ggml: fix backend split scheduler race condition (#26040 )

* ggml: fix backend split scheduler race condition splits without input were running concurrently with other splits, while potentially reusing memory the other split is accessing * only sync when split has no inputs
convert: fix get block count error for Nemotron 3 Ultra (#27101 )
2026-08-20 11:42:39 +02:00 · 2026-08-20 10:42:33 +02:00 · 2026-08-20 10:35:28 +03:00 · 2026-08-20 10:27:51 +03:00 · 2026-08-20 10:00:35 +03:00 · 2026-08-20 10:00:16 +03:00
282 changed files with 6123 additions and 2565 deletions
@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2.1
-ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.36.3
-ARG IGC_VERSION_FULL=2_2.36.3+21719
-ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.33.0
-ARG NPU_DRIVER_FULL=v1.33.0.20260529-26625960453
-ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+ARG NPU_DRIVER_VERSION=v1.35.0
+ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
+ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1

 # Optional proxy build arguments
 ARG http_proxy=
@@ -170,7 +170,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
    fi; \
    DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
    if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
    fi; \
    mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
    apt-get update; \
@@ -1,20 +0,0 @@
-name: "Linux - Setup Vulkan SDK"
-description: "Setup Vulkan SDK for Linux"
-inputs:
-  path:
-    description: "Installation path"
-    required: true
-  version:
-    description: "Vulkan SDK version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup Vulkan SDK
-      id: setup
-      uses: ./.github/actions/unarchive-tar
-      with:
-        url: https://sdk.lunarg.com/sdk/download/${{ inputs.version }}/linux/vulkan_sdk.tar.xz
-        path: ${{ inputs.path }}
-        strip: 1
@@ -6,8 +6,7 @@ inputs:
    required: true
  cuda_arch:
    description: "CUDA target architecture"
-    required: false
-    default: "x64"
+    required: true

 runs:
  using: "composite"
@@ -10,33 +10,6 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  ubuntu-24-vulkan-cache:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
-
  #ubuntu-24-spacemit-cache:
  #  runs-on: ubuntu-24.04

@@ -67,9 +40,9 @@ jobs:
    runs-on: ubuntu-24.04

    env:
-      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -96,8 +69,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -21,6 +21,7 @@ on:
    paths: [
      '.github/workflows/build-cpu.yml',
      '.github/workflows/build-cmake-pkg.yml',
+      'ggml/src/ggml-rpc/**',
      '**/CMakeLists.txt',
      '**/.cmake',
      '**/*.h',
@@ -123,7 +124,6 @@ jobs:
    env:
      OPENBLAS_VERSION: 0.3.23
      SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.357.0

    strategy:
      matrix:
@@ -134,9 +134,6 @@ jobs:
          - build: 'x64-openblas'
            arch: 'x64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'x64-vulkan'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
          - build: 'arm64'
            arch: 'arm64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
@@ -167,15 +164,6 @@ jobs:
          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll

-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == 'x64-vulkan' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
-          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
-          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
-          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
-
      - name: Install Ninja
        id: install_ninja
        run: |
@@ -22,6 +22,7 @@ env:

 jobs:
  cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    runs-on: windows-2022

    permissions:
@@ -29,7 +30,16 @@ 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
@@ -39,12 +49,13 @@ 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: Install Cuda Toolkit
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -54,26 +65,21 @@ jobs:
      - 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" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DGGML_RPC=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% -t ggml
-          cmake --build build --config Release
+          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 }}

  hip:
    runs-on: windows-2022
@@ -39,8 +39,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -81,7 +81,7 @@ jobs:
        # TODO: fix and re-enable the `test-llama-archs` test below
        run: |
          cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000

      - name: Test (GPU)
        id: cmake_test_gpu
@@ -89,15 +89,15 @@ jobs:
        run: |
          cd ${{ github.workspace }}
          export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000

  openvino-windows-2022:
    runs-on: windows-2022

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -166,4 +166,4 @@ jobs:
          call "%OPENVINO_ROOT%\setupvars.bat"

          cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
@@ -1,66 +0,0 @@
-name: CI (rpc)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-rpc.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-rpc.yml',
-      'ggml/src/ggml-rpc/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  ubuntu-24-rpc:
-    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libssl-dev ninja-build
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose
@@ -288,8 +288,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -93,19 +93,13 @@ jobs:
        run: |
          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"

-      - name: Use Vulkan SDK Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
+        id: setup
+        uses: ./.github/actions/unarchive-tar
        with:
+          url: https://sdk.lunarg.com/sdk/download/${{ env.VULKAN_SDK_VERSION }}/linux/vulkan_sdk.tar.xz
          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
+          strip: 1

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -133,3 +127,56 @@ jobs:
          # This is using llvmpipe and runs slower than other backends
          # test-backend-ops is too slow on llvmpipe, skip it
          ctest -L main -E test-backend-ops --verbose --timeout 900
+
+  windows:
+    runs-on: windows-2025
+
+    env:
+      VULKAN_VERSION: 1.4.357.0
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cpu-windows-2025-x64-vulkan
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" `
+            -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_NATIVE=OFF `
+            -DLLAMA_BUILD_SERVER=ON `
+            -DGGML_RPC=ON `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_VULKAN=ON `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main -C Release --verbose --timeout 900
@@ -394,6 +394,11 @@ jobs:
    name: Create shared tags from digests
    needs: [prepare_matrices, push_to_registry, create_tag]
    runs-on: ubuntu-24.04
+    permissions:
+      contents: read
+      packages: write
+      id-token: write
+      attestations: write
    strategy:
      fail-fast: false
      matrix:
@@ -428,6 +433,7 @@ jobs:
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Create tags from digests
+        id: create_tags
        shell: bash
        run: |
          set -euo pipefail
@@ -439,6 +445,7 @@ jobs:
          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
          BUILD_DATE="${{ steps.build_date.outputs.date }}"
          COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
+          echo "image_repo=${IMAGE_REPO}" >> "$GITHUB_OUTPUT"
          TAGS="${{ matrix.config.tag }}"
          ARCHES="${{ matrix.config.arches }}"
          DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -505,6 +512,16 @@ jobs:

              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
              docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
+
+              if [[ "$tag_name" == "${TAGS%% *}" ]]; then
+                  local digest
+                  digest="$(docker buildx imagetools inspect "${merged_versioned_tag}" --format '{{.Manifest.Digest}}')"
+                  if [[ ! "$digest" =~ ^sha256:[0-9a-f]{64}$ ]]; then
+                      echo "Invalid digest for ${merged_versioned_tag}: ${digest}" >&2
+                      exit 1
+                  fi
+                  echo "${image_type}_digest=${digest}" >> "$GITHUB_OUTPUT"
+              fi
          }

          for tag in $TAGS; do
@@ -528,3 +545,24 @@ jobs:
          done
        env:
          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - name: Attest full image
+        if: ${{ matrix.config.full }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.full_digest }}
+
+      - name: Attest light image
+        if: ${{ matrix.config.light }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.light_digest }}
+
+      - name: Attest server image
+        if: ${{ matrix.config.server }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.server_digest }}
@@ -49,6 +49,33 @@ jobs:
          git push origin "${VERSION}"
          echo "Created and pushed tag ${VERSION}"

+      - name: Generate release description
+        id: desc
+        run: bash scripts/make-release-desc.sh "${{ steps.checks.outputs.version }}"
+        env:
+          GITHUB_REPOSITORY: ${{ github.repository }}
+
+      - name: Create release
+        if: ${{ github.event.inputs.dry_run == 'false' }}
+        uses: ggml-org/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        with:
+          tag_name: ${{ steps.checks.outputs.version }}
+          # TODO: remove the prerelease flag once the semantic versioning workflow is ready
+          # ref: https://github.com/ggml-org/ggml/discussions/1579
+          prerelease: true
+          body: |
+            > [!NOTE]
+            > Semantic versioning is still work in progress.
+            > More info can be found in https://github.com/ggml-org/ggml/discussions/1579
+
+            ${{ steps.desc.outputs.nightly }}
+
+            ## ${{ steps.desc.outputs.changelog_title }}
+
+            ${{ steps.desc.outputs.changelog }}
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
@@ -446,8 +446,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -562,8 +562,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -1439,7 +1439,9 @@ jobs:
      - name: xcodebuild for swift package
        id: xcodebuild
        run: |
-          ./build-xcframework.sh
+          # note: only macos and ios-device due to long build time
+          # ref: https://github.com/ggml-org/llama.cpp/pull/27252
+          ./build-xcframework.sh macos ios-device

      - name: Build Xcode project
        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
@@ -1567,6 +1569,8 @@ jobs:
    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
    permissions:
        contents: write # for creating release
+        id-token: write
+        attestations: write

    runs-on: ubuntu-slim

@@ -1575,14 +1579,14 @@ jobs:
      - windows
      - windows-cpu
      - windows-cuda
-      #- windows-sycl
+      - windows-sycl
      - windows-rocm
      - windows-openvino
      #- ubuntu-22-rocm
      - ubuntu-cpu
      - ubuntu-vulkan
      - ubuntu-24-openvino
-      #- ubuntu-24-sycl
+      - ubuntu-24-sycl
      - android-arm64
      - macos-cpu
      - ios-xcode
@@ -1660,6 +1664,22 @@ jobs:
        run: |
          tar -czvf release/llama-${{ steps.tag.outputs.name }}-ui.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./ui-dist .

+      - name: Attest release artifacts
+        id: attest
+        uses: actions/attest@v4
+        with:
+          subject-path: 'release/*'
+
+      - name: Create and push git tag
+        run: |
+          TAG="${{ steps.tag.outputs.name }}"
+          if git rev-parse -q --verify "refs/tags/${TAG}" >/dev/null 2>&1; then
+            echo "Tag ${TAG} already exists, skipping creation"
+          else
+            git tag "${TAG}"
+            git push origin "${TAG}"
+          fi
+
      - name: Create release
        id: create_release
        uses: ggml-org/action-create-release@v1
@@ -1677,6 +1697,9 @@ jobs:
            **Website:**
            - <https://llama.app>

+            **Attestations:**
+            - <${{ steps.attest.outputs.attestation-url }}>
+
            **macOS/iOS:**
            - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
            - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780)
@@ -2,6 +2,7 @@ You are a coding agent. Here are some very important rules that you must follow:

 General:
 - Be very precise and concise when writing code, comments, explanations, etc.
+- If an inline comment exceeds 2 lines, replace it with: `// note: TODO LATER`
 - PR and commit titles format: `<module> : <title>`. Lookup recents for examples
 - Don't try to build or run the code unless you are explicitly asked to do so
 - Use the `gh` CLI tool when querying PRs, issues, or other GitHub resources
@@ -5,7 +5,7 @@ include(CheckIncludeFileCXX)
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
 set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 0)
+set(LLAMA_VERSION_PATCH 2)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

 # whether this is a development/nightly build
@@ -224,9 +224,10 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #

+add_subdirectory(vendor)
+
 if (LLAMA_BUILD_COMMON)
    add_subdirectory(common)
-    add_subdirectory(vendor/cpp-httplib)
 endif()

 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
@@ -7,10 +7,11 @@
 <b>LLM inference in C/C++</b>

 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
+[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
-[![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)
+[![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
+[![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](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) / [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)

@@ -119,7 +120,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 ## Acknowledgements

 - [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [nothings/stb](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
 - [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
+- [mackron/miniaudio](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
+- [sheredom/subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
@@ -1,5 +1,8 @@
 #!/usr/bin/env bash
 #
+# usage: ./build-xcframework.sh [BUILD ...] (default: all builds)
+# builds: ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device
+#
 # Options
 IOS_MIN_OS_VERSION=16.4
 MACOS_MIN_OS_VERSION=13.3
@@ -19,6 +22,43 @@ GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
 GGML_OPENMP=OFF

+# Max number of concurrent platform builds
+MAX_PARALLEL_BUILDS=1
+
+# Split the available cores between the concurrent builds (min 1)
+JOBS_PER_BUILD=$(( $(sysctl -n hw.logicalcpu) / MAX_PARALLEL_BUILDS ))
+if [[ "$JOBS_PER_BUILD" -lt 1 ]]; then
+    JOBS_PER_BUILD=1
+fi
+
+# echo "build_fn build_dir release_dir platform is_simulator min_os" for a build name
+build_spec() {
+    case "$1" in
+        ios-sim)      echo "build_ios_sim build-ios-sim Release-iphonesimulator ios true ${IOS_MIN_OS_VERSION}" ;;
+        ios-device)   echo "build_ios_device build-ios-device Release-iphoneos ios false ${IOS_MIN_OS_VERSION}" ;;
+        macos)        echo "build_macos build-macos Release macos false ${MACOS_MIN_OS_VERSION}" ;;
+        visionos)     echo "build_visionos build-visionos Release-xros visionos false ${VISIONOS_MIN_OS_VERSION}" ;;
+        visionos-sim) echo "build_visionos_sim build-visionos-sim Release-xrsimulator visionos true ${VISIONOS_MIN_OS_VERSION}" ;;
+        tvos-sim)     echo "build_tvos_sim build-tvos-sim Release-appletvsimulator tvos true ${TVOS_MIN_OS_VERSION}" ;;
+        tvos-device)  echo "build_tvos_device build-tvos-device Release-appletvos tvos false ${TVOS_MIN_OS_VERSION}" ;;
+        *)            return 1 ;;
+    esac
+}
+
+# Default: build everything
+if [[ $# -eq 0 ]]; then
+    BUILDS=(ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device)
+else
+    BUILDS=("$@")
+fi
+for b in "${BUILDS[@]}"; do
+    if ! build_spec "$b" >/dev/null; then
+        echo "Error: unknown build '$b'" >&2
+        echo "Valid builds: ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device" >&2
+        exit 1
+    fi
+done
+
 COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
 COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"

@@ -250,6 +290,7 @@ combine_static_libraries() {
        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
+        "${base_dir}/${build_dir}/vendor/hash/${release_dir}/libvendor-hash.a"
    )

    # Create temporary directory for processing
@@ -401,148 +442,189 @@ combine_static_libraries() {
    rm -rf "${temp_dir}"
 }

-echo "Building for iOS simulator..."
-cmake -B build-ios-sim -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-    -DIOS=ON \
-    -DCMAKE_SYSTEM_NAME=iOS \
-    -DCMAKE_OSX_SYSROOT=iphonesimulator \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_ios_sim() {
+    echo "Building for iOS simulator..."
+    cmake -B build-ios-sim -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+        -DIOS=ON \
+        -DCMAKE_SYSTEM_NAME=iOS \
+        -DCMAKE_OSX_SYSROOT=iphonesimulator \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-ios-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for iOS devices..."
-cmake -B build-ios-device -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=iOS \
-    -DCMAKE_OSX_SYSROOT=iphoneos \
-    -DCMAKE_OSX_ARCHITECTURES="arm64" \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_ios_device() {
+    echo "Building for iOS devices..."
+    cmake -B build-ios-device -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+        -DCMAKE_SYSTEM_NAME=iOS \
+        -DCMAKE_OSX_SYSROOT=iphoneos \
+        -DCMAKE_OSX_ARCHITECTURES="arm64" \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-ios-device --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for macOS..."
-cmake -B build-macos -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -S .
-cmake --build build-macos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_macos() {
+    echo "Building for macOS..."
+    cmake -B build-macos -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -S .
+    cmake --build build-macos --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for visionOS..."
-cmake -B build-visionos -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
-    -DCMAKE_OSX_ARCHITECTURES="arm64" \
-    -DCMAKE_SYSTEM_NAME=visionOS \
-    -DCMAKE_OSX_SYSROOT=xros \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_BUILD_SERVER=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_visionos() {
+    echo "Building for visionOS..."
+    cmake -B build-visionos -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+        -DCMAKE_OSX_ARCHITECTURES="arm64" \
+        -DCMAKE_SYSTEM_NAME=visionOS \
+        -DCMAKE_OSX_SYSROOT=xros \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DLLAMA_BUILD_SERVER=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-visionos --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for visionOS simulator..."
-cmake -B build-visionos-sim -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DCMAKE_SYSTEM_NAME=visionOS \
-    -DCMAKE_OSX_SYSROOT=xrsimulator \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_BUILD_SERVER=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_visionos_sim() {
+    echo "Building for visionOS simulator..."
+    cmake -B build-visionos-sim -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DCMAKE_SYSTEM_NAME=visionOS \
+        -DCMAKE_OSX_SYSROOT=xrsimulator \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DLLAMA_BUILD_SERVER=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-visionos-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

 # Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
-echo "Building for tvOS simulator..."
-cmake -B build-tvos-sim -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=tvOS \
-    -DCMAKE_OSX_SYSROOT=appletvsimulator \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DGGML_METAL=ON \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_tvos_sim() {
+    echo "Building for tvOS simulator..."
+    cmake -B build-tvos-sim -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+        -DCMAKE_SYSTEM_NAME=tvOS \
+        -DCMAKE_OSX_SYSROOT=appletvsimulator \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DGGML_METAL=ON \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-tvos-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for tvOS devices..."
-cmake -B build-tvos-device -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=tvOS \
-    -DCMAKE_OSX_SYSROOT=appletvos \
-    -DCMAKE_OSX_ARCHITECTURES="arm64" \
-    -DGGML_METAL=ON \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_tvos_device() {
+    echo "Building for tvOS devices..."
+    cmake -B build-tvos-device -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+        -DCMAKE_SYSTEM_NAME=tvOS \
+        -DCMAKE_OSX_SYSROOT=appletvos \
+        -DCMAKE_OSX_ARCHITECTURES="arm64" \
+        -DGGML_METAL=ON \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-tvos-device --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}
+
+run_builds_parallel() {
+    local -a pids=()
+    local -a names=()
+    local name i
+    for name in "$@"; do
+        # Wait for the oldest running build to free a slot
+        if [[ "${#pids[@]}" -ge "$MAX_PARALLEL_BUILDS" ]]; then
+            if ! wait "${pids[0]}"; then
+                echo "ERROR: build '${names[0]}' failed, log follows (${names[0]}.log):" >&2
+                kill "${pids[@]}" 2>/dev/null || true
+                cat "${names[0]}.log" >&2
+                exit 1
+            fi
+            pids=("${pids[@]:1}")
+            names=("${names[@]:1}")
+        fi
+        echo "Starting build: $name (log: ${name}.log, -j ${JOBS_PER_BUILD})"
+        "$name" > "${name}.log" 2>&1 &
+        pids+=("$!")
+        names+=("$name")
+    done
+    # Wait for the remaining builds
+    for i in "${!pids[@]}"; do
+        if ! wait "${pids[$i]}"; then
+            echo "ERROR: build '${names[$i]}' failed, log follows (${names[$i]}.log):" >&2
+            kill "${pids[@]}" 2>/dev/null || true
+            cat "${names[$i]}.log" >&2
+            exit 1
+        fi
+    done
+}
+
+BUILD_FNS=()
+for b in "${BUILDS[@]}"; do
+    read -r fn _ < <(build_spec "$b")
+    BUILD_FNS+=("$fn")
+done
+echo "Building: ${BUILDS[*]} (max ${MAX_PARALLEL_BUILDS} at a time, -j ${JOBS_PER_BUILD} each)..."
+run_builds_parallel "${BUILD_FNS[@]}"

