggml_nelements(V));
+        V_ptr = Vf_pool->get();
+        {
+            const char * V_data = (const char *)V->data;
+            const bool v_non_dense = ((int64_t)V->ne[1] * V->nb[1] != V->nb[2]) && V->ne[2] > 1;
+            const bool v_gemma = v_non_dense &&
+                ((int64_t)V->nb[2] < (int64_t)V->ne[1] * (int64_t)V->nb[1]);
+            if (ggml_is_contiguously_allocated(V) && !v_non_dense) {
+                to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(V->type, dst);
+                to_fp16(V_data, V_ptr, ggml_nelements(V), stream);
+            } else {
+                const size_t bs = ggml_blck_size(V->type);
+                const size_t ts = ggml_type_size(V->type);
+                to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(V->type);
+                int64_t s01, s02, s03;
+                if (v_gemma) {
+                    const int64_t blk_per_row = (int64_t)V->ne[0] / bs;
+                    s01 = (int64_t)V->ne[2] * blk_per_row;
+                    s02 = blk_per_row;
+                    s03 = (int64_t)V->ne[1] * s01;
+                } else {
+                    s01 = (int64_t)V->nb[1] / ts;
+                    s02 = (int64_t)V->nb[2] / ts;
+                    s03 = (int64_t)V->nb[3] / ts;
+                }
+                to_fp16(V_data, V_ptr,
+                        V->ne[0], V->ne[1], V->ne[2], V->ne[3],
+                        s01, s02, s03, stream);
+            }
+        }
+    } else {
+        // F32: strided copy to dense F16 via cont_to_f16_sycl<float>.
+        Kf_pool.emplace(ctx.pool(), ggml_nelements(K));
+        K_ptr = Kf_pool->get();
+        cont_to_f16_sycl<float>((const char *) K->data, K_ptr, K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                                K->nb[1], K->nb[2], K->nb[3], stream);
+        Vf_pool.emplace(ctx.pool(), ggml_nelements(V));
+        V_ptr = Vf_pool->get();
+        cont_to_f16_sycl<float>((const char *) V->data, V_ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3],
+                                V->nb[1], V->nb[2], V->nb[3], stream);
+    }

    // divide-by-(1/scale) reproduces ggml's score *= kq_scale on the proven probe graph.
    //
@@ -244,8 +354,8 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso

    auto id2ptr = [&](size_t r) -> void * {
        if (r == E.id_q)     return Qf.get();
-        if (r == E.id_k)     return Kf.get();
-        if (r == E.id_v)     return Vf.get();
+        if (r == E.id_k)     return K_ptr;
+        if (r == E.id_v)     return V_ptr;
        if (r == E.id_scale) return scale_dev;
        if (r == E.id_mask)  return (void *) mask->data;
        return nullptr;
@@ -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>
@@ -97,7 +97,7 @@ static void ggml_sycl_flash_attn_ext_vec(ggml_backend_sycl_context & ctx, ggml_t
 enum best_fattn_kernel {
    BEST_FATTN_KERNEL_NONE     =   0,
    BEST_FATTN_KERNEL_VEC      = 100,
-    BEST_FATTN_KERNEL_ONEDNN   = 150, // added enum for onednn==150
+    BEST_FATTN_KERNEL_ONEDNN   = 150, // oneDNN SDPA: native F16 (PR #25222)
    BEST_FATTN_KERNEL_TILE     = 200,
    BEST_FATTN_KERNEL_MKL      = 300,
 };
@@ -130,6 +130,14 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const

    bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;

+    // XMX-accelerated path: oneDNN SDPA (native F16 and dequant+non-F16).
+    // ONEDNN requires min 32 query tokens — short-circuit decode to avoid
+    // calling _supported() on every decode FA call.
+    if (Q->ne[1] >= 32
+        && ggml_sycl_flash_attn_ext_onednn_supported(dst)) {
+        return BEST_FATTN_KERNEL_ONEDNN;
+    }
+
    // MKL path: XMX-accelerated GEMM for prompt processing (all KV cache types).
    // The MKL kernel converts non-F16 K/V to F16 via to_fp16_sycl before GEMM,
    // so quantized, F16, BF16, and F32 caches all benefit from XMX acceleration.
@@ -167,7 +175,6 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
            return BEST_FATTN_KERNEL_MKL;
        }
    }
-
    for (const ggml_tensor * t : {Q, K, V, mask}) {
        if (t == nullptr || ggml_is_quantized(t->type)) {
            continue;
@@ -215,6 +222,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
    switch (K->type) {
        case GGML_TYPE_F32:
        case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
            break;
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -233,8 +241,11 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
        return BEST_FATTN_KERNEL_NONE;
    }

-    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
-    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes.
+    // BF16 is excluded: the VEC kernel has no BF16 template (it needs GGML_SYCL_FA_ALL_QUANTS for non-F16/Q4_0/Q8_0).
+    const bool has_bf16 = (K->type == GGML_TYPE_BF16 || V->type == GGML_TYPE_BF16);
+    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0
+        && !has_bf16;