 # Setup frameworks and copy binaries and headers
 echo "Setting up framework structures..."
-setup_framework_structure "build-ios-sim" ${IOS_MIN_OS_VERSION} "ios"
-setup_framework_structure "build-ios-device" ${IOS_MIN_OS_VERSION} "ios"
-setup_framework_structure "build-macos" ${MACOS_MIN_OS_VERSION} "macos"
-setup_framework_structure "build-visionos" ${VISIONOS_MIN_OS_VERSION} "visionos"
-setup_framework_structure "build-visionos-sim" ${VISIONOS_MIN_OS_VERSION} "visionos"
-setup_framework_structure "build-tvos-sim" ${TVOS_MIN_OS_VERSION} "tvos"
-setup_framework_structure "build-tvos-device" ${TVOS_MIN_OS_VERSION} "tvos"
+for b in "${BUILDS[@]}"; do
+    read -r _ bdir _ platform _ min_os < <(build_spec "$b")
+    setup_framework_structure "$bdir" "$min_os" "$platform"
+done

 # Create dynamic libraries from static libraries
 echo "Creating dynamic libraries from static libraries..."
-combine_static_libraries "build-ios-sim" "Release-iphonesimulator" "ios" "true"
-combine_static_libraries "build-ios-device" "Release-iphoneos" "ios" "false"
-combine_static_libraries "build-macos" "Release" "macos" "false"
-combine_static_libraries "build-visionos" "Release-xros" "visionos" "false"
-combine_static_libraries "build-visionos-sim" "Release-xrsimulator" "visionos" "true"
-combine_static_libraries "build-tvos-sim" "Release-appletvsimulator" "tvos" "true"
-combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
+for b in "${BUILDS[@]}"; do
+    read -r _ bdir rdir platform is_sim _ < <(build_spec "$b")
+    combine_static_libraries "$bdir" "$rdir" "$platform" "$is_sim"
+done

 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
+XCFW_ARGS=()
+for b in "${BUILDS[@]}"; do
+    read -r _ bdir _ _ _ _ < <(build_spec "$b")
+    XCFW_ARGS+=(-framework "$(pwd)/${bdir}/framework/llama.framework")
+    XCFW_ARGS+=(-debug-symbols "$(pwd)/${bdir}/dSYMs/llama.dSYM")
+done
 xcrun xcodebuild -create-xcframework \
-    -framework $(pwd)/build-ios-sim/framework/llama.framework \
-    -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-ios-device/framework/llama.framework \
-    -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-visionos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-visionos-sim/framework/llama.framework \
-    -debug-symbols $(pwd)/build-visionos-sim/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-tvos-device/framework/llama.framework \
-    -debug-symbols $(pwd)/build-tvos-device/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-tvos-sim/framework/llama.framework \
-    -debug-symbols $(pwd)/build-tvos-sim/dSYMs/llama.dSYM \
-    -output $(pwd)/build-apple/llama.xcframework
+    "${XCFW_ARGS[@]}" \
+    -output "$(pwd)/build-apple/llama.xcframework"
@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

    # TODO: fix and re-enable the `test-llama-archs` test below
-    CTEST_EXTRA="-E test-llama-archs"
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi

 ## helpers
@@ -126,7 +126,8 @@ set_target_properties(${TARGET} PROPERTIES
    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )

-target_include_directories(${TARGET} PUBLIC . ../vendor)
+target_include_directories(${TARGET} PUBLIC .)
+target_link_libraries     (${TARGET} PUBLIC vendor::nlohmann vendor::sheredom)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)

 if (LLAMA_SUBPROCESS)
@@ -3362,7 +3362,7 @@ 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, get_info\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, 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);
@@ -4658,6 +4658,12 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
        [](common_params &, int) { /* unused */ }
    ).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());

+    args.push_back(common_arg(
+        {"dedup-cache-models"}, "0|1",
+        "in server router mode, hide a cached model from the model list when this preset resolves to the same model file",
+        [](common_params &, const std::string &) { /* unused */ }
+    ).set_env(COMMON_ARG_PRESET_DEDUP_CACHE_MODELS).set_preset_only());
+
    // args.push_back(common_arg(
    //     {"pin"},
    //     "in server router mode, do not unload this model if models_max is exceeded",
@@ -11,8 +11,9 @@
 #include <memory>

 // pseudo-env variable to identify preset-only arguments
-#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
-#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_LOAD_ON_STARTUP    "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT       "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"

 //
 // CLI argument parsing
@@ -1778,6 +1778,8 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
    struct ggml_threadpool_params tpp =
            ggml_threadpool_params_from_cpu_params(params.cpuparams);

+    // each pool needs to match the respective n_threads exactly
+    // see: https://github.com/ggml-org/llama.cpp/pull/27138#issuecomment-5332307332
    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
        if (!threadpool_batch) {
@@ -989,6 +989,26 @@ std::vector<common_cached_model_info> common_list_cached_models() {
    return result;
 }

+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file) {
+    auto [repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
+
+    auto files = hf_cache::get_cached_files(repo);
+    if (files.empty()) {
+        return "";
+    }
+
+    if (!hf_file.empty()) {
+        for (const auto & f : files) {
+            if (f.path == hf_file) {
+                return f.local_path;
+            }
+        }
+        return "";
+    }
+
+    return find_best_model(files, tag).local_path;
+}
+
 bool common_download_remove(const std::string & hf_repo_with_tag) {
    namespace fs = std::filesystem;

@@ -85,6 +85,10 @@ std::vector<std::string> common_download_get_all_parts(const std::string & url);
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();

+// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
+// returns an empty string if the model is not present in the cache
+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
+
 // download single file from url to local path
 // returns status code or -1 on error
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
@@ -926,6 +926,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
    const bool is_dspark;

+    // dspark speculators
+    bool sample_from_anchor = true;
+
    const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
    uint32_t        target_layer_ids_n = 0;

@@ -960,16 +963,20 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            if (llama_model_meta_val_str(model_dft, "dflash.block_size", buf, sizeof(buf)) >= 0) {
                block_size = std::atoi(buf);
            }
+            if (llama_model_meta_val_str(model_dft, "dflash.sample_from_anchor", buf, sizeof(buf)) >= 0) {
+                sample_from_anchor = std::strcmp(buf, "true") == 0;
+            }
        }
        mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft));

        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(type).c_str());
        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
-        LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u\n", __func__, block_size, mask_token_id, target_layer_ids_n);
+        LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u, sample_from_anchor=%s\n", __func__,
+                block_size, mask_token_id, target_layer_ids_n, sample_from_anchor ? "true" : "false");

        // DFlash input is [id_last, <mask> * (block_size-1)]: in-place denoising yields at most
-        // block_size-1 draft tokens, DSpark yield a full block_size draft tokens
-        const int32_t n_draft_max = is_dspark ? block_size : block_size - 1;
+        // block_size-1 draft tokens, anchor-first DSpark yields a full block_size draft tokens
+        const int32_t n_draft_max = is_dspark && sample_from_anchor ? block_size : block_size - 1;
        if (this->params.n_max > n_draft_max || this->params.n_min > n_draft_max) {
            LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained block size %d -- clamping to %d\n",
                    __func__, this->params.n_max, this->params.n_min, block_size, n_draft_max);
@@ -1175,7 +1182,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

            const int32_t n_draft = params.n_max;

-            const int32_t n_block_tokens = n_draft + (is_dspark ? 0 : 1);
+            const int32_t n_block_tokens = n_draft + (is_dspark && sample_from_anchor ? 0 : 1);
            i_block_beg[seq_id] = batch.n_tokens;
            n_block    [seq_id] = n_block_tokens;
            for (int32_t i = 0; i < n_block_tokens; ++i) {
@@ -1208,11 +1215,11 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            auto & result = *dp.result;

            if (is_dspark) {
-                // DSpark predicts the next token from position 0 and optionally truncates
-                // at the first position below the confidence threshold.
+                // DSpark: read from the first draft slot, truncate below the confidence threshold
                const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
-
-                for (int32_t i = 0; i < n_block_tokens; ++i) {
+                // bonus-anchor drafts read the mask positions only, like DFlash
+                const int32_t i_draft_beg = sample_from_anchor ? 0 : 1;
+                for (int32_t i = i_draft_beg; i < n_block_tokens; ++i) {
                    const int32_t idx = beg + i;

                    if (conf && conf[(size_t) idx * n_embd_dec] < params.p_min) {
@@ -2642,6 +2649,10 @@ void common_speculative_draft(common_speculative * spec) {
        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) dparams.size(); ++seq_id) {
            auto & dp = dparams[seq_id];

+            if (!dp.drafting) {
+                continue;
+            }
+
            auto & result = *dp.result;

            // a new draft has been sampled
@@ -55,6 +55,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "DeepseekV32ForCausalLM": "deepseek",
    "DFlashDraftModel": "qwen",
    "Qwen3DSparkModel": "qwen",
+    "DSparkDraftModel": "qwen",
+    "DSparkSpeculator": "qwen",
    "DeepseekV4ForCausalLM": "deepseek",
    "DeepseekV4DSparkModel": "deepseek",
    "DistilBertForMaskedLM": "bert",
@@ -107,6 +109,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
+    "GraniteSWAForCausalLM": "granite",
+    "GraniteMoeSWAForCausalLM": "granite",
    "Grok1ForCausalLM": "grok",
    "GrokForCausalLM": "grok",
    "GroveMoeForCausalLM": "grovemoe",
@@ -74,6 +74,108 @@ class GraniteModel(LlamaModel):
        return super().filter_tensors(item)


+@ModelBase.register("GraniteSWAForCausalLM")
+class GraniteSWAModel(GraniteModel):
+    """Conversion for IBM's GraniteSWAForCausalLM (interleaved sliding window attention)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if name.endswith("sinks"):
+            name += ".weight"
+
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        """GraniteSWA uses Granite parameters plus sliding window configuration."""
+        super().set_gguf_parameters()
+
+        # Add sliding_window from config
+        sliding_window = self.hparams.get("sliding_window", 128)
+        self.gguf_writer.add_sliding_window(sliding_window)
+        logger.info("gguf: (granite_swa) sliding_window = %s", sliding_window)
+
+        # Derive sliding_window_pattern from layer_types
+        if layer_types := self.hparams.get("layer_types"):
+            is_swa = [t == "sliding_attention" for t in layer_types]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern = %d SWA layers / %d total",
+                        sum(is_swa), len(is_swa))
+        else:
+            # Fall back to period-based pattern: i % 4 != 0
+            # This matches the transformers default pattern
+            n_layers = self.block_count
+            is_swa = [i % 4 != 0 for i in range(n_layers)]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern (inferred) = %d SWA layers / %d total",
+                        sum(is_swa), n_layers)
+
+        # Add rope_pattern from no_rope_layers
+        if no_rope_layers := self.hparams.get("no_rope_layers"):
+            # Convert 1/0 to bool (1 = use RoPE, 0 = NoPE)
+            rope_pattern = [bool(x) for x in no_rope_layers]
+            self.gguf_writer.add_rope_pattern(rope_pattern)
+            logger.info("gguf: (granite_swa) rope_pattern = %d RoPE layers / %d total",
+                        sum(rope_pattern), len(rope_pattern))
+
+
+@ModelBase.register("GraniteMoeSWAForCausalLM")
+class GraniteMoeSWAModel(GraniteSWAModel):
+    """Conversion for IBM's GraniteMoeSWAForCausalLM (unified dense + MoE with iSWA)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if shared_intermediate_size := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_intermediate_size)
+            logger.info("gguf: (granitemoewa) shared_intermediate_size = %s", shared_intermediate_size)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        """Split merged MoE tensors (gate+up) following standard MoE pattern."""
+
+        # Handle expert FFN tensors (merged gate+up) - swash format: experts.gate_up_proj
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.experts.gate_up_proj"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, f"Merged FFN tensor size must be 2 * intermediate_size, got {data_torch.shape[-2]}"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle expert FFN down projection - swash format: experts.down_proj
+        if name.endswith("block_sparse_moe.experts.down_proj"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_EXP, bid), bid)
+            return
+
+        # Handle expert FFN tensors (merged gate+up) - standard granite format: input_linear.weight
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.input_linear.weight"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle shared expert FFN tensors (if present) - kept fused since
+        # inference (build_ffn) supports a single ffn_up_shexp tensor with
+        # LLM_FFN_SWIGLU for the shared expert.
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams.get("shared_intermediate_size", self.hparams["intermediate_size"])
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid)
+            return
+
+        # Handle shared expert output (if present)
+        if name.endswith("shared_mlp.output_linear.weight"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, bid), bid)
+            return
+
+        # Pass through to parent for all other tensors (including sinks)
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
@ModelBase.example("ibm-granite/granite-3.1-3b-a800m-instruct")
 class GraniteMoeModel(GraniteModel):
@@ -207,7 +207,9 @@ class NemotronHModel(GraniteHybridModel):
        # calling the parent __init__. This is because the parent constructor
        # uses self.model_arch to build the tensor name map, and all MoE-specific
        # mappings would be missed if it were called with the default non-MoE arch.
-        hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
        has_moe_params = (
            "num_experts_per_tok" in hparams
            or (isinstance(hparams.get("llm_config"), dict) and "num_experts_per_tok" in hparams["llm_config"])
@@ -215,8 +217,11 @@ class NemotronHModel(GraniteHybridModel):
        if has_moe_params:
            self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
            self.is_moe = True
+        layers_block_type = hparams.get("layers_block_type")
+        if layers_block_type is not None:
+            hparams["num_hidden_layers"] = len(layers_block_type)

-        super().__init__(*args, **kwargs)
+        super().__init__(*args, hparams=hparams, **kwargs)

        # Save the top-level head_dim for later
        self.head_dim = self.hparams.get("head_dim", self.hparams.get("attention_head_dim"))
@@ -4,12 +4,13 @@ import json

 from typing import Any, Callable, Iterable, TYPE_CHECKING

+import numpy as np
 import torch

 if TYPE_CHECKING:
    from torch import Tensor

-from .base import ModelBase, TextModel, gguf, logger
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf, logger


@ModelBase.register("QWenLMHeadModel")
@@ -708,22 +709,82 @@ class DFlashModel(Qwen3Model):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Qwen3DSparkModel")
+@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator")
@ModelBase.example("satgeze/Qwen3.6-27B-DSpark")
 class DSparkModel(DFlashModel):
-    # DSpark = DFlash + a semi-autoregressive Markov head
+    # DSpark = DFlash + a semi-autoregressive Markov head.
    model_arch = gguf.MODEL_ARCH.DFLASH

-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # normalize the flat DeepSpec schema to DFlash's nested dflash_config
-        self.hparams.setdefault("dflash_config", {
-            k: self.hparams[k] for k in ("target_layer_ids", "mask_token_id") if k in self.hparams
-        })
+    def __init__(self, dir_model, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, False)
+
+        # EAGLE3-style exports use the 1+N bonus-anchor block, DFlash-lineage exports sample from the anchor
+        self._sample_from_anchor = hparams.get(
+            "sample_from_anchor",
+            "transformer_layer_config" not in hparams and "aux_hidden_state_layer_ids" not in hparams)
+        if "transformer_layer_config" in hparams:
+            hparams = {**hparams, **hparams["transformer_layer_config"]}
+
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
+        # normalize both schemas to DFlash's nested dflash_config
+        if "aux_hidden_state_layer_ids" in self.hparams:
+            self.hparams.setdefault("dflash_config", {
+                "mask_token_id": self.hparams.get("mask_token_id"),
+                "target_layer_ids": [i - 1 for i in self.hparams["aux_hidden_state_layer_ids"]],
+            })
+        else:
+            self.hparams.setdefault("dflash_config", {
+                k: self.hparams[k] for k in ("target_layer_ids", "mask_token_id") if k in self.hparams
+            })
+
+        if (markov_head_type := self.hparams.get("markov_head_type", "vanilla")) != "vanilla":
+            raise ValueError(f"unsupported markov_head_type {markov_head_type!r} (only 'vanilla' is supported)")
+
+        n_vocab = self.hparams["vocab_size"]
+        self._n_vocab_draft = self.hparams.get("draft_vocab_size") or n_vocab
+        if self._n_vocab_draft > n_vocab:
+            raise ValueError(f"draft_vocab_size {self._n_vocab_draft} exceeds vocab_size {n_vocab}")
+        self._d2t: Tensor | None = None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_sample_from_anchor(self._sample_from_anchor)

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-        if name.endswith(("embed_tokens.weight", "lm_head.weight")):
+        if item[0] == "t2d":  # not used at runtime
            return None
-        return super().filter_tensors((name, gen))
+        return super().filter_tensors(item)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "model.d2t":
+            self._d2t = data_torch
+            return
+
+        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith(("embed_tokens.weight", "lm_head.weight")):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        n_vocab = self.hparams["vocab_size"]
+        if self._n_vocab_draft < n_vocab and self._d2t is None:
+            raise ValueError(f"draft_vocab_size {self._n_vocab_draft} < vocab_size {n_vocab} but no d2t table found")
+
+        # write d2t as absolute target token ids
+        if self._d2t is not None:
+            data = LazyTorchTensor.to_eager(self._d2t).to(torch.int64).cpu().numpy().reshape(-1)
+            if data.size != self._n_vocab_draft:
+                raise ValueError(f"d2t size {data.size} does not match draft_vocab_size {self._n_vocab_draft}")
+            data = data + np.arange(data.size, dtype=np.int64)
+            if np.any((data < 0) | (data >= n_vocab)):
+                raise ValueError(f"d2t target ids out of range for target vocab size {n_vocab}")
+            if np.unique(data).size != data.size:
+                raise ValueError("d2t contains duplicate target ids")
+            logger.info(f"{'d2t,':<30} --> I64, shape = {{{data.size}}}")
+            self.gguf_writer.add_tensor("d2t", data, raw_dtype=gguf.GGMLQuantizationType.I64)
@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.2.1"
-OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.

 </details>

@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================

-set "OPENVINO_VERSION_MAJOR=2026.2.1"
-set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"

 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.

 </details>

@@ -106,6 +106,10 @@ acceptance (from the draft's confidence head, if present) falls below `P` (defau
 Currently only drafts with a Qwen3 backbone are supported; support for other backbones
 (e.g. Gemma4) is planned.

+DSpark drafts exported in the [speculators](https://github.com/vllm-project/speculators) format
+(for example [`RedHatAI/gemma-4-31B-it-speculator.dspark`](https://huggingface.co/RedHatAI/gemma-4-31B-it-speculator.dspark))
+convert the same way.
+
 See:

 - #25173
@@ -2,21 +2,5 @@ set(TARGET llama-gguf-hash)
 add_executable(${TARGET} gguf-hash.cpp)
 install(TARGETS ${TARGET} RUNTIME)

-# clibs dependencies
-include_directories(deps/)
-
-add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
-target_link_libraries(${TARGET} PRIVATE xxhash)
-
-add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
-target_link_libraries(${TARGET} PRIVATE sha1)
-if (NOT MSVC)
-    # disable warnings in 3rd party code
-    target_compile_options(sha1 PRIVATE -w)
-endif()
-
-add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
-target_link_libraries(${TARGET} PRIVATE sha256)
-
-target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE vendor::hash ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
@@ -1,13 +0,0 @@
-{
-  "name": "rotate-bits",
-  "version": "0.1.1",
-  "repo": "jb55/rotate-bits.h",
-  "description": "rotate bits",
-  "keywords": ["rotl", "rotr"],
-  "src": ["rotate-bits.h"],
-  "license": "Public Domain",
-  "development": {
-    "thlorenz/tap.c": "*"
-  }
-}
-
@@ -1,9 +0,0 @@
-{
-  "name": "sha1",
-  "version": "0.0.1",
-  "repo": "clibs/sha1",
-  "description": "sha1 hash algorithm",
-  "keywords": ["sha1", "hash"],
-  "license": "public domain",
-  "src": ["sha1.c", "sha1.h"]
-}
@@ -1,15 +0,0 @@
-{
-  "name": "sha256",
-  "version": "0.0.2",
-  "repo": "jb55/sha256.c",
-  "description": "sha256 in c",
-  "keywords": ["sha256", "sha2"],
-  "src": ["sha256.c", "sha256.h"],
-  "dependencies": {
-    "jb55/rotate-bits.h": "0.1.1"
-  },
-  "development": {
-    "thlorenz/tap.c": "*"
-  }
-}
-
@@ -1,12 +0,0 @@
-{
-  "name": "xxhash",
-  "version": "0.8.2",
-  "repo": "Cyan4973/xxhash",
-  "description": "Extremely fast non-cryptographic hash algorithm",
-  "keywords": ["xxhash", "hashing"],
-  "license": "BSD-2-Clause",
-  "src": [
-    "xxhash.c",
-    "xxhash.h"
-  ]
-}
@@ -17,14 +17,17 @@
 extern "C" {
 #endif

-#include "xxhash/xxhash.h"
-#include "sha1/sha1.h"
-#include "sha256/sha256.h"
+#include "hash/xxhash/xxhash.h"
+#include "hash/sha256/sha256.h"

 #ifdef __cplusplus
 }
 #endif

+// sha1 is compiled as C++ and lives in a namespace, see scripts/sync_vendor.py
+#include "hash/sha1/sha1.h"
+using namespace vendor_hash;
+

 // uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp')
 #define UUID_NAMESPACE_LLAMA_CPP "ef001206-dadc-5f6d-a15f-3359e577d4e5"
@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_PATCH 2)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -7,7 +7,7 @@ extern "C" {
 #endif

 #define RPC_PROTO_MAJOR_VERSION    5
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    1
 #define RPC_PROTO_PATCH_VERSION    0

 #ifdef  __cplusplus
@@ -1981,6 +1981,14 @@ extern "C" {
            float                 beta_fast,
            float                 beta_slow);

+    // set the offset dims for RoPE
+    // a must be GGML_OP_ROPE or GGML_OP_ROPE_BACK
+    // vision RoPE is not supported
+    // example: (marking: x = rotated, 0 = unrotated)
+    //     n_embd = 10, n_dims = 4, offset = 2 --> [00xxxx0000]
+    GGML_API struct ggml_tensor * ggml_rope_set_offset(
+            struct ggml_tensor  * a,
+            int                   n_offs);

    // clamp
    // in-place, returns view(a)
@@ -83,6 +83,7 @@ extern "C" {
    GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
    GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
    GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API void                  ggml_backend_meta_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);

    //
    // Backend (meta)
@@ -1118,7 +1118,6 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
 }

 static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(const struct ggml_tensor * tensor, bool assume_sync) {
-    GGML_ASSERT(ggml_backend_buffer_is_meta(tensor->buffer));
    ggml_backend_meta_buffer_context * buf_ctx = (ggml_backend_meta_buffer_context *) tensor->buffer->context;
    return ggml_backend_meta_get_split_state(buf_ctx->get_simple_tensor_container(tensor), tensor, assume_sync);
 }
@@ -1178,7 +1177,15 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor_impl(ggml_backend_m
        t_ij->flags = tensor->flags;
        memcpy(t_ij->op_params, tensor->op_params, sizeof(tensor->op_params));
        ggml_set_name(t_ij, tensor->name);
+
        t_ij->buffer = simple_buf;
+        if (simple_buf) {
+            // the backend that owns the buffer will set .extra
+            ggml_backend_buffer_init_tensor(simple_buf, t_ij);
+        } else {
+            t_ij->extra = tensor->extra;
+        }
+
        t_ij->view_src = tensor->view_src;
        t_ij->view_offs = tensor->view_offs;
        if (t_ij->view_src != nullptr && ggml_backend_buffer_is_meta(t_ij->view_src->buffer)) {
@@ -1209,7 +1216,6 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor_impl(ggml_backend_m
            t_ij->data = (char *) ggml_backend_buffer_get_base(simple_buf)
                + size_t(tensor->data) - size_t(ggml_backend_buffer_get_base(tensor->buffer));
        }
-        t_ij->extra = tensor->extra;
        for (int i = 0; i < GGML_MAX_SRC; i++) {
            t_ij->src[i] = tensor->src[i];
            if (tensor->src[i] == tensor) {
@@ -1502,6 +1508,16 @@ bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf) {
    return buf != nullptr && buf->iface.free_buffer == ggml_backend_meta_buffer_iface.free_buffer;
 }

+void ggml_backend_meta_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(ggml_backend_buffer_is_meta(buffer));
+    ggml_backend_meta_buffer_context * buf_ctx = (ggml_backend_meta_buffer_context *) buffer->context;
+    for (size_t i = 0; i < buf_ctx->bufs.size(); i++) {
+        if (buf_ctx->bufs[i]) {
+            ggml_backend_buffer_set_usage(buf_ctx->bufs[i].get(), usage);
+        }
+    }
+}
+
 static ggml_backend_buffer_t ggml_backend_meta_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    const size_t n_simple_bufts = ggml_backend_meta_buft_n_bufts(buft);

@@ -182,6 +182,8 @@ void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backe
    // FIXME: add a generic callback to the buffer interface
    if (ggml_backend_buffer_is_multi_buffer(buffer)) {
        ggml_backend_multi_buffer_set_usage(buffer, usage);
+    } else if (ggml_backend_buffer_is_meta(buffer)) {
+        ggml_backend_meta_buffer_set_usage(buffer, usage);
    }
 }

@@ -1599,11 +1601,23 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
    std::vector<int32_t> ids;
    std::vector<ggml_bitset_t> used_ids;

+    int prev_backend_id = -1;
+
    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
        struct ggml_backend_sched_split * split = &splits[split_id];
        int split_backend_id = split->backend_id;
        ggml_backend_t split_backend = sched->backends[split_backend_id];

+        // ensure the previous split's async work has completed before we start
+        // this split, the allocator may have reused buffer regions across splits
+        if (split->n_inputs == 0 && prev_backend_id >= 0 && prev_backend_id != split_backend_id) {
+            if (sched->events[prev_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_synchronize(sched->events[prev_backend_id][sched->cur_copy]);
+            } else {
+                ggml_backend_synchronize(sched->backends[prev_backend_id]);
+            }
+        }
+
        // copy the input tensors to the split backend
        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1766,12 +1780,12 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
            }
        }

-        // record the event of this copy
-        if (split->n_inputs > 0) {
-            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
-            }
+        // record the event of this split
+        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+            ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
        }
+
+        prev_backend_id = split_backend_id;
    }

    return GGML_STATUS_SUCCESS;
@@ -2534,6 +2534,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
            }
        case GGML_OP_ROPE:
            {
+                if (((const int32_t *) op->op_params)[15] != 0) {
+                    return false; // FIXME: support ggml_rope_set_offset
+                }
                if (op->src[0]->ne[0] > 896) {
                    return false;
                }
@@ -5979,6 +5979,8 @@ static void ggml_compute_forward_rope_flt(
    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);

+    const int n_offs = ((int32_t *) dst->op_params)[15];
+
    GGML_TENSOR_UNARY_OP_LOCALS

    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
@@ -5995,6 +5997,10 @@ static void ggml_compute_forward_rope_flt(
    GGML_ASSERT(n_dims <= ne0);
    GGML_ASSERT(n_dims % 2 == 0);

+    GGML_ASSERT(n_offs >= 0);
+    GGML_ASSERT(n_offs % 2 == 0);
+    GGML_ASSERT(n_offs + n_dims <= ne0);
+
    // rows per thread
    const int dr = (nr + nth - 1)/nth;

@@ -6020,6 +6026,7 @@ static void ggml_compute_forward_rope_flt(

    if (is_vision) {
        GGML_ASSERT(n_dims == ne0/2);
+        GGML_ASSERT(n_offs == 0);
    }

    const float * freq_factors = NULL;
@@ -6068,12 +6075,12 @@ static void ggml_compute_forward_rope_flt(

                switch (mode) {
                    case GGML_ROPE_TYPE_NORMAL:
-                        rotate_pairs<T>(n_dims, 1, cache, src, dst_data, 1);
+                        rotate_pairs<T>(n_dims, 1, cache, src + n_offs, dst_data + n_offs, 1);
                        break;
                    case GGML_ROPE_TYPE_NEOX:
                    case GGML_ROPE_TYPE_MROPE:
                    case GGML_ROPE_TYPE_IMROPE:
-                        rotate_pairs<T>(n_dims, n_dims/2, cache, src, dst_data);
+                        rotate_pairs<T>(n_dims, n_dims/2, cache, src + n_offs, dst_data + n_offs);
                        break;
                    case GGML_ROPE_TYPE_VISION:
                        rotate_pairs<T>(ne0, n_dims, cache, src, dst_data);
@@ -6084,7 +6091,11 @@ static void ggml_compute_forward_rope_flt(

                if (!is_vision) {
                    // fill the remain channels with data from src tensor
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+                        if (i0 == n_offs) {
+                            i0 += n_dims - 2; // skip the rotated channels
+                            continue;
+                        }
                        const T * const src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                        T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);

@@ -29,13 +29,15 @@ extern "C" {
 // FP16 to FP32 conversion

 // 16-bit float
-// on Arm, we use __fp16
+// on Arm, we use __fp16, which requires the IEEE fp16 format: implied on
+// AArch64, selected by -mfp16-format=ieee on 32 bit Arm, where the compiler
+// may otherwise reject the type
 // on x86, we use uint16_t
 //
 // for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
 // for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
 //
-#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+#if defined(__ARM_NEON) && defined(__ARM_FP16_FORMAT_IEEE) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)

@@ -326,7 +328,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
    #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
 #endif

-#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) && defined(__ARM_FP16_FORMAT_IEEE)

 #define GGML_SIMD

@@ -1418,7 +1418,9 @@ struct ggml_backend_cuda_context {
    cudaEvent_t copy_event = nullptr;

    cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
-    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
+    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {nullptr};
+    void * cublas_workspaces[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {nullptr};
+    size_t cublas_workspace_sizes[GGML_CUDA_MAX_DEVICES] = {0};

    int curr_stream_no = 0;

@@ -1495,17 +1497,22 @@ struct ggml_backend_cuda_context {

    ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }

-    cublasHandle_t cublas_handle(int device) {
-        if (cublas_handles[device] == nullptr) {
-            ggml_cuda_set_device(device);
-            CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
-            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
-        }
-        return cublas_handles[device];
-    }
-
    cublasHandle_t cublas_handle() {
-        return cublas_handle(device);
+        if (cublas_handles[device][curr_stream_no] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUBLAS_CHECK(cublasCreate(&cublas_handles[device][curr_stream_no]));
+            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device][curr_stream_no], CUBLAS_TF32_TENSOR_OP_MATH));
+            CUBLAS_CHECK(cublasSetStream(cublas_handles[device][curr_stream_no], stream()));
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUBLAS_VER_MAJOR > 11 || (CUBLAS_VER_MAJOR == 11 && CUBLAS_VER_MINOR >= 2))
+            if (cublas_workspace_sizes[device] == 0) {
+                const int cc = ggml_cuda_info().devices[device].cc;
+                cublas_workspace_sizes[device] = (cc >= GGML_CUDA_CC_HOPPER) ? 32 * 1024 * 1024 : 4 * 1024 * 1024;
+            }
+            CUDA_CHECK(cudaMalloc(&cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
+            CUBLAS_CHECK(cublasSetWorkspace(cublas_handles[device][curr_stream_no], cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
+#endif
+        }
+        return cublas_handles[device][curr_stream_no];
    }

    // pool
@@ -711,9 +711,12 @@ ggml_backend_cuda_context::~ggml_backend_cuda_context() {
            if (streams[i][j] != nullptr) {
                CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
            }
-        }
-        if (cublas_handles[i] != nullptr) {
-            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+            if (cublas_handles[i][j] != nullptr) {
+                CUBLAS_CHECK(cublasDestroy(cublas_handles[i][j]));
+            }
+            if (cublas_workspaces[i][j] != nullptr) {
+                CUDA_CHECK(cudaFree(cublas_workspaces[i][j]));
+            }
        }
    }
 }
@@ -1416,7 +1419,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const

    const int64_t ne_dst = ggml_nelements(dst);
    cudaStream_t main_stream = ctx.stream();
-    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
+    cublasHandle_t cublas_h = ctx.cublas_handle();

    const size_t src0_ts = ggml_type_size(src0->type);
    GGML_ASSERT(nb00 == src0_ts);
@@ -1539,14 +1542,14 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
    //     probably because the internal kernel selection logic is suboptimal.
    if (compute_type == GGML_TYPE_F32 && ne12 == 1 && ne13 == 1) {
        CUBLAS_CHECK(
-            cublasSgemm(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasSgemm(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                    ne01, ne11, ne10,
                    (const float *) alpha, (const float *) src0_ptr, s01,
                                           (const float *) src1_ptr, s11,
                    (const float *) beta,  (float       *)  dst_ptr, ne0));
    } else if (ne12 == 1 && ne13 == 1) {
        CUBLAS_CHECK(
-            cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasGemmEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                    ne01, ne11, ne10,
                    alpha, src0_ptr, cu_data_type_a, s01,
                           src1_ptr, cu_data_type_b, s11,
@@ -1561,7 +1564,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
        // use cublasGemmStridedBatchedEx
        CUBLAS_CHECK(
-        cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmStridedBatchedEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, src0_ptr, cu_data_type_a, s01, sma,     // strideA
                       src1_ptr, cu_data_type_b, s11, smb,     // strideB
@@ -1599,7 +1602,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
        CUDA_CHECK(cudaGetLastError());

        CUBLAS_CHECK(
-        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmBatchedEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, s01,
                       (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
@@ -2723,6 +2726,12 @@ static bool ggml_cuda_should_fuse_rms_norm_mul_rope(const ggml_tensor * rms_norm
        return false;
    }

+    // ggml_rope_set_offset is not yet supported in the fused kernel
+    const int n_offs = ((const int32_t *) rope->op_params)[15];
+    if (n_offs != 0) {
+        return false;
+    }
+
    return true;
 }

@@ -4770,7 +4779,7 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
    }

 // ref: https://github.com/ggml-org/llama.cpp/pull/17368
-#if defined(__linux__)
+#if defined(__linux__) && !defined(GGML_USE_HIP)
    // Check if this is a UMA (Unified Memory Architecture) system
    cudaDeviceProp prop;
    CUDA_CHECK(cudaGetDeviceProperties(&prop, ggml_cuda_get_physical_device(ctx->device)));
@@ -4790,7 +4799,7 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
            GGML_LOG_ERROR("%s: /proc/meminfo reading failed, using cudaMemGetInfo\n", __func__);
        }
    }
-#endif // defined(__linux__)
+#endif // defined(__linux__) && !defined(GGML_USE_HIP)

    // virtual devices sharing one physical GPU share its memory pool; split it between them
    const int share_count = ggml_cuda_physical_device_share_count(ctx->device);
@@ -4,6 +4,7 @@
 #include "vecdotq.cuh"

 #include <cstdint>
+#include <type_traits>

 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);

@@ -69,7 +70,8 @@ enum mmvq_parameter_table_id {
    MMVQ_PARAMETERS_GCN,
    MMVQ_PARAMETERS_RDNA2,
    MMVQ_PARAMETERS_RDNA3_0,
-    MMVQ_PARAMETERS_RDNA4
+    MMVQ_PARAMETERS_RDNA4,
+    MMVQ_PARAMETERS_GB10
 };

 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
@@ -83,6 +85,8 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
    return MMVQ_PARAMETERS_GCN;
 #elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING && __CUDA_ARCH__ < GGML_CUDA_CC_AMPERE
    return MMVQ_PARAMETERS_TURING;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+    return MMVQ_PARAMETERS_GB10;
 #else
    return MMVQ_PARAMETERS_GENERIC;
 #endif
@@ -104,6 +108,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING && ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_AMPERE) {
        return MMVQ_PARAMETERS_TURING;
    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_DGX_SPARK) {
+        return MMVQ_PARAMETERS_GB10;
+    }
    return MMVQ_PARAMETERS_GENERIC;
 }

@@ -351,7 +358,7 @@ static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
 #endif
 }

-static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id, bool small_k = false, bool halve_iters = false) {
    if (table_id == MMVQ_PARAMETERS_GENERIC) {
        switch (ncols_dst) {
            case 1:
@@ -454,11 +461,32 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
                return 1;
        }
    }
+    if (table_id == MMVQ_PARAMETERS_GB10) {
+        const int generic = calc_nwarps(type, ncols_dst, MMVQ_PARAMETERS_GENERIC);
+        // Only worth the wider block when it actually retires the K loop in half the trips (Observation)
+        if (ncols_dst == 1 && !small_k && halve_iters) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                    return 2 * generic;
+                default:
+                    break;
+            }
+        }
+        return generic;
+    }
    return 1;
 }

 static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
-    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING || table_id == MMVQ_PARAMETERS_GB10) {
        switch (ncols_dst) {
            case 1:
                return small_k ? nwarps : 1;
@@ -477,8 +505,8 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
    return 1;
 }

-template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
-__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false, bool halve_iters = false>
+__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id(), small_k, halve_iters)*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
        const void * vx_ptr, const void * vy_ptr, const int32_t * ids_ptr, const ggml_cuda_mm_fusion_args_device fusion, float * dst_ptr,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -495,7 +523,7 @@ static __global__ void mul_mat_vec_q(
    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
    constexpr int vdr = get_vdr_mmvq(type);
    constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
+    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id, small_k, halve_iters);
    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();

@@ -773,8 +801,8 @@ static __global__ void mul_mat_vec_q_moe(
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
-        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false) {
-    const int nwarps = calc_nwarps(type, ncols_dst, table_id);
+        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false, const bool halve_iters = false) {
+    const int nwarps = calc_nwarps(type, ncols_dst, table_id, small_k, halve_iters);
    const int rpb = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
    const int64_t nblocks = (nrows_x + rpb - 1) / rpb;
    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
@@ -782,7 +810,7 @@ static std::pair<dim3, dim3> calc_launch_params(
    return {block_nums, block_dims};
 }

-template<ggml_type type, int c_ncols_dst, bool small_k = false>
+template<ggml_type type, int c_ncols_dst, bool small_k = false, bool halve_iters = false>
 static void mul_mat_vec_q_switch_fusion(
        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -797,7 +825,7 @@ static void mul_mat_vec_q_switch_fusion(
    if constexpr (c_ncols_dst == 1) {
        if (has_fusion) {
            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
-            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k>, launch_params,
+            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k, halve_iters>, launch_params,
                 vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -808,7 +836,7 @@ static void mul_mat_vec_q_switch_fusion(
    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");

    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
-    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k>, launch_params,
+    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k, halve_iters>, launch_params,
        vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -860,16 +888,18 @@ static void mul_mat_vec_q_switch_ncols_dst(

    const bool has_ids = ids != nullptr;

+    // How the K loop divides up at the baseline block width, both decisions below use these.
+    constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr                   = get_vdr_mmvq(type);
+    const int     blocks_per_row_x      = ncols_x / qk;
+    const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+
    const auto should_use_small_k = [&](int c_ncols_dst) {
        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
-        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
-        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
-        constexpr int vdr                   = get_vdr_mmvq(type);
-        const int     blocks_per_row_x      = ncols_x / qk;
-        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
-        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
-        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+        const int  nwarps = calc_nwarps(type, c_ncols_dst, table_id);
+        bool       use    = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;

        constexpr std::array<ggml_type, 2> iq_slow_turing = {
            GGML_TYPE_IQ3_XXS,
@@ -902,6 +932,28 @@ static void mul_mat_vec_q_switch_ncols_dst(
        return use;
    };