    // Fused-XMX path: oneDNN Graph SDPA (flash attention). Strictly
    // additive -- taken only when statically supported, otherwise falls through to VEC/TILE below.
@@ -276,6 +287,7 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
        const char * kname = "TILE";
        best_fattn_kernel k = ggml_sycl_get_best_fattn_kernel(ctx.device, dst);
        if (k == BEST_FATTN_KERNEL_MKL)  kname = "MKL";
+        if (k == BEST_FATTN_KERNEL_ONEDNN)  kname = "ONEDNN";
        if (k == BEST_FATTN_KERNEL_VEC)  kname = "VEC";
        int64_t delta = 0;
        if (Dk == 256) {
@@ -292,7 +304,8 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
                (long long)V_dbg->ne[1]);
    }

-    switch (ggml_sycl_get_best_fattn_kernel(ggml_sycl_get_device(), dst)) {
+    const best_fattn_kernel fk = ggml_sycl_get_best_fattn_kernel(ggml_sycl_get_device(), dst);
+    switch (fk) {
        case BEST_FATTN_KERNEL_NONE:
            GGML_ABORT("Not support Flash-Attention");
        case BEST_FATTN_KERNEL_ONEDNN:
@@ -331,6 +344,7 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
            q->wait();
            const char * kname = "???";
            best_fattn_kernel kb = ggml_sycl_get_best_fattn_kernel(ctx.device, dst);
+            if (kb == BEST_FATTN_KERNEL_ONEDNN) kname = "ONEDNN";
            if (kb == BEST_FATTN_KERNEL_MKL) kname = "MKL";
            if (kb == BEST_FATTN_KERNEL_TILE) kname = "TILE";
            if (kb == BEST_FATTN_KERNEL_VEC) kname = "VEC";
@@ -354,6 +368,7 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
            }
        }
    }
+
 }

 bool ggml_sycl_flash_attn_ext_supported(int device, const ggml_tensor * dst) {
@@ -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];
@@ -1026,6 +1062,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_pool2d_f32;
    vk_pipeline pipeline_rwkv_wkv6_f32;
    vk_pipeline pipeline_rwkv_wkv7_f32;
+    vk_pipeline pipeline_gated_linear_attn_f32;
    // [size_idx][kda] where size_idx: 0=d16, 1=d32, 2=d64, 3=d128
    vk_pipeline pipeline_gated_delta_net[4][2];
    vk_pipeline pipeline_ssm_scan_f32_d128;
@@ -1117,6 +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;
@@ -1747,6 +1835,13 @@ struct vk_op_rwkv_wkv7_push_constants {
    uint32_t C;
    uint32_t H;
 };
+struct vk_op_gated_linear_attn_push_constants {
+    uint32_t B;
+    uint32_t T;
+    uint32_t C;
+    uint32_t H;
+    float scale;
+};
 struct vk_op_gated_delta_net_push_constants {
    uint32_t H;
    uint32_t n_tokens;
@@ -2051,6 +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) {
@@ -2463,17 +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();
@@ -3164,6 +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());
@@ -5665,6 +5801,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

    ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);

+    ggml_vk_create_pipeline(device, device->pipeline_gated_linear_attn_f32, "gated_linear_attn_f32", gated_linear_attn_f32_len, gated_linear_attn_f32_data, "main", 6, sizeof(vk_op_gated_linear_attn_push_constants), {1, 1, 1}, {}, 1);
+
    {
        const uint32_t gdn_sizes[] = {16, 32, 64, 128};
        const char * gdn_names[][2] = {
@@ -6107,6 +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;
            }
        }

@@ -6461,8 +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.
@@ -6761,6 +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);

@@ -6883,6 +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 &&
@@ -8309,7 +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);
    }
@@ -8421,7 +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);

@@ -8442,7 +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);

@@ -8477,7 +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 {
@@ -8521,7 +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);
 }
@@ -11392,6 +11549,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            return ctx->device->pipeline_rwkv_wkv7_f32;
        }
        return nullptr;
+    case GGML_OP_GATED_LINEAR_ATTN:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_gated_linear_attn_f32;
+        }
+        return nullptr;
    case GGML_OP_GATED_DELTA_NET:
        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            const uint32_t S_v = dst->src[2]->ne[0];
@@ -12422,6 +12584,41 @@ static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx,
    );
 }

+static void ggml_vk_gated_linear_attn(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
+    const size_t seq_length = dst->src[0]->ne[2];
+    const size_t n_embed    = dst->ne[0];
+    const size_t n_heads    = dst->src[0]->ne[1];
+    const size_t n_seqs     = dst->src[4]->ne[1];
+
+    float scale;
+    memcpy(&scale, dst->op_params, sizeof(float));
+
+    GGML_ASSERT(dst->buffer != nullptr);
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, dst->src[0], dst->src[1], dst->src[2], dst, dst->op);
+    GGML_ASSERT(pipeline != nullptr);
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+    vk_subbuffer src_buf[5] = {};
+    for (int i = 0; i < 5; i++) {
+        src_buf[i] = ggml_vk_tensor_subbuffer(ctx, dst->src[i]);
+    }
+
+    const vk_op_gated_linear_attn_push_constants pc = {
+        (uint32_t)n_seqs,
+        (uint32_t)seq_length,
+        (uint32_t)n_embed,
+        (uint32_t)n_heads,
+        scale,
+    };
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {src_buf[0], src_buf[1], src_buf[2], src_buf[3], src_buf[4], dst_buf},
+        pc, { (uint32_t)(n_seqs * n_heads), 1, 1 });
+}
+
 static void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
    const ggml_tensor * src_q     = dst->src[0];
    const ggml_tensor * src_v     = dst->src[2];
@@ -14216,7 +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);