+    // Whether doubling nwarps pays off on the ncols_dst == 1 path, where K sets the K loop trip count.
+    const auto should_halve_iters = [&] {
+        if (table_id != MMVQ_PARAMETERS_GB10) {
+            return false;
+        }
+
+        // Expert rows are gathered per token, so a wider block adds reduction work without reuse.
+        if (has_ids) {
+            return false;
+        }
+
+        const int blocks_per_iter = calc_nwarps(type, 1, table_id) * blocks_per_iter_1warp;
+        const int iters           = (blocks_per_row_x + blocks_per_iter - 1) /  blocks_per_iter;
+        const int iters_wide      = (blocks_per_row_x + blocks_per_iter * 2 - 1) / (blocks_per_iter * 2);
+
+        // An odd trip count leaves half the wider block idle for its last iteration, that tail is
+        // only affordable once the loop is long enough to dilute it to an eighth of the work (observation).
+        const int idle = iters_wide * 2 - iters;
+
+        return idle * 8 <= iters_wide * 2;
+    };
+
    if (has_ids && ncols_dst > 1) {
        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
        mul_mat_vec_q_moe_launch<type>(
@@ -914,26 +966,34 @@ static void mul_mat_vec_q_switch_ncols_dst(

    switch (ncols_dst) {
        case 1: {
-            constexpr int c_ncols_dst = 1;
+            // static, else MSVC lambda capture breaks the constexpr uses below
+            static constexpr int c_ncols_dst = 1;

-            bool use_small_k = should_use_small_k(c_ncols_dst);
+            // Tag types keep the flags compile-time, so __launch_bounds__ matches what is launched.
+            const auto launch = [&](auto small_k_tag, auto halve_iters_tag) {
+                constexpr bool c_small_k = decltype(small_k_tag)::value;
+                // Types the table does not promote would compile a second, identical kernel.
+                constexpr bool c_promoted =
+                    calc_nwarps(type, c_ncols_dst, MMVQ_PARAMETERS_GB10, false, true) !=
+                    calc_nwarps(type, c_ncols_dst, MMVQ_PARAMETERS_GB10, false, false);

-            if (use_small_k) {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
-                                                                        nsamples_dst, warp_size, table_id, true);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
+                constexpr bool c_halve_iters = decltype(halve_iters_tag)::value && c_promoted;
+
+                const std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                              nsamples_dst, warp_size, table_id, c_small_k, c_halve_iters);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, c_small_k, c_halve_iters>(
                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
                    stream);
+            };
+
+            if (should_use_small_k(c_ncols_dst)) {
+                launch(std::true_type{},  std::false_type{});
+            } else if (should_halve_iters()) {
+                launch(std::false_type{}, std::true_type{});
            } else {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
-                                                                        nsamples_dst, warp_size, table_id);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
-                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
-                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
-                    stream);
+                launch(std::false_type{}, std::false_type{});
            }
        } break;
        case 2: {
@@ -54,8 +54,6 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const float alpha = 1.0f;
    const float beta = 0.0f;

-    CUBLAS_CHECK(cublasSetStream(handle, stream));
-
    const int64_t lda = nb01 / sizeof(float);
    const int64_t ldc = nb1  / sizeof(float);

@@ -53,6 +53,7 @@ static __global__ void rope_norm(const T *            x,
                                 const int            s2,
                                 const int            s3,
                                 const int            n_dims,
+                                 const int            n_offs,
                                 const int32_t *      pos,
                                 const float          freq_scale,
                                 const float          ext_factor,
@@ -61,7 +62,8 @@ static __global__ void rope_norm(const T *            x,
                                 const float          theta_scale,
                                 const float *        freq_factors,
                                 const int64_t *      row_indices,
-                                 const int            set_rows_stride) {
+                                 const int            set_rows_stride,
+                                 const bool           inplace) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

    if (i0 >= ne00) {
@@ -92,19 +94,24 @@ static __global__ void rope_norm(const T *            x,
            ggml_cuda_memcpy_1<4>(dst + idst, &v);
        }
    };
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        store_coaelsced(x[ix + 0], x[ix + 1]);
        return;
    }

-    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+    const int iw = i0 - n_offs; // relative idx

-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float theta_base = pos[i2]*powf(theta_scale, iw/2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw/2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor, cos_theta, sin_theta);

    const float x0 = x[ix + 0];
    const float x1 = x[ix + 1];
@@ -125,6 +132,7 @@ static __global__ void rope_neox(const T *            x,
                                 const int            s2,
                                 const int            s3,
                                 const int            n_dims,
+                                 const int            n_offs,
                                 const int32_t *      pos,
                                 const float          freq_scale,
                                 const float          ext_factor,
@@ -133,7 +141,8 @@ static __global__ void rope_neox(const T *            x,
                                 const float          theta_scale,
                                 const float *        freq_factors,
                                 const int64_t *      row_indices,
-                                 const int            set_rows_stride) {
+                                 const int            set_rows_stride,
+                                 const bool           inplace) {
    ggml_cuda_pdl_lc();
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

@@ -158,27 +167,33 @@ static __global__ void rope_neox(const T *            x,
        idst += row_indices[i2] * set_rows_stride;
    }

-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        dst[idst + i0 / 2 + 0] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 0]);
        dst[idst + i0 / 2 + 1] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 1]);

        return;
    }

-    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+    const int iw = i0 - n_offs; // relative idx

-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float theta_base = pos[i2]*powf(theta_scale, iw/2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw/2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor, cos_theta, sin_theta);

-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims/2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs/2 + 0];
+    const float x1 = x[ix + n_offs/2 + n_dims/2];

-    dst[idst + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
-    dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
+    dst[idst + n_offs/2 + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_offs/2 + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

 template <bool forward, bool has_ff, typename T>
@@ -194,6 +209,7 @@ static __global__ void rope_multi(const T *            x,
                                  const int            s2,
                                  const int            s3,
                                  const int            n_dims,
+                                  const int            n_offs,
                                  const int32_t *      pos,
                                  const float          freq_scale,
                                  const float          ext_factor,
@@ -202,7 +218,8 @@ static __global__ void rope_multi(const T *            x,
                                  const float          theta_scale,
                                  const float *        freq_factors,
                                  const mrope_sections sections,
-                                  const bool           is_imrope) {
+                                  const bool           is_imrope,
+                                  const bool           inplace) {
    const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);

    if (i0 >= ne00) {
@@ -219,52 +236,58 @@ static __global__ void rope_multi(const T *            x,
    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    ggml_cuda_pdl_sync();
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];

        return;
    }

+    const int iw = i0 - n_offs; // relative idx
+
    const int sect_dims = sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
    const int sec_w = sections.v[1] + sections.v[0];
-    const int sector = (i0 / 2) % sect_dims;
+    const int sector = (iw / 2) % sect_dims;

    float theta_base = 0.0;
    if (is_imrope) {
        if (sector % 3 == 1 && sector < 3 * sections.v[1]) {         // h
-            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {  // w
-            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {  // t
-            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * powf(theta_scale, iw / 2.0f);
        } else {
-            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, iw / 2.0f);
        }
    } else {
        if (sector < sections.v[0]) {
-            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, iw / 2.0f);
        }
    }

-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float freq_factor = has_ff ? freq_factors[iw/2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor, cos_theta, sin_theta);

-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims/2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs/2 + 0];
+    const float x1 = x[ix + n_offs/2 + n_dims/2];

-    dst[idst + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
+    dst[idst + n_offs/2 + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[idst + n_offs/2 + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }

 template <bool forward, bool has_ff, typename T>
@@ -344,6 +367,7 @@ static void rope_norm_cuda(const T *            x,
                           const int            s2,
                           const int            s3,
                           const int            n_dims,
+                           const int            n_offs,
                           const int            nr,
                           const int32_t *      pos,
                           const float          freq_scale,
@@ -354,6 +378,7 @@ static void rope_norm_cuda(const T *            x,
                           const float *        freq_factors,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
+                           const bool           inplace,
                           cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -364,12 +389,12 @@ static void rope_norm_cuda(const T *            x,

    if (freq_factors == nullptr) {
        rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    } else {
        rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    }
 }

@@ -386,6 +411,7 @@ static void rope_neox_cuda(const T *            x,
                           const int            s2,
                           const int            s3,
                           const int            n_dims,
+                           const int            n_offs,
                           const int            nr,
                           const int32_t *      pos,
                           const float          freq_scale,
@@ -396,6 +422,7 @@ static void rope_neox_cuda(const T *            x,
                           const float *        freq_factors,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
+                           const bool           inplace,
                           cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -407,12 +434,12 @@ static void rope_neox_cuda(const T *            x,

    if (freq_factors == nullptr) {
        ggml_cuda_kernel_launch(rope_neox<forward, false, T, D>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    } else {
        ggml_cuda_kernel_launch(rope_neox<forward, true, T, D>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    }
 }

@@ -429,6 +456,7 @@ static void rope_multi_cuda(const T *            x,
                            const int            s2,
                            const int            s3,
                            const int            n_dims,
+                            const int            n_offs,
                            const int            nr,
                            const int32_t *      pos,
                            const float          freq_scale,
@@ -439,6 +467,7 @@ static void rope_multi_cuda(const T *            x,
                            const float *        freq_factors,
                            const mrope_sections sections,
                            const bool           is_imrope,
+                            const bool           inplace,
                            cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -450,13 +479,13 @@ static void rope_multi_cuda(const T *            x,
    if (freq_factors == nullptr) {
        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
        ggml_cuda_kernel_launch(rope_multi<forward, false, T>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope, inplace);
    } else {
        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
        ggml_cuda_kernel_launch(rope_multi<forward, true, T>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope, inplace);
    }
 }

@@ -552,8 +581,12 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    const int mode       = ((int32_t *) dst->op_params)[2];
    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];
    mrope_sections sections;

+    // when dst aliases src0, the channels outside the rotated window already hold the correct data
+    const bool inplace = dst_d == src0->data;
+
    // RoPE alteration for extended context
    float freq_base;
    float freq_scale;
@@ -581,6 +614,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,

    if (is_vision) {
        GGML_ASSERT(n_dims == ne00/2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
    }

    const int32_t * pos = (const int32_t *) src1_d;
@@ -597,31 +631,31 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    if (is_neox) {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                  set_rows_stride, stream);
+                                                  set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                 set_rows_stride, stream);
+                                                 set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                set_rows_stride, stream);
+                                                set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
    } else if (is_mrope && !is_vision) {
        if (src0->type == GGML_TYPE_F32) {
            rope_multi_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, stream);
+                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16) {
            rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, stream);
+                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -640,19 +674,19 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    } else {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                  set_rows_stride, stream);
+                                                  set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                 set_rows_stride, stream);
+                                                 set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                set_rows_stride, stream);
+                                                set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -65,15 +65,13 @@ static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
    get_batch_pointers<<<(total_batches + 255) / 256, 256, 0, stream>>>(A, X, A_ptrs_dev, X_ptrs_dev, ne02,
                                                                        total_batches, s02, s03, s2, s3);

-    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
-
    // Yes, this is necessary, without this we get RMSE errors
-    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_DEFAULT_MATH));
-    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(id), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(), CUBLAS_DEFAULT_MATH));
+    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
                                    CUBLAS_DIAG_NON_UNIT, k, n, &alpha, A_ptrs_dev, n, X_ptrs_dev, k, total_batches));

    // revert to standard mode from common.cuh
-    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_TF32_TENSOR_OP_MATH));
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(), CUBLAS_TF32_TENSOR_OP_MATH));

    GGML_UNUSED_VARS(s12, s13);
 }
@@ -632,7 +632,6 @@ static void ssm_scan_ssd_f32_cuda(
    // Step 3: chunked SSD loop
    // Per chunk: pre_matmul (incl. M) + 4 cuBLAS (CB, Y, S@C, state update) + scale_state
    cublasHandle_t handle = ctx.cublas_handle();
-    CUBLAS_CHECK(cublasSetStream(handle, stream));
    const float alpha_one  = 1.0f;
    const float beta_zero  = 0.0f;
    const float beta_one   = 1.0f;
@@ -1061,9 +1061,11 @@ static bool ggml_backend_et_device_supports_op(ggml_backend_dev_t dev, const ggm
                const bool zero_view_offset = op->src[0]->view_src == nullptr || op->src[0]->view_offs == 0;
                const bool has_sections = ggml_get_op_params_i32(op, 11) > 0 || ggml_get_op_params_i32(op, 12) > 0 ||
                                          ggml_get_op_params_i32(op, 13) > 0;
+                // FIXME: support ggml_rope_set_offset
+                const bool zero_rot_offset  = ggml_get_op_params_i32(op, 15) == 0;

                supported =
-                    zero_view_offset && ndims <= 512 &&
+                    zero_view_offset && zero_rot_offset && ndims <= 512 &&
                    (is_normal || (is_neox && ndims % 16 == 0) || (is_imrope && ndims % 16 == 0 && has_sections));
            } else {
                supported = false;
@@ -3180,6 +3180,10 @@ static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * s
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
    const int32_t * op_params = &op->op_params[0];

+    if (op_params[15] != 0) {
+        return false; // FIXME: support ggml_rope_set_offset
+    }
+
    int mode = op_params[2];

    // n_dims == ne0/2, so the rotation spans the full row
@@ -132,8 +132,8 @@ struct hmx_fa_context {
    __fp16 *     vtcm_v_tiles[2];      // V tiles (column-major, double-buffered)
    __fp16 *     vtcm_s_tiles[2];      // S = QK^T [g_br, Bc] (double-buffered)
    __fp16 *     vtcm_p_tiles[2];      // P = softmax(S) [g_br, Bc]
-    __fp16 *     vtcm_d_tiles;         // Diagonal rescale [g_br, g_br]
-    __fp16 *     vtcm_d_inv_l;         // Diagonal rescale (1/l) [g_br, g_br]
+    __fp16 *     vtcm_d_tiles[2];      // Diagonal rescale, g_br/32 packed diagonal tiles (double-buffered)
+    __fp16 *     vtcm_d_inv_l;         // Diagonal rescale (1/l), same packed layout
    HVX_Vector * vtcm_m_vec;           // Row max [g_br]
    HVX_Vector * vtcm_l_vec;           // Row sum [g_br]
    HVX_Vector * vtcm_s_rowmax;        // Softmax intermediate [g_br]
@@ -782,13 +782,14 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
            }
        }

-        // Initialize vtcm_d_tiles and vtcm_d_inv_l to 0
+        // Zero the whole rescale region: vtcm_d_tiles[0], the optional vtcm_d_tiles[1]
+        // and vtcm_d_inv_l are equal-sized and allocated back to back, so one run covers
+        // them all.  The scatter only ever writes the diagonal, ignore the rest.
        const size_t d_bytes_per_t = hex_align_up(d_tile_bytes / n, 128);
        const size_t d_start       = i * d_bytes_per_t;
        const size_t d_end         = hex_smin(d_start + d_bytes_per_t, d_tile_bytes);
        if (d_start < d_tile_bytes) {
-            hvx_splat_u8_a((char *) factx->vtcm_d_tiles + d_start, 0, d_end - d_start);
-            hvx_splat_u8_a((char *) factx->vtcm_d_inv_l + d_start, 0, d_end - d_start);
+            hvx_splat_u8_a((char *) factx->vtcm_d_tiles[0] + d_start, 0, d_end - d_start);
        }
    }

@@ -1432,17 +1433,19 @@ static inline void fa_softmax_impl(
        const HVX_VectorPred q_32_mask = Q6_Q_vsetq_R(32 * sizeof(__fp16));
        HVX_Vector           v_exp_m_diff = exp_m_diff_f16;

+        __fp16 * const d_tiles_out = factx->vtcm_d_tiles[args->buf_idx];
+
        size_t t0 = r_vec_idx * 2;
        if (t0 < args->n_row_tiles) {
            const HVX_Vector v_content = v_exp_m_diff;
-            __fp16 *         out_base  = factx->vtcm_d_tiles + t0 * (args->n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+            __fp16 *         out_base  = d_tiles_out + t0 * HMX_FP16_TILE_N_ELMS;
            Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
        }

        size_t t1 = r_vec_idx * 2 + 1;
        if (t1 < args->n_row_tiles) {
            const HVX_Vector v_content = Q6_V_vror_VR(v_exp_m_diff, 64);
-            __fp16 *         out_base  = factx->vtcm_d_tiles + t1 * (args->n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+            __fp16 *         out_base  = d_tiles_out + t1 * HMX_FP16_TILE_N_ELMS;
            Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
        }
    }
@@ -1506,7 +1509,7 @@ static __attribute__((noinline)) void fa_build_d_diag_inv_l(struct hmx_fa_contex
            v_content = Q6_V_vror_VR(v_content, 64);
        }

-        __fp16 * out_base = factx->vtcm_d_inv_l + i * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        __fp16 * out_base = factx->vtcm_d_inv_l + i * HMX_FP16_TILE_N_ELMS;
        Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
    }
 }
@@ -1615,7 +1618,7 @@ static void hmx_fa_o_update_worker(void * data) {
    const size_t o_stride = n_row_tiles_g_br * HMX_FP16_TILE_N_ELMS;
    const size_t v_stride = n_tiles_per_bc * HMX_FP16_TILE_N_ELMS;
    for (size_t r = 0; r < n_row_tiles; ++r) {
-        const __fp16 * d_diag     = d_tiles + r * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        const __fp16 * d_diag     = d_tiles + r * HMX_FP16_TILE_N_ELMS;
        const __fp16 * p_tile_in  = p_tiles + (r * n_tiles_per_bc) * HMX_FP16_TILE_N_ELMS;
        const __fp16 * o_rc       = o_prev + r * HMX_FP16_TILE_N_ELMS;
        const __fp16 * v_tile_in  = v_tiles;
@@ -1654,7 +1657,7 @@ static void hmx_fa_o_norm_worker(void * data) {
    asm volatile(HMX_SET_BIAS("%0") :: "r"((unsigned int)job->hmx_scales));
    const size_t o_stride = n_row_tiles_g_br * HMX_FP16_TILE_N_ELMS;
    for (size_t r = 0; r < n_row_tiles; ++r) {
-        const __fp16 * d_diag = d_tiles + r * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        const __fp16 * d_diag = d_tiles + r * HMX_FP16_TILE_N_ELMS;
        const __fp16 * o_rc = o_prev + r * HMX_FP16_TILE_N_ELMS;
        __fp16 *       o_out = o_curr + r * DV_tiles * HMX_FP16_TILE_N_ELMS;

@@ -1882,7 +1885,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    factx.vtcm_s_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_s_tiles[1], pipeline);
    factx.vtcm_p_tiles[0]     = VTCM_LAYOUT_PTR(__fp16, base, L.off_p_tiles[0]);
    factx.vtcm_p_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_p_tiles[1], pipeline);
-    factx.vtcm_d_tiles        = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_tiles);
+    factx.vtcm_d_tiles[0]     = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_tiles[0]);
+    factx.vtcm_d_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_d_tiles[1], pipeline);
    factx.vtcm_d_inv_l        = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_inv_l);
    factx.vtcm_m_vec          = VTCM_LAYOUT_PTR(HVX_Vector, base, L.off_m_vec);
    factx.vtcm_l_vec          = VTCM_LAYOUT_PTR(HVX_Vector, base, L.off_l_vec);
@@ -2039,7 +2043,30 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            }
                        }

-                        // ---- 3. Pop and run K-prep for next block & push next QK-dot ----
+                        // ---- 3. Start HMX O update for block kv_blk - 1 (reads P[1 - buf_idx], V[1 - buf_idx], D) ----
+                        // O update relys on the previous block's P and V tiles.
+                        // O update MUST be pushed before the next block's QK-dot: hmx_queue_pop() retires the
+                        // oldest descriptor, so push order alone decides which pop waits for which job.
+                        // If OU went in after QK(i+1), the pop below would retire QK(i+1) and leave
+                        // OU(i-1) in flight into the next iteration, where V-prep overwrites V[prev_buf].
+                        if (kv_blk > 0) {
+                            const size_t prev_buf        = 1 - buf_idx;
+                            ou_job[prev_buf].o_curr      = o_tile_curr;
+                            ou_job[prev_buf].o_prev      = o_tile_prev;
+                            ou_job[prev_buf].p_tiles     = factx.vtcm_p_tiles[prev_buf];
+                            ou_job[prev_buf].v_tiles     = factx.vtcm_v_tiles[prev_buf];
+                            ou_job[prev_buf].d_tiles     = factx.vtcm_d_tiles[prev_buf];
+                            ou_job[prev_buf].hmx_scales  = factx.vtcm_hmx_scales_id;
+                            ou_job[prev_buf].n_row_tiles = n_row_tiles;
+                            ou_job[prev_buf].n_col_tiles =
+                                hmx_ceil_div(hex_smin(Bc, nek1 - (kv_blk - 1) * Bc), HMX_FP16_TILE_N_COLS);
+                            ou_job[prev_buf].n_row_tiles_g_br = n_row_tiles_g_br;
+                            ou_job[prev_buf].n_tiles_per_bc   = n_tiles_per_bc;
+                            ou_job[prev_buf].DV               = DV;
+                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
+                        }
+
+                        // ---- 4. Pop and run K-prep for next block & push next QK-dot ----
                        if (kv_blk + 1 < factx.n_kv_blocks) {
                            const uint32_t next_start = (kv_blk + 1) * Bc;
                            const uint32_t next_rows  = hex_smin(Bc, nek1 - next_start);
@@ -2059,10 +2086,10 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_qk_dot_worker, &qk_job[next_buf]));
                        }

-                        // ---- 4. Wait for current block's QK-dot to finish ----
+                        // ---- 5. Wait for current block's QK-dot to finish ----
                        hmx_queue_pop(hmx_q);

-                        // ---- 5. Phase 2: softmax + build_D ----
+                        // ---- 6. Phase 2: softmax + build_D ----
                        fa_softmax_args_t sargs;
                        memset(&sargs, 0, sizeof(sargs));
                        sargs.factx                = &factx;
@@ -2085,23 +2112,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                        sargs.mask_vtcm_row_stride = factx.mask_buf_row_stride;
                        sargs.slopes               = factx.vtcm_slopes;

-                        // Start HMX O update for block kv_blk - 1 (reads P[1 - buf_idx], V[1 - buf_idx])
-                        if (kv_blk > 0) {
-                            const size_t prev_buf = 1 - buf_idx;
-                            ou_job[prev_buf].o_curr           = o_tile_curr;
-                            ou_job[prev_buf].o_prev           = o_tile_prev;
-                            ou_job[prev_buf].p_tiles          = factx.vtcm_p_tiles[prev_buf];
-                            ou_job[prev_buf].v_tiles          = factx.vtcm_v_tiles[prev_buf];
-                            ou_job[prev_buf].d_tiles          = factx.vtcm_d_tiles;
-                            ou_job[prev_buf].hmx_scales       = factx.vtcm_hmx_scales_id;
-                            ou_job[prev_buf].n_row_tiles      = n_row_tiles;
-                            ou_job[prev_buf].n_col_tiles      = hmx_ceil_div(hex_smin(Bc, nek1 - (kv_blk - 1) * Bc), HMX_FP16_TILE_N_COLS);
-                            ou_job[prev_buf].n_row_tiles_g_br = n_row_tiles_g_br;
-                            ou_job[prev_buf].n_tiles_per_bc   = n_tiles_per_bc;
-                            ou_job[prev_buf].DV               = DV;
-                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
-                        }
-
                        // Run Softmax on HVX (blocking call)
                        fa_phase_softmax_and_build_d(&factx, &sargs, n_row_tiles, n_row_tiles_g_br);

@@ -2128,7 +2138,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                        ou_job[0].o_prev           = o_tile_prev;
                        ou_job[0].p_tiles          = factx.vtcm_p_tiles[1 - buf_idx];
                        ou_job[0].v_tiles          = factx.vtcm_v_tiles[1 - buf_idx];
-                        ou_job[0].d_tiles          = factx.vtcm_d_tiles;
+                        ou_job[0].d_tiles          = factx.vtcm_d_tiles[1 - buf_idx];
                        ou_job[0].hmx_scales       = factx.vtcm_hmx_scales_id;
                        ou_job[0].n_row_tiles      = n_row_tiles;
                        ou_job[0].n_col_tiles      = last_cols;
@@ -2232,7 +2242,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            ou_job.o_prev           = o_tile_prev;
                            ou_job.p_tiles          = factx.vtcm_p_tiles[0];
                            ou_job.v_tiles          = factx.vtcm_v_tiles[0];
-                            ou_job.d_tiles          = factx.vtcm_d_tiles;
+                            ou_job.d_tiles          = factx.vtcm_d_tiles[0];
                            ou_job.hmx_scales       = factx.vtcm_hmx_scales_id;
                            ou_job.n_row_tiles      = n_row_tiles;
                            ou_job.n_col_tiles      = n_col_tiles;
@@ -109,7 +109,7 @@ struct hmx_fa_vtcm_layout {
    size_t off_v_tiles[2];
    size_t off_s_tiles[2];
    size_t off_p_tiles[2];
-    size_t off_d_tiles;
+    size_t off_d_tiles[2];
    size_t off_d_inv_l;
    size_t off_m_vec;
    size_t off_l_vec;
@@ -125,7 +125,7 @@ struct hmx_fa_vtcm_layout {
    size_t q_tile_bytes;
    size_t o_tile_bytes;
    size_t s_tile_bytes;       // S and P tiles (same size)
-    size_t d_tile_bytes;
+    size_t d_tile_bytes;       // d_tiles[0..1] + d_inv_l, allocated back to back
    size_t m_line_bytes;       // one mask row
    size_t m_buf_slot_bytes;   // one dma_cache slot = align_up(Br * m_line_bytes, 4096)
    size_t col_vec_bytes;
@@ -149,7 +149,12 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
    const size_t k_tile_size  = hex_align_up(Bc   * DK   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
    const size_t v_tile_size  = hex_align_up(Bc   * DV   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
    const size_t s_tile_size  = hex_align_up(g_br * Bc   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
-    const size_t d_tile_size  = hex_align_up(g_br * g_br * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
+
+    // The rescale matrices are diagonal: the HMX kernels only ever load the g_br/32
+    // tiles that sit on the diagonal, so store just those, packed back to back with
+    // a stride of one tile.  The old [g_br, g_br] square layout allocated g_br/32
+    // times more than it used, which is also why a second D buffer was unaffordable.
+    const size_t d_tile_size = (g_br / HMX_FP16_TILE_N_ROWS) * HTP_FA_HMX_TILE_SIZE;

    const size_t q_dma_size   = hex_align_up(g_br * DK * (is_q_fp32 ? sizeof(float) : sizeof(__fp16)), 128);
    const size_t k_dma_size   = hex_align_up(Bc * hex_round_up(DK * sizeof(__fp16), 128), 128);
@@ -167,7 +172,8 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
    VTCM_LAYOUT_ALLOC(off, off_q_tiles,       q_tile_size);
    VTCM_LAYOUT_ALLOC(off, off_o_tiles[0],    o_tile_size);
    VTCM_LAYOUT_ALLOC(off, off_o_tiles[1],    o_tile_size);
-    VTCM_LAYOUT_ALLOC(off, off_d_tiles,       d_tile_size);
+    VTCM_LAYOUT_ALLOC(off, off_d_tiles[0],    d_tile_size);
+    VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_d_tiles[1], d_tile_size, pipeline);
    VTCM_LAYOUT_ALLOC(off, off_d_inv_l,       d_tile_size);

    // Group B & C share start offset (Group B tiles must be 2KB aligned)
@@ -213,7 +219,10 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
    L->o_tile_bytes        = o_tile_size;
    L->col_vec_bytes       = col_vec_size;
    L->s_tile_bytes        = s_tile_size;
-    L->d_tile_bytes        = d_tile_size;
+    // Measured from the actual offsets rather than assumed to be N * d_tile_size, so
+    // that inserting a region between them (or adding padding to VTCM_LAYOUT_ALLOC)
+    // cannot silently leave the tail of the run unzeroed.
+    L->d_tile_bytes        = (L->off_d_inv_l + d_tile_size) - L->off_d_tiles[0];
    L->m_line_bytes        = m_line_size;
    L->m_buf_slot_bytes    = m_buf_slot;
    L->row_buf_stride      = row_vec_size / 128;
@@ -329,6 +329,7 @@ typedef struct {
    uint64_t nb3;
    int32_t  n_past;
    int32_t  n_dims;
+    int32_t  n_offs;
    int32_t  n_ctx_orig;
    float    freq_base;
    float    freq_scale;
@@ -341,6 +342,7 @@ typedef struct {
    int32_t  sect_2;
    int32_t  sect_3;
    bool     src2;
+    bool     inplace;
 } ggml_metal_kargs_rope;

 typedef struct {
@@ -3884,6 +3884,11 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
    const int sect_2 = ((const int32_t *) op->op_params)[13];
    const int sect_3 = ((const int32_t *) op->op_params)[14];

+    const int n_offs = ((const int32_t *) op->op_params)[15];
+
+    // when dst aliases src0, the channels outside the rotated window already hold the correct data
+    const bool inplace = op->data == op->src[0]->data;
+
    ggml_metal_kargs_rope args = {
        /*.ne00        =*/ ne00,
        /*.ne01        =*/ ne01,
@@ -3903,6 +3908,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
        /*.nb3         =*/ nb3,
        /*.n_past      =*/ n_past,
        /*.n_dims      =*/ n_dims,
+        /*.n_offs      =*/ n_offs,
        /*.n_ctx_orig  =*/ n_ctx_orig,
        /*.freq_base   =*/ freq_base,
        /*.freq_scale  =*/ freq_scale,
@@ -3915,6 +3921,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
        /* sect_2      =*/ sect_2,
        /* sect_3      =*/ sect_3,
        /* src2        =*/ op->src[2] != nullptr,
+        /* inplace     =*/ inplace,
    };

    auto pipeline = ggml_metal_library_get_pipeline_rope(lib, op);
@@ -656,13 +656,13 @@ void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & r

 template <typename type4x4>
 void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
-    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
    const float d = xb->d;

    float4x4 reg_f;

-    for (int i = 0; i < 16; i++) {
-        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    for (int i = 0; i < 4; ++i) {
+        reg_f[i] = float4(qs[4*il + i]) * d;
    }

    reg = (type4x4) reg_f;
@@ -670,12 +670,10 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg

 template <typename type4>
 void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
-    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
    const float d = xb->d;

-    for (int i = 0; i < 4; i++) {
-        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
-    }
+    reg = (type4) (float4(qs[il]) * d);
 }

 template <typename type4x4>
@@ -4688,14 +4686,15 @@ kernel void kernel_rope_norm(
    float sin_theta;

    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
-        if (i0 < args.n_dims) {
-            const int ic = i0/2;
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;

-            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);

            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;

-            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);

            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
@@ -4706,6 +4705,10 @@ kernel void kernel_rope_norm(
            dst_data[0] = x0*cos_theta - x1*sin_theta;
            dst_data[1] = x0*sin_theta + x1*cos_theta;
        } else {
+            if (args.inplace) {
+                continue;
+            }
+
            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);

@@ -4741,17 +4744,18 @@ kernel void kernel_rope_neox(
    float sin_theta;

    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
-        if (i0 < args.n_dims) {
-            const int ic = i0/2;
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;

-            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);

            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;

-            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);

-            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
-            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);

            const float x0 = src[0];
            const float x1 = src[args.n_dims/2];
@@ -4759,6 +4763,10 @@ kernel void kernel_rope_neox(
            dst_data[0]             = x0*cos_theta - x1*sin_theta;
            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
        } else {
+            if (args.inplace) {
+                continue;
+            }
+
            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);

@@ -4793,8 +4801,9 @@ kernel void kernel_rope_multi(
    float sin_theta;

    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
-        if (i0 < args.n_dims) {
-            const int ic = i0/2;
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;

            // mrope theta calculations
            // note: the rest is the same as kernel_rope_neox
@@ -4827,14 +4836,14 @@ kernel void kernel_rope_multi(
            }
            // end of mrope

-            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);

            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;

-            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);

-            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
-            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);

            const float x0 = src[0];
            const float x1 = src[args.n_dims/2];
@@ -4842,6 +4851,10 @@ kernel void kernel_rope_multi(
            dst_data[0]             = x0*cos_theta - x1*sin_theta;
            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
        } else {
+            if (args.inplace) {
+                continue;
+            }
+
            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);

@@ -202,6 +202,7 @@ set(GGML_OPENCL_KERNELS
    sqr
    sqrt
    ssm_conv
+    ssm_scan
    gated_delta_net
    sub
    sum_rows
@@ -866,6 +866,9 @@ struct ggml_backend_opencl_context {
    // [size_idx][kda][tgpp] where size_idx: 0=S_V=16, 1=32, 2=64, 3=128; kda: 0 or 1.
    // tgpp 0 = TG variant (COLS_PER_LANE_GROUP=1), tgpp 1 = prefill variant (COLS_PER_LANE_GROUP=4).
    cl_kernel kernel_gated_delta_net_f32[4][2][2] = {};
+    cl_kernel kernel_ssm_scan_f32_mamba2_d128 = nullptr;
+    cl_kernel kernel_ssm_scan_f32_mamba2_d256 = nullptr;
+
    cl_kernel kernel_timestep_embedding;
    cl_kernel kernel_gemv_moe_q4_0_f32_ns, kernel_gemm_moe_q4_0_f32_ns, kernel_gemm_moe_q4_0_f32_ns_bin;
    cl_kernel kernel_gemm_moe_q8_0_f32_ns;
@@ -892,6 +895,7 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_gemm_moe_q4_0_q8_1_dp4a = nullptr;    // dp4a (int8) q4_0 MoE prefill GEMM
    cl_kernel kernel_moe_reorder_b;
    cl_kernel kernel_moe_histogram, kernel_moe_scan, kernel_moe_fill, kernel_moe_scatter;
+    cl_kernel kernel_moe_scatter_stable = nullptr;   // deterministic slot assignment
    cl_kernel kernel_moe_combine_f32 = nullptr;   // fused router-weight mul + cross-expert sum
    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
    cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
@@ -3154,6 +3158,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        GGML_LOG_CONT(".");
    }

+    // ssm_scan (Mamba-2 fused per-token recurrent step; d_state in {128, 256})
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "ssm_scan.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("ssm_scan.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_ssm_scan_f32_mamba2_d128 = clCreateKernel(prog, "kernel_ssm_scan_f32_mamba2_d128", &err), err));
+        CL_CHECK((backend_ctx->kernel_ssm_scan_f32_mamba2_d256 = clCreateKernel(prog, "kernel_ssm_scan_f32_mamba2_d256", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
    // gated_delta_net: one kernel per (S_V, KDA, tgpp) triple.
    {
    #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -4442,6 +4464,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        CL_CHECK((backend_ctx->kernel_moe_scan = clCreateKernel(prog, "kernel_moe_scan", &err), err));
        CL_CHECK((backend_ctx->kernel_moe_fill = clCreateKernel(prog, "kernel_moe_fill", &err), err));
        CL_CHECK((backend_ctx->kernel_moe_scatter = clCreateKernel(prog, "kernel_moe_scatter", &err), err));
+        CL_CHECK((backend_ctx->kernel_moe_scatter_stable = clCreateKernel(prog, "kernel_moe_scatter_stable", &err), err));
        CL_CHECK(clReleaseProgram(prog));
        GGML_LOG_CONT(".");
    }
@@ -7301,6 +7324,23 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                   (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
        case GGML_OP_SSM_CONV:
            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_SSM_SCAN: {
+            // Mamba-2 fused per-token scan. Requires src3->ne[0] == 1 (scalar
+            // A per head); d_state in {128, 256}; all sources f32. Falls back
+            // to CPU otherwise (incl. Mamba-1 element-wise A).
+            for (int i = 0; i < 6; ++i) {
+                if (op->src[i]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+            }
+            if (op->type != GGML_TYPE_F32) {
+                return false;
+            }
+            const int K = ggml_get_op_params_i32(op, 0);
+            const int d_state = (int) op->src[0]->ne[0];
+            const bool is_mamba2 = (op->src[3]->ne[0] == 1);
+            return is_mamba2 && (d_state == 128 || d_state == 256) && (K == 1);
+        }
        case GGML_OP_GATED_DELTA_NET:
            {
                // Match the Vulkan backend: only F32 -> F32, S_v in {16, 32, 64, 128}.
@@ -7376,6 +7416,9 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
        case GGML_OP_DIAG_MASK_INF:
            return op->ne[3] == 1;
        case GGML_OP_ROPE: {
+            if (((const int32_t *) op->op_params)[15] != 0) {
+                return false; // FIXME: support ggml_rope_set_offset
+            }
            const int mode = ((const int32_t *) op->op_params)[2];
            const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
            const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
@@ -12257,6 +12300,103 @@ static void ggml_cl_mean(ggml_backend_t backend, const ggml_tensor * src0, const
    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }

+static void ggml_cl_ssm_scan(ggml_backend_t backend, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // s
+    const ggml_tensor * src1 = dst->src[1]; // x
+    const ggml_tensor * src2 = dst->src[2]; // dt
+    const ggml_tensor * src3 = dst->src[3]; // A
+    const ggml_tensor * src4 = dst->src[4]; // B
+    const ggml_tensor * src5 = dst->src[5]; // C
+    const ggml_tensor * src6 = dst->src[6]; // ids
+
+    GGML_ASSERT(src0 && src1 && src2 && src3 && src4 && src5 && src6 && dst);
+
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *) backend->context;
+
+    ggml_tensor_extra_cl * e0 = (ggml_tensor_extra_cl *) src0->extra;
+    ggml_tensor_extra_cl * e1 = (ggml_tensor_extra_cl *) src1->extra;
+    ggml_tensor_extra_cl * e2 = (ggml_tensor_extra_cl *) src2->extra;
+    ggml_tensor_extra_cl * e3 = (ggml_tensor_extra_cl *) src3->extra;
+    ggml_tensor_extra_cl * e4 = (ggml_tensor_extra_cl *) src4->extra;
+    ggml_tensor_extra_cl * e5 = (ggml_tensor_extra_cl *) src5->extra;
+    ggml_tensor_extra_cl * e6 = (ggml_tensor_extra_cl *) src6->extra;
+    ggml_tensor_extra_cl * ed = (ggml_tensor_extra_cl *) dst->extra;
+
+    cl_ulong o0 = e0->offset + src0->view_offs;
+    cl_ulong o1 = e1->offset + src1->view_offs;
+    cl_ulong o2 = e2->offset + src2->view_offs;
+    cl_ulong o3 = e3->offset + src3->view_offs;
+    cl_ulong o4 = e4->offset + src4->view_offs;
+    cl_ulong o5 = e5->offset + src5->view_offs;
+    cl_ulong o6 = e6->offset + src6->view_offs;
+    cl_ulong od = ed->offset + dst->view_offs;
+
+    const int d_state  = (int) src0->ne[0];
+    const int head_dim = (int) src0->ne[1];
+    const int n_head   = (int) src1->ne[1];
+    const int n_group  = (int) src4->ne[1];
+    const int n_tokens = (int) src1->ne[2];
+    const int n_seqs   = (int) src1->ne[3];
+
+    // Mirror CPU ref: s_off = ggml_nelements(src1) * sizeof(float)
+    const cl_ulong s_off_bytes = (cl_ulong) ggml_nelements(src1) * sizeof(float);
+
+    cl_kernel kernel = (d_state == 128)
+        ? backend_ctx->kernel_ssm_scan_f32_mamba2_d128
+        : backend_ctx->kernel_ssm_scan_f32_mamba2_d256;
+    GGML_ASSERT(kernel != nullptr);
+
+    cl_ulong s0_nb2 = src0->nb[2];
+    cl_ulong s0_nb3 = src0->nb[3];
+    cl_ulong x_nb2  = src1->nb[2];
+    cl_ulong x_nb3  = src1->nb[3];
+    cl_ulong dt_nb1 = src2->nb[1];
+    cl_ulong dt_nb2 = src2->nb[2];
+    cl_ulong A_nb1  = src3->nb[1];
+    cl_ulong B_nb2  = src4->nb[2];
+    cl_ulong B_nb3  = src4->nb[3];
+    cl_ulong C_nb2  = src5->nb[2];
+    cl_ulong C_nb3  = src5->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &e0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &o0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &e1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &o1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &e2->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &o2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &e3->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &o3));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_mem),   &e4->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &o4));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_mem),   &e5->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &o5));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_mem),   &e6->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &o6));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &od));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &s0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &s0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &x_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &x_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &dt_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &dt_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &A_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &B_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &B_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong), &C_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &C_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &s_off_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &head_dim));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(int),      &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_group));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &n_tokens));
+
+    size_t global_work_size[] = { (size_t)n_head * head_dim * 64, (size_t)n_seqs, 1 };
+    size_t local_work_size[]  = { 64, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_ssm_conv(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0);
    GGML_ASSERT(src0->extra);
@@ -20725,18 +20865,42 @@ static void moe_router_reoerder(ggml_backend_t backend, const ggml_tensor * src,
    size_t fill_local_size[] = {64, 1, 1};
    backend_ctx->enqueue_ndrange_kernel(kernel, 3, fill_global_size, fill_local_size, src);