@@ -14418,7 +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);

@@ -14704,7 +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);

@@ -15421,6 +15618,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr

        break;

+    case GGML_OP_GATED_LINEAR_ATTN:
+        ggml_vk_gated_linear_attn(ctx, compute_ctx, node);
+
+        break;
+
    case GGML_OP_GATED_DELTA_NET:
        ggml_vk_gated_delta_net(ctx, compute_ctx, node);

@@ -15498,7 +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 {
@@ -16109,12 +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,
@@ -16131,7 +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;
@@ -16703,6 +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";
@@ -16794,6 +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;
@@ -16801,8 +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
@@ -17056,13 +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;
@@ -17078,13 +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++) {
@@ -18128,6 +18382,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_RWKV_WKV6:
        case GGML_OP_RWKV_WKV7:
            return true; // all inputs are contiguous, see ggml.c
+        case GGML_OP_GATED_LINEAR_ATTN:
+            // the shader block size is hardcoded to head_size 64
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 && op->src[0]->ne[0] == 64;
        case GGML_OP_GATED_DELTA_NET:
            {
                const uint32_t S_v = op->src[2]->ne[0];
@@ -18308,7 +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) {
@@ -19117,6 +19374,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
        } else if (tensor->op == GGML_OP_RWKV_WKV7) {
            tensor_clone = ggml_rwkv_wkv7(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3],
            src_clone[4], src_clone[5], src_clone[6]);
+        } else if (tensor->op == GGML_OP_GATED_LINEAR_ATTN) {
+            const float * op_params = (const float *)tensor->op_params;
+            tensor_clone = ggml_gated_linear_attn(ggml_ctx, src_clone[0], src_clone[1],
+            src_clone[2], src_clone[3], src_clone[4], op_params[0]);
        } else if (tensor->op == GGML_OP_GATED_DELTA_NET) {
            tensor_clone = ggml_gated_delta_net(ggml_ctx, src_clone[0], src_clone[1],
            src_clone[2], src_clone[3], src_clone[4], src_clone[5],
@@ -0,0 +1,82 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+#define BLOCK_SIZE 64
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform Parameters {
+    uint B;
+    uint T;
+    uint C;
+    uint H;
+    float scale;
+};
+
+layout(binding = 0) readonly buffer KBuf { A_TYPE k[]; };
+layout(binding = 1) readonly buffer VBuf { A_TYPE v[]; };
+layout(binding = 2) readonly buffer QBuf { A_TYPE q[]; };
+layout(binding = 3) readonly buffer GBuf { A_TYPE g[]; };
+layout(binding = 4) readonly buffer StateBuf { A_TYPE state_in[]; };
+layout(binding = 5) buffer DstBuf { A_TYPE dst[]; };
+
+shared A_TYPE _k[BLOCK_SIZE], _q[BLOCK_SIZE], _g[BLOCK_SIZE];
+
+void main() {
+    const uint head_size = BLOCK_SIZE;
+    const uint batch_id = gl_WorkGroupID.x / H;
+    const uint head_id = gl_WorkGroupID.x % H;
+    const uint tid = gl_LocalInvocationID.x;
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    // state[i] holds column tid of this head's state matrix: S[i][tid]
+    A_TYPE state[BLOCK_SIZE];
+    [[unroll]] for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        barrier();
+        _k[tid] = k[t];
+        _q[tid] = q[t];
+        _g[tid] = g[t];
+        barrier();
+
+        const A_TYPE v_val = v[t];
+        A_TYPE y = 0.0;
+
+        [[unroll]] for (uint i = 0; i < head_size; i += 4) {
+            vec4 k_vec = vec4(_k[i], _k[i+1], _k[i+2], _k[i+3]);
+            vec4 q_vec = vec4(_q[i], _q[i+1], _q[i+2], _q[i+3]);
+            vec4 g_vec = vec4(_g[i], _g[i+1], _g[i+2], _g[i+3]);
+            vec4 s_vec = vec4(state[i], state[i+1], state[i+2], state[i+3]);
+
+            vec4 kv = k_vec * v_val;
+
+            s_vec = s_vec * g_vec + kv;
+            y += dot(q_vec, s_vec);
+
+            state[i]   = s_vec.x;
+            state[i+1] = s_vec.y;
+            state[i+2] = s_vec.z;
+            state[i+3] = s_vec.w;
+        }
+
+        dst[t] = y * scale;
+    }
+
+    [[unroll]] for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
--- a/Show More
+++ b/Show More