-    // Scatter
-    kernel = backend_ctx->kernel_moe_scatter;
-    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
-    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
-    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
-    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
-    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &slot_counter_buf));
-    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne21));
-    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne20));
-    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne02));
+    // Scatter. The deterministic variant is the default: kernel_moe_scatter derives
+    // each token's slot from an atomic counter, so the packing inside an expert - and
+    // with it the output of the ragged prefill GEMM - changes from run to run. Set
+    // GGML_OPENCL_MOE_STABLE_SCATTER=0 to restore the atomic version.
+    static const bool stable_scatter = []{
+        const char * e = getenv("GGML_OPENCL_MOE_STABLE_SCATTER");
+        return !e || e[0] == '\0' || e[0] != '0';
+    }();

-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);
+    if (stable_scatter) {
+        kernel = backend_ctx->kernel_moe_scatter_stable;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne21));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne20));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02));
+
+        // one workgroup (one wave) per expert; each ranks its own tokens
+        size_t scatter_global_size[] = {64, (size_t)ne02};
+        size_t scatter_local_size[]  = {64, 1};
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, scatter_global_size, scatter_local_size, src);
+    } else {
+        kernel = backend_ctx->kernel_moe_scatter;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &slot_counter_buf));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne21));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne20));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne02));
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);
+    }

    // [MOE_TILES] env-gated padding probe: read back total_tiles (= Sum_e
    // ceil(k_e/n_tile_size)) and compare to the ideal tile count for the real
@@ -24743,6 +24907,14 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
            }
            func = ggml_cl_ssm_conv;
            break;
+        case GGML_OP_SSM_SCAN:
+            if (!any_on_device) {
+                return false;
+            }
+            // SSM_SCAN has 7 source tensors, so it cannot use the standard
+            // (src0, src1, dst) func signature. Dispatch directly and return.
+            ggml_cl_ssm_scan(backend, tensor);
+            return true;
        case GGML_OP_GATED_DELTA_NET:
            if (!any_on_device) {
                return false;
@@ -118,6 +118,17 @@ __kernel void flash_attn_f16(
    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+#if WG_SIZE > FA_SG
+        // WAR on l_k/l_v: a thread that finishes the compute below early — either
+        // it skipped it (my_query_row >= n_q, the continue) or its subgroup simply
+        // ran ahead — wraps around and reloads the tiles while another subgroup is
+        // still reading them. Any WG that is exactly one lockstep subgroup
+        // (WG_SIZE == FA_SG) cannot diverge and hides this; a WG spanning multiple
+        // subgroups (Intel sg=32, or BLOCK_M > 64 on Adreno) corrupts the result.
+        // All threads reach this each iteration (no-op on the first), so it does
+        // not diverge with the continue. Compiled out when WG == one subgroup.
+        barrier(CLK_LOCAL_MEM_FENCE);
+#endif
        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
            const int row = i / DK_VEC;
            const int col = i % DK_VEC;
@@ -119,13 +119,15 @@ __kernel void flash_attn_f32(
    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
-#if FA_SG < 64
-        // WAR on l_k/l_v: threads with my_query_row >= n_q skip the compute below
-        // (continue) and would race ahead to reload the tiles while active threads
-        // still read them. A single 64-wide Adreno subgroup (WG == sg) runs lockstep
-        // and hides this; a WG that spans multiple narrower subgroups (Intel sg=32)
-        // corrupts the result. All threads reach this each iteration (no-op on the
-        // first), so it does not diverge with the continue. Compiled out at sg=64.
+#if WG_SIZE > FA_SG
+        // WAR on l_k/l_v: a thread that finishes the compute below early — either
+        // it skipped it (my_query_row >= n_q, the continue) or its subgroup simply
+        // ran ahead — wraps around and reloads the tiles while another subgroup is
+        // still reading them. Any WG that is exactly one lockstep subgroup
+        // (WG_SIZE == FA_SG) cannot diverge and hides this; a WG spanning multiple
+        // subgroups (Intel sg=32, or BLOCK_M > 64 on Adreno) corrupts the result.
+        // All threads reach this each iteration (no-op on the first), so it does
+        // not diverge with the continue. Compiled out when WG == one subgroup.
        barrier(CLK_LOCAL_MEM_FENCE);
 #endif
        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
@@ -68,6 +68,79 @@ __kernel void kernel_moe_scatter(
    emap[tile_idx] = val;
 }

+// Deterministic replacement for kernel_moe_scatter.
+//
+// kernel_moe_scatter takes each token's slot from atomic_inc(slot_counter[expert]),
+// so the token -> slot packing inside an expert depends on which work-item wins the
+// atomic and changes from run to run. The ragged prefill GEMM path is sensitive to
+// that packing (the non-ragged path is not, since its padded slots alias slot 0 and
+// are overwritten last), which makes MoE prompt processing non-reproducible: the same
+// binary on the same prompt returns one of several outputs.
+//
+// Here the slot is the token's rank in flat (n, k) order among the tokens routed to
+// the same expert - a fixed function of the routing input. One workgroup per expert
+// walks the flat routing list in blocks of 64 and ranks its own tokens with a
+// workgroup scan, carrying a running count between blocks. Cost is one pass over the
+// routing list per expert; the list is a few KiB and stays in cache.
+__kernel void kernel_moe_scatter_stable(
+    __global const int * input,
+    __global int * post_router,
+    __global ushort * emap,
+    __global const int * tile_offset,
+    int N,
+    int topK,
+    uint n_experts
+) {
+    const int e   = get_group_id(1);
+    const int lid = get_local_id(0);
+    const int M   = N * topK;
+
+    __local int scan[64];
+    __local int running;
+
+    if (lid == 0) {
+        running = 0;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    for (int base = 0; base < M; base += 64) {
+        const int j = base + lid;
+
+        int pred = 0;
+        if (j < M) {
+            const int n = j / topK;
+            const int k = j - n * topK;
+            pred = (input[n * (int)n_experts + k] == e) ? 1 : 0;
+        }
+
+        scan[lid] = pred;
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        // Hillis-Steele inclusive scan over the 64 lanes
+        for (int off = 1; off < 64; off <<= 1) {
+            int add = (lid >= off) ? scan[lid - off] : 0;
+            barrier(CLK_LOCAL_MEM_FENCE);
+            scan[lid] += add;
+            barrier(CLK_LOCAL_MEM_FENCE);
+        }
+
+        if (pred) {
+            const int local_slot = running + (scan[lid] - 1);   // exclusive rank
+            const int tile_idx   = tile_offset[e] + (local_slot >> 5);
+            const int lane       = local_slot & 31;
+
+            post_router[tile_idx * 32 + lane] = j;
+            emap[tile_idx] = (ushort)e;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+        if (lid == 63) {
+            running += scan[63];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+}
+
 __kernel void kernel_moe_fill(
    __global int * post_router,
    __global int * total_tiles,
@@ -0,0 +1,216 @@
+// Mamba2 fused SSM scan kernel. One workgroup per (head, dim, seq); WG size =
+// 64 threads. Each thread owns c_factor = d_state/64 state elements in
+// private registers; the state stays resident across the n_tokens t-loop
+//
+// References:
+//   ggml/src/ggml-cuda/ssm-scan.cu:117 ssm_scan_f32_group
+//   ggml/src/ggml-cpu/ops.cpp:9368 ggml_compute_forward_ssm_scan_f32
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_khr_subgroups
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#else
+#define REQD_SUBGROUP_SIZE_64
+#endif
+
+inline float softplus_f32(float x) {
+    return (x <= 20.0f) ? log(1.0f + exp(x)) : x;
+}
+
+// d_state = 128 (most Mamba-2 models, e.g. mamba2-2.7B, Codestral-Mamba).
+// WG = 64 threads, each holds 2 state elements (tid and tid+64).
+REQD_SUBGROUP_SIZE_64
+kernel void kernel_ssm_scan_f32_mamba2_d128(
+    global const char * src0_base, ulong src0_off,
+    global const char * src1_base, ulong src1_off,
+    global const char * src2_base, ulong src2_off,
+    global const char * src3_base, ulong src3_off,
+    global const char * src4_base, ulong src4_off,
+    global const char * src5_base, ulong src5_off,
+    global const char * src6_base, ulong src6_off,
+    global       char * dst_base,  ulong dst_off,
+    ulong s0_nb2, ulong s0_nb3,
+    ulong x_nb2,  ulong x_nb3,
+    ulong dt_nb1, ulong dt_nb2,
+    ulong A_nb1,
+    ulong B_nb2,  ulong B_nb3,
+    ulong C_nb2,  ulong C_nb3,
+    ulong s_off_bytes,
+    int   head_dim, int n_head, int n_group, int n_tokens
+) {
+    const int d_state = 128;
+
+    const int tid     = (int) get_local_id(0);
+    const int wg_x    = (int) get_group_id(0);
+    const int seq_id  = (int) get_group_id(1);
+
+    const int head_id = wg_x / head_dim;
+    const int dim_id  = wg_x - head_id * head_dim;
+    const int g       = head_id / (n_head / n_group);
+
+    src0_base += src0_off;
+    src1_base += src1_off;
+    src2_base += src2_off;
+    src3_base += src3_off;
+    src4_base += src4_off;
+    src5_base += src5_off;
+    src6_base += src6_off;
+    dst_base  += dst_off;
+
+    const int seq_slot = ((global const int *) src6_base)[seq_id];
+
+    const ulong state_base_off = (ulong)seq_slot * s0_nb3 + (ulong)head_id * s0_nb2
+                                + (ulong)dim_id * d_state * sizeof(float);
+    global const float * s0_warp = (global const float *)(src0_base + state_base_off);
+    const ulong state_out_off = (ulong)seq_id * s0_nb3 + (ulong)head_id * s0_nb2
+                              + (ulong)dim_id * d_state * sizeof(float);
+    global float * s_warp = (global float *)(dst_base + s_off_bytes + state_out_off);
+
+    global const char * x_seq  = src1_base + (ulong)seq_id * x_nb3;
+    global const char * dt_seq = src2_base + (ulong)seq_id * dt_nb2;
+    global const char * B_seq  = src4_base + (ulong)seq_id * B_nb3 + (ulong)g * d_state * sizeof(float);
+    global const char * C_seq  = src5_base + (ulong)seq_id * C_nb3 + (ulong)g * d_state * sizeof(float);
+
+    const ulong y_dim_total = (ulong)n_head * head_dim;
+    global float * y_seq = (global float *)dst_base
+                           + (ulong)seq_id * (ulong)n_tokens * y_dim_total;
+
+    const float A_val = ((global const float *)src3_base)[(ulong)head_id * A_nb1 / sizeof(float)];
+
+    // c_factor = 2: each thread owns 2 state elements (tid and tid+64).
+    float state0 = s0_warp[tid];
+    float state1 = s0_warp[tid + 64];
+
+    for (int t = 0; t < n_tokens; ++t) {
+        const float dt_h        = ((global const float *)(dt_seq + (ulong)t * dt_nb1))[head_id];
+        const float dt_softplus = softplus_f32(dt_h);
+        const float dA          = exp(dt_softplus * A_val);
+        const float x_val       = ((global const float *)(x_seq + (ulong)t * x_nb2))[(ulong)head_id * head_dim + dim_id];
+        const float x_dt        = x_val * dt_softplus;
+
+        const float B0 = ((global const float *)(B_seq + (ulong)t * B_nb2))[tid];
+        const float B1 = ((global const float *)(B_seq + (ulong)t * B_nb2))[tid + 64];
+        const float C0 = ((global const float *)(C_seq + (ulong)t * C_nb2))[tid];
+        const float C1 = ((global const float *)(C_seq + (ulong)t * C_nb2))[tid + 64];
+
+        state0 = state0 * dA + B0 * x_dt;
+        state1 = state1 * dA + B1 * x_dt;
+        const float partial = state0 * C0 + state1 * C1;
+
+        const float sum = sub_group_reduce_add(partial);
+        if (tid == 0) {
+            y_seq[(ulong)t * y_dim_total + (ulong)head_id * head_dim + dim_id] = sum;
+        }
+    }
+
+    s_warp[tid]      = state0;
+    s_warp[tid + 64] = state1;
+}
+
+// d_state = 256 (Falcon-H1). WG = 64 threads, each holds 4 state elements.
+REQD_SUBGROUP_SIZE_64
+kernel void kernel_ssm_scan_f32_mamba2_d256(
+    global const char * src0_base, ulong src0_off,
+    global const char * src1_base, ulong src1_off,
+    global const char * src2_base, ulong src2_off,
+    global const char * src3_base, ulong src3_off,
+    global const char * src4_base, ulong src4_off,
+    global const char * src5_base, ulong src5_off,
+    global const char * src6_base, ulong src6_off,
+    global       char * dst_base,  ulong dst_off,
+    ulong s0_nb2, ulong s0_nb3,
+    ulong x_nb2,  ulong x_nb3,
+    ulong dt_nb1, ulong dt_nb2,
+    ulong A_nb1,
+    ulong B_nb2,  ulong B_nb3,
+    ulong C_nb2,  ulong C_nb3,
+    ulong s_off_bytes,
+    int   head_dim, int n_head, int n_group, int n_tokens
+) {
+    const int d_state = 256;
+
+    const int tid     = (int) get_local_id(0);
+    const int wg_x    = (int) get_group_id(0);
+    const int seq_id  = (int) get_group_id(1);
+
+    const int head_id = wg_x / head_dim;
+    const int dim_id  = wg_x - head_id * head_dim;
+    const int g       = head_id / (n_head / n_group);
+
+    src0_base += src0_off;
+    src1_base += src1_off;
+    src2_base += src2_off;
+    src3_base += src3_off;
+    src4_base += src4_off;
+    src5_base += src5_off;
+    src6_base += src6_off;
+    dst_base  += dst_off;
+
+    const int seq_slot = ((global const int *) src6_base)[seq_id];
+
+    const ulong state_base_off = (ulong)seq_slot * s0_nb3 + (ulong)head_id * s0_nb2
+                                + (ulong)dim_id * d_state * sizeof(float);
+    global const float * s0_warp = (global const float *)(src0_base + state_base_off);
+    const ulong state_out_off = (ulong)seq_id * s0_nb3 + (ulong)head_id * s0_nb2
+                              + (ulong)dim_id * d_state * sizeof(float);
+    global float * s_warp = (global float *)(dst_base + s_off_bytes + state_out_off);
+
+    global const char * x_seq  = src1_base + (ulong)seq_id * x_nb3;
+    global const char * dt_seq = src2_base + (ulong)seq_id * dt_nb2;
+    global const char * B_seq  = src4_base + (ulong)seq_id * B_nb3 + (ulong)g * d_state * sizeof(float);
+    global const char * C_seq  = src5_base + (ulong)seq_id * C_nb3 + (ulong)g * d_state * sizeof(float);
+
+    const ulong y_dim_total = (ulong)n_head * head_dim;
+    global float * y_seq = (global float *)dst_base
+                           + (ulong)seq_id * (ulong)n_tokens * y_dim_total;
+
+    const float A_val = ((global const float *)src3_base)[(ulong)head_id * A_nb1 / sizeof(float)];
+
+    // c_factor = 4: each thread owns 4 state elements.
+    float state0 = s0_warp[tid];
+    float state1 = s0_warp[tid + 64];
+    float state2 = s0_warp[tid + 128];
+    float state3 = s0_warp[tid + 192];
+
+    for (int t = 0; t < n_tokens; ++t) {
+        const float dt_h        = ((global const float *)(dt_seq + (ulong)t * dt_nb1))[head_id];
+        const float dt_softplus = softplus_f32(dt_h);
+        const float dA          = exp(dt_softplus * A_val);
+        const float x_val       = ((global const float *)(x_seq + (ulong)t * x_nb2))[(ulong)head_id * head_dim + dim_id];
+        const float x_dt        = x_val * dt_softplus;
+
+        global const float * B_t = (global const float *)(B_seq + (ulong)t * B_nb2);
+        global const float * C_t = (global const float *)(C_seq + (ulong)t * C_nb2);
+
+        const float B0 = B_t[tid];
+        const float B1 = B_t[tid + 64];
+        const float B2 = B_t[tid + 128];
+        const float B3 = B_t[tid + 192];
+        const float C0 = C_t[tid];
+        const float C1 = C_t[tid + 64];
+        const float C2 = C_t[tid + 128];
+        const float C3 = C_t[tid + 192];
+
+        state0 = state0 * dA + B0 * x_dt;
+        state1 = state1 * dA + B1 * x_dt;
+        state2 = state2 * dA + B2 * x_dt;
+        state3 = state3 * dA + B3 * x_dt;
+        const float partial = state0 * C0 + state1 * C1 + state2 * C2 + state3 * C3;
+
+        const float sum = sub_group_reduce_add(partial);
+        if (tid == 0) {
+            y_seq[(ulong)t * y_dim_total + (ulong)head_id * head_dim + dim_id] = sum;
+        }
+    }
+
+    s_warp[tid]       = state0;
+    s_warp[tid + 64]  = state1;
+    s_warp[tid + 128] = state2;
+    s_warp[tid + 192] = state3;
+}
@@ -1227,6 +1227,10 @@ static bool is_op_unsupported_case(const ggml_tensor * op) {
        const int32_t * op_params = op->op_params;
        const int n_dims = op_params[1];
        const int mode = op_params[2];
+        if (op_params[15] != 0) {
+            // FIXME: support ggml_rope_set_offset
+            return true;
+        }
        if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX && mode != GGML_ROPE_TYPE_IMROPE) {
            // GGML_LOG_WARN("OpenVINO backend does not support ROPE with mode %d\n", mode);
            return true;
@@ -47,7 +47,7 @@ struct rpc_tensor {
    uint64_t data;
    char name[GGML_MAX_NAME];

-    char padding[4];
+    int32_t use_count;
 };

 static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
@@ -447,7 +447,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {

    // Avoid sending uninitialized data over the wire
    memset(result.name, 0, sizeof(result.name));
-    memset(result.padding, 0, sizeof(result.padding));
+    result.use_count = 0;

    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
    return result;
@@ -675,7 +675,7 @@ static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
    // this is no-op because we don't have any async operations
 }

-static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
+static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
    if (tensor == nullptr) {
        return;
    }
@@ -684,10 +684,15 @@ static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors,
    }
    visited.insert(tensor);
    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        add_tensor(tensor->src[i], tensors, visited);
+        add_tensor(tensor->src[i], cgraph, tensors, visited);
    }
-    add_tensor(tensor->view_src, tensors, visited);
-    tensors.push_back(serialize_tensor(tensor));
+    add_tensor(tensor->view_src, cgraph, tensors, visited);
+    rpc_tensor result = serialize_tensor(tensor);
+    const size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, tensor);
+    if (hash_pos != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        result.use_count = cgraph->use_counts[hash_pos];
+    }
+    tensors.push_back(result);
 }

 static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
@@ -695,7 +700,7 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
    std::vector<rpc_tensor> tensors;
    std::unordered_set<ggml_tensor*> visited;
    for (uint32_t i = 0; i < n_nodes; i++) {
-        add_tensor(cgraph->nodes[i], tensors, visited);
+        add_tensor(cgraph->nodes[i], cgraph, tensors, visited);
    }
    // serialization format:
    // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
@@ -1451,6 +1456,10 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input) {
            GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id);
            return false;
        }
+        if (graph->nodes[i] != nullptr) {
+            const size_t hash_pos = ggml_hash_insert(&graph->visited_hash_set, graph->nodes[i]);
+            graph->use_counts[hash_pos] = tensor_ptrs.at(id)->use_count;
+        }
    }
    ggml_status status = ggml_backend_graph_compute(backends[device], graph);
    GGML_ASSERT(status == GGML_STATUS_SUCCESS && "Unsuccessful graph computations are not supported with RPC");
@@ -0,0 +1,119 @@
+#include "fwht.hpp"
+
+#include <cmath>
+
+template <int N>
+static void fwht_kernel(const float * __restrict__ src, float * __restrict__ dst, const int64_t n_rows,
+                        const float scale, const sycl::nd_item<2> & item) {
+    const sycl::sub_group sg = item.get_sub_group();
+
+    const int64_t r = item.get_global_id(0);
+    if (r >= n_rows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    constexpr int el_w = N / WARP_SIZE;
+    static_assert(el_w >= 1 && N % WARP_SIZE == 0, "row must be a whole number of sub-group widths");
+
+    float     reg[el_w];
+    const int lane = sg.get_local_linear_id();
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        reg[i] = src[i * WARP_SIZE + lane] * scale;
+    }
+
+    // Butterflies inside the sub-group. The partner of a lane with bit h clear is the
+    // lower index of the pair, so it takes the sum and the upper takes lower - upper.
+#pragma unroll
+    for (int h = 1; h < WARP_SIZE; h *= 2) {
+#pragma unroll
+        for (int j = 0; j < el_w; ++j) {
+            const float val  = reg[j];
+            const float val2 = dpct::permute_sub_group_by_xor(sg, val, h, WARP_SIZE);
+
+            reg[j] = (lane & h) == 0 ? val + val2 : val2 - val;
+        }
+    }
+
+    // Butterflies across registers: h is a multiple of WARP_SIZE, so the partner of
+    // element i*WARP_SIZE + lane lives in reg[i + h/WARP_SIZE] on the same lane.
+#pragma unroll
+    for (int h = WARP_SIZE; h < N; h *= 2) {
+        const int step = h / WARP_SIZE;
+#pragma unroll
+        for (int j = 0; j < el_w; j += 2 * step) {
+#pragma unroll
+            for (int k = 0; k < step; ++k) {
+                const float x = reg[j + k];
+                const float y = reg[j + k + step];
+
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        dst[i * WARP_SIZE + lane] = reg[i];
+    }
+}
+
+template <int N>
+static void launch_fwht(const float * src, float * dst, const int64_t n_rows, const float scale,
+                        dpct::queue_ptr stream) {
+    constexpr int rows_per_block = 4;
+
+    const int64_t num_blocks = (n_rows + rows_per_block - 1) / rows_per_block;
+
+    // dim 1 is the fastest-varying, so a sub-group is exactly one row's WARP_SIZE lanes.
+    const sycl::range<2> global(num_blocks * rows_per_block, WARP_SIZE);
+    const sycl::range<2> local(rows_per_block, WARP_SIZE);
+
+    stream->parallel_for(sycl::nd_range<2>(global, local),
+                         [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             fwht_kernel<N>(src, dst, n_rows, scale, item);
+                         });
+}
+
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
+    if (src->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!ggml_are_same_shape(src, dst)) {
+        return false;
+    }
+    if (!ggml_is_contiguous(src) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    const int     n    = (int) src->ne[0];
+    const int64_t rows = ggml_nrows(src);
+
+    const float *   src_d  = (const float *) src->data;
+    float *         dst_d  = (float *) dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    const float scale = 1.0f / std::sqrt((float) n);
+
+    switch (n) {
+        case 64:
+            launch_fwht<64>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 128:
+            launch_fwht<128>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 256:
+            launch_fwht<256>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 512:
+            launch_fwht<512>(src_d, dst_d, rows, scale, stream);
+            return true;
+        default:
+            return false;
+    }
+}
@@ -0,0 +1,12 @@
+#ifndef GGML_SYCL_FWHT_HPP
+#define GGML_SYCL_FWHT_HPP
+
+#include "common.hpp"
+
+// Fast Walsh-Hadamard transform, the fast path for a MUL_MAT whose src0 ggml has
+// tagged GGML_HINT_SRC0_IS_HADAMARD. src0 is not read at all. Returns false if the
+// shape is not one this can serve, in which case the caller must fall through to the
+// ordinary mat-mul dispatch.
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst);
+
+#endif  // GGML_SYCL_FWHT_HPP
@@ -58,6 +58,7 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
+#include "ggml-sycl/fwht.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml-sycl/norm.hpp"
@@ -108,7 +109,14 @@ int g_ggml_sycl_enable_host_pinned_mem = 1;
 static ggml_sycl_device_info ggml_sycl_init() {
    ggml_sycl_device_info info = {};

-    info.device_count = dpct::dev_mgr::instance().device_count();
+    // Do not hard crash when there exists no SYCL devices.
+    // We want to allow the user to use non-SYCL tools when SYCL is compiled (such as llama-quantize)
+    try {
+        info.device_count = dpct::dev_mgr::instance().device_count();
+    } catch (sycl::exception const &exc) {
+        GGML_LOG_INFO("%s: no SYCL device available: %s\n", __func__, exc.what());
+        info.device_count = 0;
+    }
    if (info.device_count == 0) {
        GGML_LOG_ERROR("%s: failed to initialize: %s\n", GGML_SYCL_NAME, __func__);
        return info;
@@ -4473,6 +4481,18 @@ static bool can_use_mul_mat_vec_q(const ggml_tensor * src0, const ggml_tensor *

 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    // Handle HADAMARAD hint given from further up the pipeline and pass it to the correct
+    // kernel.
+    //
+    // The op check is not redundant: this backend also routes MUL_MAT_ID through here with a
+    // stack copy of dst, which carries MUL_MAT_ID's own op_params. ggml_mul_mat_set_hint()
+    // asserts GGML_OP_MUL_MAT for the same reason.
+    if (dst->op == GGML_OP_MUL_MAT && ggml_get_op_params_i32(dst, 1) == GGML_HINT_SRC0_IS_HADAMARD &&
+        ggml_sycl_op_fwht(ctx, src1, dst)) {
+        return;
+    }
+
    const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
    int64_t min_compute_capability = INT_MAX;

@@ -6222,6 +6242,8 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
        }
        case GGML_OP_ROPE:
        case GGML_OP_ROPE_BACK:
+            // FIXME: support ggml_rope_set_offset
+            return ((const int32_t *) op->op_params)[15] == 0;
        case GGML_OP_IM2COL:
        case GGML_OP_IM2COL_3D:
        case GGML_OP_UPSCALE:
@@ -200,8 +200,11 @@ if (Vulkan_FOUND)
    set (_ggml_vk_header     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp")
    set (_ggml_vk_input_dir  "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders")
    set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv")
+    set (_ggml_vk_generated_shader_files ${_ggml_vk_header})

    file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
+    set_source_files_properties(${_ggml_vk_shader_files} PROPERTIES HEADER_FILE_ONLY TRUE)
+    target_sources(ggml-vulkan PRIVATE ${_ggml_vk_shader_files})

    # Because external projects do not provide source-level tracking,
    # the vulkan-shaders-gen sources need to be explicitly added to
@@ -241,8 +244,11 @@ if (Vulkan_FOUND)
            COMMENT "Generate vulkan shaders for ${file}"
        )
        target_sources(ggml-vulkan PRIVATE ${_ggml_vk_target_cpp})
+        list(APPEND _ggml_vk_generated_shader_files ${_ggml_vk_target_cpp})
    endforeach()

+    source_group("Vulkan shaders" FILES ${_ggml_vk_shader_files})
+    source_group("Generated Vulkan shaders" FILES ${_ggml_vk_generated_shader_files})
 else()
    message(WARNING "Vulkan not found")
 endif()
@@ -913,6 +913,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_quantize_q8_1_x4;

    vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_transpose[GGML_TYPE_COUNT]; // fused dequant+transpose for FA quant-KV
    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
    vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT];
@@ -962,6 +963,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
    vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
    vk_pipeline pipeline_cpy_transpose_16, pipeline_cpy_transpose_32;
+    vk_pipeline pipeline_cpy_transpose_02_16, pipeline_cpy_transpose_02_32;
    // [src0 0=fp32,1=fp16][dst]
    vk_pipeline pipeline_set_rows_i32[2][GGML_TYPE_COUNT];
    vk_pipeline pipeline_set_rows_i64[2][GGML_TYPE_COUNT];
@@ -1644,6 +1646,7 @@ struct vk_op_rope_push_constants {
    uint32_t rope_mode;
    uint32_t nrows;
    uint32_t n_dims;
+    uint32_t n_offs;
    float freq_scale;
    float freq_base;
    float ext_factor;
@@ -3382,10 +3385,10 @@ static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
    // Arbitrary frequency to cleanup/reuse command buffers
    static constexpr uint32_t cleanup_frequency = 10;

-    if (device->compute_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
+    if (device->compute_queue && device->compute_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
        ggml_vk_command_pool_cleanup(device, device->compute_queue->cmd_pool);
    }
-    if (device->transfer_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
+    if (device->transfer_queue && device->transfer_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
        ggml_vk_command_pool_cleanup(device, device->transfer_queue->cmd_pool);
    }
 }
@@ -5389,6 +5392,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_0], "dequant_q5_0", dequant_q5_0_len, dequant_q5_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_1], "dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_transpose[GGML_TYPE_Q8_0], "dequant_q8_0_transpose", dequant_q8_0_transpose_len, dequant_q8_0_transpose_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_k", dequant_q2_k_len, dequant_q2_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_TQ2_0], "dequant_tq2_0", dequant_tq2_0_len, dequant_tq2_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_k", dequant_q3_k_len, dequant_q3_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
@@ -5525,6 +5529,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_32, "cpy_transpose_32", cpy_transpose_32_len, cpy_transpose_32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_16, "cpy_transpose_16", cpy_transpose_16_len, cpy_transpose_16_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_02_32, "cpy_transpose_02_32", cpy_transpose_02_32_len, cpy_transpose_02_32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_02_16, "cpy_transpose_02_16", cpy_transpose_02_16_len, cpy_transpose_02_16_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q1_0], "cpy_f32_q1_0", cpy_f32_q1_0_len, cpy_f32_q1_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q2_0], "cpy_f32_q2_0", cpy_f32_q2_0_len, cpy_f32_q2_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
@@ -8931,6 +8937,18 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
        }
    }

+    // Same, for a 0<->2 swap: src dim2 is the innermost dimension.
+    bool transpose02 = dst && !contig && src->nb[2] == ggml_type_size(to) &&
+                       ggml_is_contiguous(dst) && ggml_are_same_shape(dst, src);
+
+    if (transpose02 && src->type == to) {
+        if (ggml_type_size(to) == 4) {
+            return ctx->device->pipeline_cpy_transpose_02_32;
+        } else if (ggml_type_size(to) == 2) {
+            return ctx->device->pipeline_cpy_transpose_02_16;
+        }
+    }
+
    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
        if (contig) {
            return ctx->device->pipeline_contig_cpy_f32_f32;
@@ -10807,9 +10825,32 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx

    const bool f32acc = !ctx->device->fp16 || dst->op_params[3] == GGML_PREC_F32 || k->type == GGML_TYPE_BF16;

+    // dequant K/V once into an f16 scratch, reordered KV layout so FA can read without a stride
+    auto is_dense_kv_cache = [](const ggml_tensor * t) {
+        return t->nb[0] == ggml_type_size(t->type) &&
+               t->nb[2] == ggml_row_size(t->type, t->ne[0]) &&
+               t->nb[1] == t->nb[2] * t->ne[2] &&
+               t->nb[3] == t->nb[1] * t->ne[1];
+    };
+    const bool k_quant = k->type != GGML_TYPE_F16 && k->type != GGML_TYPE_BF16 && k->type != GGML_TYPE_F32;
+    const bool v_quant = v->type != GGML_TYPE_F16 && v->type != GGML_TYPE_BF16 && v->type != GGML_TYPE_F32;
+    const bool use_dequant_kv = k_quant && v_quant && neq1 >= 64 &&
+                                is_dense_kv_cache(k) && is_dense_kv_cache(v) &&
+                                (uint64_t)ggml_nelements(k) * sizeof(ggml_fp16_t) <= ctx->device->properties.limits.maxStorageBufferRange &&
+                                (uint64_t)ggml_nelements(v) * sizeof(ggml_fp16_t) <= ctx->device->properties.limits.maxStorageBufferRange &&
+                                ctx->device->pipeline_dequant_transpose[k->type] != nullptr &&
+                                ctx->device->pipeline_dequant_transpose[v->type] != nullptr &&
+                                // coopmat2 path does not benefit from the f16 scratch
+                                !ctx->device->coopmat2 &&
+                                // Intel Xe1 regresses, see PR 25494
+                                (ctx->device->vendor_id != VK_VENDOR_ID_INTEL ||
+                                 (ctx->device->coopmat_support && ctx->device->architecture != vk_device_architecture::INTEL_XE1));
+    const ggml_type k_type_eff = use_dequant_kv ? GGML_TYPE_F16 : k->type;
+    const ggml_type v_type_eff = use_dequant_kv ? GGML_TYPE_F16 : v->type;
+
    // For scalar/coopmat1 FA, we can use the "large" size to accommodate qga.
    // For coopmat2 FA, we always use the small size (which is still pretty large for gqa).
-    vk_fa_tuning_params tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, 512, KV, k->type, v->type, f32acc);
+    vk_fa_tuning_params tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, 512, KV, k_type_eff, v_type_eff, f32acc);
    const uint32_t max_gqa = std::min(tuning_params.block_rows, 32u);

    if (N <= 8 && qk_ratio > 1 && qk_ratio <= max_gqa &&
@@ -10822,7 +10863,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
        workgroups_y /= gqa_ratio;
    }

-    tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, N, KV, k->type, v->type, f32acc);
+    tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, N, KV, k_type_eff, v_type_eff, f32acc);

    const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type));
    uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type));
@@ -10836,6 +10877,17 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
        v_stride /= 4;
    }

+    uint32_t nbk2_eff = (uint32_t)nbk2, nbk3_eff = (uint32_t)nbk3;
+    uint32_t nbv2_eff = (uint32_t)nbv2, nbv3_eff = (uint32_t)nbv3;
+    if (use_dequant_kv) {
+        k_stride = HSK;
+        v_stride = HSV;
+        nbk2_eff = (uint32_t)((uint64_t)HSK * KV * sizeof(ggml_fp16_t));
+        nbk3_eff = (uint32_t)((uint64_t)HSK * KV * nek2 * sizeof(ggml_fp16_t));
+        nbv2_eff = (uint32_t)((uint64_t)HSV * KV * sizeof(ggml_fp16_t));
+        nbv3_eff = (uint32_t)((uint64_t)HSV * KV * nev2 * sizeof(ggml_fp16_t));
+    }
+
    const uint32_t alignment = tuning_params.block_cols;
    bool aligned = (KV % alignment) == 0 &&
                   // the "aligned" shader variant will forcibly align strides, for performance
@@ -10862,7 +10914,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
    bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768 && nem0 >= tuning_params.block_cols * 16
                        && (ctx->device->architecture != vk_device_architecture::AMD_GCN || HSK > 256 || HSV > 256);
    vk_fa_pipeline_state fa_pipeline_state = get_fa_pipeline_state(ctx->device, tuning_params, HSK, HSV, aligned, f32acc,
-                                                                   mask != nullptr, use_mask_opt, logit_softcap != 0, k->type, v->type);
+                                                                   mask != nullptr, use_mask_opt, logit_softcap != 0, k_type_eff, v_type_eff);

    vk_pipeline pipeline = nullptr;

@@ -10966,6 +11018,34 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
    vk_subbuffer sinks_buf = sinks ? ggml_vk_tensor_subbuffer(ctx, sinks) : q_buf;
    vk_subbuffer mask_opt_buf = use_mask_opt ? ggml_vk_subbuffer(ctx, ctx->prealloc_y, 0) : q_buf;

+    if (use_dequant_kv) {
+        const uint64_t fp = sizeof(ggml_fp16_t);
+        const uint64_t k_f16_sz = (uint64_t)ggml_nelements(k) * fp;
+        const uint64_t v_f16_sz = (uint64_t)ggml_nelements(v) * fp;
+        if (ctx->prealloc_size_x < k_f16_sz + v_f16_sz) {
+            ctx->prealloc_size_x = k_f16_sz + v_f16_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
+        }
+        vk_pipeline tr_k = ctx->device->pipeline_dequant_transpose[k->type];
+        vk_pipeline tr_v = ctx->device->pipeline_dequant_transpose[v->type];
+        ggml_pipeline_request_descriptor_sets(ctx, tr_k, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, tr_v, 1);
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+        vk_subbuffer k_dst = vk_subbuffer{ ctx->prealloc_x, 0,        k_f16_sz };
+        vk_subbuffer v_dst = vk_subbuffer{ ctx->prealloc_x, k_f16_sz, v_f16_sz };
+        const uint32_t k_nel = (uint32_t)ggml_nelements(k);
+        const uint32_t v_nel = (uint32_t)ggml_nelements(v);
+        { const std::vector<uint32_t> pc = { (uint32_t)HSK, (uint32_t)nek2, (uint32_t)KV, 0, k_nel };
+          ggml_vk_dispatch_pipeline(ctx, subctx, tr_k, { k_buf, k_dst }, pc, { k_nel, 1, 1 }); }
+        { const std::vector<uint32_t> pc = { (uint32_t)HSV, (uint32_t)nev2, (uint32_t)KV, 0, v_nel };
+          ggml_vk_dispatch_pipeline(ctx, subctx, tr_v, { v_buf, v_dst }, pc, { v_nel, 1, 1 }); }
+        ggml_vk_sync_buffers(ctx, subctx);
+        k_buf = k_dst;
+        v_buf = v_dst;
+    }
+
    uint32_t mask_n_head_log2 = ((sinks != nullptr) << 24) | n_head_log2;

    if (use_mask_opt)
@@ -10995,8 +11075,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                              (uint32_t)nev2, (uint32_t)nev3,
                                              nem1, nem2, nem3,
                                              q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
-                                              k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
-                                              v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
+                                              k_stride, nbk2_eff, nbk3_eff,
+                                              v_stride, nbv2_eff, nbv3_eff,
                                              scale, max_bias, logit_softcap,
                                              mask_n_head_log2, m0, m1,
                                              gqa_ratio, split_kv, split_k };
@@ -11038,6 +11118,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                    {q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf, mask_opt_buf},
                                    pc, { workgroups_x, workgroups_y, workgroups_z });
    }
+
+    if (use_dequant_kv) {
+        ctx->prealloc_x_need_sync = true;
+    }
 }

 static vk_conv_shapes ggml_vk_conv_select_shape(ggml_backend_vk_context * ctx, uint32_t K, uint32_t NPQ) {
@@ -12192,7 +12276,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                elements = { ne, 1, 1 };
            }

-            if (pipeline == ctx->device->pipeline_cpy_transpose_32 ||
+            if (pipeline == ctx->device->pipeline_cpy_transpose_02_32 ||
+                pipeline == ctx->device->pipeline_cpy_transpose_02_16) {
+                // 32x32 tiles over dims 0 and 2; dim1 and dim3 are the batch
+                elements[0] = (uint32_t)CEIL_DIV(dst->ne[0], 32);
+                elements[1] = (uint32_t)CEIL_DIV(dst->ne[2], 32);
+                elements[2] = (uint32_t)(dst->ne[1]*dst->ne[3]);
+                elements[0] = std::min(elements[0], ctx->device->properties.limits.maxComputeWorkGroupCount[0]);
+                elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+                elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+            } else if (pipeline == ctx->device->pipeline_cpy_transpose_32 ||
                pipeline == ctx->device->pipeline_cpy_transpose_16) {
                // 32x32 tiles
                elements[0] = (uint32_t)CEIL_DIV(dst->ne[0], 32);
@@ -13120,6 +13213,7 @@ static uint32_t ggml_vk_rms_partials_size(ggml_backend_vk_context * ctx, const g
 static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *dst, const ggml_tensor *src0, const bool has_ff, bool backprop, const uint32_t set_rows_stride) {
    const int n_dims        = ((const int32_t *) dst->op_params)[1];
    const int mode          = ((const int32_t *) dst->op_params)[2];
+    const int n_offs        = ((const int32_t *) dst->op_params)[15];
    // const int n_ctx         = ((const int32_t *) dst->op_params)[3];
    const int n_ctx_orig    = ((const int32_t *) dst->op_params)[4];
    const float freq_base   = ((const float *)   dst->op_params)[5];
@@ -13149,7 +13243,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
    uint32_t nb13 = dst->nb[3] / ggml_type_size(dst->type);

    vk_op_rope_push_constants rope {
-        (uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale,
+        (uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, (uint32_t)n_offs, freq_scale,
        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff,
        { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,

@@ -19195,6 +19289,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
                    tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
                }
            }
+            const int n_offs = ((int32_t *) tensor->op_params)[15];
+            if (n_offs != 0) {
+                tensor_clone = ggml_rope_set_offset(tensor_clone, n_offs);
+            }
        } else if (tensor->op == GGML_OP_UNARY) {
            switch (ggml_get_unary_op(tensor)) {
            case GGML_UNARY_OP_EXP:
@@ -0,0 +1,61 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+// workgroup does 32x32 tile, but uses 32x8 threads
+#define TILE_DIM 32
+layout(local_size_x = 32, local_size_y = 8, local_size_z = 1) in;
+
+// +1 padding avoids shared-memory bank conflicts on the transposed read
+shared uint sh[TILE_DIM][TILE_DIM + 1];
+
+void iter(uvec3 wg_id) {
+    const uint tile_i0 = wg_id.x;   // tiles dst ne10 (== src ne00)
+    const uint tile_i2 = wg_id.y;   // tiles dst ne12 (== src ne02)
+
+    const uint tid_col = gl_LocalInvocationID.x;
+    const uint tid_row = gl_LocalInvocationID.y;
+
+    const uint i1 = wg_id.z % p.ne11;
+    const uint i3 = wg_id.z / p.ne11;
+    const uint i01 = i1;
+    const uint i03 = i3;
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i00 = tile_i0 * TILE_DIM + tid_row + 8 * y;
+        const uint i02 = tile_i2 * TILE_DIM + tid_col;
+        if (i00 < p.ne00 && i01 < p.ne01 && i02 < p.ne02 && i03 < p.ne03) {
+            const uint src_idx = i00 * p.nb00 + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+            sh[tid_row + 8 * y][tid_col] = uint(data_a[get_aoffset() + src_idx]);
+        }
+    }
+
+    barrier();
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i0 = tile_i0 * TILE_DIM + tid_col;
+        const uint i2 = tile_i2 * TILE_DIM + tid_row + 8 * y;
+        if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) {
+            const uint dst_idx = i0 * p.nb10 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+            data_d[get_doffset() + dst_idx] = D_TYPE(sh[tid_col][tid_row + 8 * y]);
+        }
+    }
+}
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+void main() {
+    bool need_barrier = false;
+    for (uint z = gl_WorkGroupID.z; z < p.ne11 * p.ne13; z += gl_NumWorkGroups.z) {
+        for (uint y = gl_WorkGroupID.y; y < CEIL_DIV(p.ne12, TILE_DIM); y += gl_NumWorkGroups.y) {
+            for (uint x = gl_WorkGroupID.x; x < CEIL_DIV(p.ne10, TILE_DIM); x += gl_NumWorkGroups.x) {
+                if (need_barrier) {
+                    barrier();
+                }
+                need_barrier = true;
+                iter(uvec3(x, y, z));
+            }
+        }
+    }
+}
@@ -18,7 +18,18 @@ void main() {
        return;
    }

+#ifdef DEQUANT_TRANSPOSE
+    // read [HS, NH, KV, NS], write [HS, KV, NH, NS]
+    const uint HS = p.M, NH = p.K, KVn = p.stride_a;
+    const uint e0 = ib * 32;
+    const uint b_idx = (e0 % HS)
+                     + ((e0 / (HS * NH)) % KVn) * HS
+                     + ((e0 / HS) % NH) * (HS * KVn)
+                     + (e0 / (HS * NH * KVn)) * (HS * KVn * NH)
+                     + 16 * il;
+#else
    const uint b_idx = 1024*i + 32*ir + 16*il;
+#endif

    const float d = float(data_a[ib].d);

@@ -50,19 +50,21 @@ void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_
    }
    idst += p.d_offset;

-    if (i0 >= p.n_dims) {
+    if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
        rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
        rope_data_d[idst + 1] = ROPE_D_TYPE(rope_data_a[ix + 1]);

        return;
    }

-    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
+    const uint iw = i0 - p.n_offs; // relative idx

-    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, iw/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;

    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+    rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);

    const float x0 = float(rope_data_a[ix + 0]);
    const float x1 = float(rope_data_a[ix + 1]);
@@ -87,25 +89,28 @@ void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_
    }
    idst += p.d_offset;

-    if (i0 >= p.n_dims) {
+    if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);

        return;
    }

-    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
+    const uint iw = i0 - p.n_offs; // relative idx

-    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, iw/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;

    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+    rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);

-    const float x0 = float(rope_data_a[ix + 0]);
-    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is p.n_offs + iw/2 = i0/2 + p.n_offs/2
+    const float x0 = float(rope_data_a[ix + p.n_offs/2 + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_offs/2 + p.n_dims/2]);

-    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
-    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+    rope_data_d[idst + p.n_offs/2 + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_offs/2 + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }


@@ -125,53 +130,56 @@ void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope
    }
    idst += p.d_offset;

-    if (i0 >= p.n_dims) {
+    if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);

        return;
    }

+    const uint iw = i0 - p.n_offs; // relative idx
+
    const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3];
    const int sec_w = p.sections[1] + p.sections[0];
-    const uint sector = (i0 / 2) % sect_dims;
+    const uint sector = (iw / 2) % sect_dims;

    float theta_base = 0.0;
    if (p.is_imrope != 0) {
        if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, iw/2.0f);
        } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, iw/2.0f);
        } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
-            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, iw/2.0f);
        } else {
-            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, iw/2.0f);
        }
    } else {
        if (sector < p.sections[0]) {
-            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, iw/2.0f);
        }
        else if (sector >= p.sections[0] && sector < sec_w) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, iw/2.0f);
        }
        else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, iw/2.0f);
        }
        else if (sector >= sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, iw/2.0f);
        }
    }

-    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;

    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+    rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);

-    const float x0 = float(rope_data_a[ix + 0]);
-    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is p.n_offs + iw/2 = i0/2 + p.n_offs/2
+    const float x0 = float(rope_data_a[ix + p.n_offs/2 + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_offs/2 + p.n_dims/2]);

-    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
-    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+    rope_data_d[idst + p.n_offs/2 + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_offs/2 + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }

 void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
@@ -5,6 +5,7 @@ struct rope_params {
    uint rope_mode;
    uint nrows;
    uint n_dims;
+    uint n_offs;
    float freq_scale;
    float freq_base;
    float ext_factor;
@@ -780,6 +780,10 @@ void process_shaders() {
        if (tname != "f16" && tname != "bf16") {
            string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
        }
+        // Fused dequant+transpose variant for FA quant-KV (per-head-contiguous f16 scratch).
+        if (tname == "q8_0") {
+            string_to_spv("dequant_" + tname + "_transpose", "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}, {"DEQUANT_TRANSPOSE", "1"}}));
+        }

        shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";

@@ -826,6 +830,8 @@ void process_shaders() {

    string_to_spv("cpy_transpose_16", "copy_transpose.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
    string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
+    string_to_spv("cpy_transpose_02_16", "copy_transpose_02.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
+    string_to_spv("cpy_transpose_02_32", "copy_transpose_02.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});

    for (std::string t : {"q1_0", "q2_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
        string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"S_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -954,10 +954,11 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key {
    int       vectorized;
    uint32_t  num_cols;
    bool      use_mmvq;
+    bool      src_overlap;

    bool operator==(const ggml_webgpu_mul_mat_vec_pipeline_key & other) const {
        return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized &&
-               num_cols == other.num_cols && use_mmvq == other.use_mmvq;
+               num_cols == other.num_cols && use_mmvq == other.use_mmvq && src_overlap == other.src_overlap;
    }
 };

@@ -969,6 +970,7 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key_hash {
        ggml_webgpu_hash_combine(seed, key.vectorized);
        ggml_webgpu_hash_combine(seed, key.num_cols);
        ggml_webgpu_hash_combine(seed, key.use_mmvq);
+        ggml_webgpu_hash_combine(seed, key.src_overlap);
        return seed;
    }
 };
@@ -977,6 +979,7 @@ struct ggml_webgpu_mul_mat_vec_shader_decisions {
    uint32_t wg_size;
    uint32_t outputs_per_wg;
    uint32_t vec_size;
+    bool     src_overlap = false;
 };

 struct ggml_webgpu_quantize_q8_pipeline_key {
@@ -998,10 +1001,11 @@ struct ggml_webgpu_mul_mat_pipeline_key {
    ggml_type src1_type;
    int       vectorized;
    int       use_subgroup_matrix;
+    bool      src_overlap;

    bool operator==(const ggml_webgpu_mul_mat_pipeline_key & other) const {
        return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized &&
-               use_subgroup_matrix == other.use_subgroup_matrix;
+               use_subgroup_matrix == other.use_subgroup_matrix && src_overlap == other.src_overlap;
    }
 };

@@ -1012,6 +1016,7 @@ struct ggml_webgpu_mul_mat_pipeline_key_hash {
        ggml_webgpu_hash_combine(seed, key.src1_type);
        ggml_webgpu_hash_combine(seed, key.vectorized);
        ggml_webgpu_hash_combine(seed, key.use_subgroup_matrix);
+        ggml_webgpu_hash_combine(seed, key.src_overlap);
        return seed;
    }
 };
@@ -1034,6 +1039,7 @@ struct ggml_webgpu_mul_mat_shader_decisions {
    uint32_t subgroup_matrix_n;

    uint32_t mul_mat_wg_size;
+    bool     src_overlap = false;
 };

 /** MUL_MAT_ID **/
@@ -1950,7 +1956,7 @@ class ggml_webgpu_shader_lib {
        return quantize_q8_pipelines[key];
    }

-    webgpu_pipeline get_mul_mat_vec_pipeline(const ggml_webgpu_shader_lib_context & context) {
+    webgpu_pipeline get_mul_mat_vec_pipeline(const ggml_webgpu_shader_lib_context & context, bool src_overlap) {
        ggml_webgpu_mul_mat_vec_pipeline_key key = {};
        key.src0_type                            = context.src0->type;
        key.src1_type                            = context.src1->type;
@@ -1961,6 +1967,7 @@ class ggml_webgpu_shader_lib {
        key.num_cols   = context.dst->ne[1];
        key.use_mmvq =
            ggml_webgpu_can_use_mmvq(context.src0, context.src1, context.supports_dot_product, context.vendor);
+        key.src_overlap = src_overlap;

        auto it = mul_mat_vec_pipelines.find(key);
        if (it != mul_mat_vec_pipelines.end()) {
@@ -2068,6 +2075,11 @@ class ggml_webgpu_shader_lib {
            defines.push_back("Q8_1_T");
        }

+        if (key.src_overlap) {
+            defines.push_back("SRC_OVERLAP");
+            variant += "_src_overlap";
+        }
+
        defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
        defines.push_back(std::string("OUTPUTS_PER_WG=") + std::to_string(outputs_per_wg));
        defines.push_back(context.supports_subgroups ? "USE_SUBGROUP_REDUCTION" : "USE_WORKGROUP_REDUCTION");
@@ -2089,7 +2101,7 @@ class ggml_webgpu_shader_lib {
        return mul_mat_vec_pipelines[key];
    }

-    webgpu_pipeline get_mul_mat_fast_pipeline(const ggml_webgpu_shader_lib_context & context) {
+    webgpu_pipeline get_mul_mat_fast_pipeline(const ggml_webgpu_shader_lib_context & context, bool src_overlap) {
        ggml_webgpu_mul_mat_pipeline_key key = {};
        key.src0_type                        = context.src0->type;
        key.src1_type                        = context.src1->type;
@@ -2098,6 +2110,7 @@ class ggml_webgpu_shader_lib {
                                      1 :
                                      0;
        key.use_subgroup_matrix = context.supports_subgroup_matrix;
+        key.src_overlap         = src_overlap;

        auto it = mul_mat_fast_pipelines.find(key);
        if (it != mul_mat_fast_pipelines.end()) {
@@ -2216,6 +2229,11 @@ class ggml_webgpu_shader_lib {
            variant += "_vectorized";
        }

+        if (key.src_overlap) {
+            defines.push_back("SRC_OVERLAP");
+            variant += "_src_overlap";
+        }
+
        if (!key.use_subgroup_matrix) {
            defines.push_back("WORKGROUP_SIZE_M=" + std::to_string(WEBGPU_MUL_MAT_WG_SIZE_M) + "u");
            defines.push_back("WORKGROUP_SIZE_N=" + std::to_string(WEBGPU_MUL_MAT_WG_SIZE_N) + "u");
@@ -1628,48 +1628,65 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context & ctx,
    // Get or create pipeline
    webgpu_pipeline                   pipeline;
    std::vector<webgpu_dispatch_desc> dispatches;
+    const bool src_overlap = ggml_webgpu_tensor_binding_overlap(ctx->global_ctx, src0, src1) && !use_mmvq;

    if (use_mat_vec) {
        if (use_mmvq) {
            ggml_webgpu_quantize_q8_dispatch(ctx, src0, src1, dst, dispatches);
        }
-        pipeline = ctx->shader_lib->get_mul_mat_vec_pipeline(shader_lib_ctx);
+        pipeline = ctx->shader_lib->get_mul_mat_vec_pipeline(shader_lib_ctx, src_overlap);
    } else {
-        pipeline = ctx->shader_lib->get_mul_mat_fast_pipeline(shader_lib_ctx);
+        pipeline = ctx->shader_lib->get_mul_mat_fast_pipeline(shader_lib_ctx, src_overlap);
+    }
+
+    uint32_t offset_src0   = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type));
+    uint32_t offset_src1   = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type));
+    size_t   merged_offset = 0;
+    size_t   merged_size   = 0;
+    if (src_overlap) {
+        const ggml_webgpu_merged_binding_range merged_range =
+            ggml_webgpu_tensor_merged_binding_range(ctx, { src0, src1 });
+        merged_offset = merged_range.offset;
+        merged_size   = merged_range.size;
+        offset_src0   = ggml_webgpu_tensor_merged_element_offset(src0, merged_range);
+        offset_src1   = ggml_webgpu_tensor_merged_element_offset(src1, merged_range);
    }

    // Build params
-    std::vector<uint32_t> params = {
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
-        (uint32_t) dst->ne[0],
-        (uint32_t) dst->ne[1],
-        (uint32_t) src0->ne[0],
-        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
-        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
-        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
-        (uint32_t) src0->ne[2],
-        (uint32_t) src0->ne[3],
-        (uint32_t) (src1->ne[2] / src0->ne[2]),
-        (uint32_t) (src1->ne[3] / src0->ne[3])
-    };
+    std::vector<uint32_t> params = { offset_src0,
+                                     offset_src1,
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+                                     (uint32_t) dst->ne[0],
+                                     (uint32_t) dst->ne[1],
+                                     (uint32_t) src0->ne[0],
+                                     (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
+                                     (uint32_t) src0->ne[2],
+                                     (uint32_t) src0->ne[3],
+                                     (uint32_t) (src1->ne[2] / src0->ne[2]),
+                                     (uint32_t) (src1->ne[3] / src0->ne[3]) };

    // Build bind group entries
    std::vector<wgpu::BindGroupEntry> entries = {};
-
-    entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
    if (use_mmvq) {
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
        auto & mmvq_qq8_entry = dispatches[0].bind_group_entries[1];
        entries.push_back(ggml_webgpu_make_bind_group_entry(1, ggml_webgpu_tensor_buf(dst), mmvq_qq8_entry.offset,
                                                            mmvq_qq8_entry.size));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
+    } else if (src_overlap) {
+        entries.push_back(
+            ggml_webgpu_make_bind_group_entry(0, ggml_webgpu_tensor_buf(src0), merged_offset, merged_size));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, dst));
    } else {
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
    }
-    entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));

    // Calculate workgroup dimensions
    uint32_t       wg_x           = 1;
@@ -4455,7 +4472,9 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
            supports_op = (op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32) && ggml_is_contiguous_rows(src0);
            break;
        case GGML_OP_ROPE:
-            supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
+            // FIXME: support ggml_rope_set_offset
+            supports_op =
+                (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && ((const int32_t *) op->op_params)[15] == 0;
            break;
        case GGML_OP_GLU:
            switch (ggml_get_glu_op(op)) {
@@ -1,3 +1,7 @@
+#ifndef SRC0
+#define SRC0 src0
+#endif
+
 #ifdef BYTE_HELPERS
 fn get_byte(value: u32, index: u32) -> u32 {
    return (value >> (index * 8)) & 0xFF;
@@ -46,7 +50,7 @@ fn load_f16_as_f32_at_src(byte_offset: u32) -> f32 {

 #ifdef DECLARE_BYTE_LOADERS_SRC0
 fn load_u16_at_src0(byte_offset: u32) -> u32 {
-    let word = src0[byte_offset / 4u];
+    let word = SRC0[byte_offset / 4u];
    let shift = (byte_offset & 0x2u) * 8u;
    return (word >> shift) & 0xFFFFu;
 }
@@ -55,14 +59,14 @@ fn load_u16_at_src0(byte_offset: u32) -> u32 {
 // Caller extracts the 16-bit half it needs via & 0xFFFFu or >> 16u.
 // this is used in k-quants for better performance
 fn load_u32_at_src0_aligned(byte_offset: u32) -> u32 {
-    return src0[(byte_offset & ~3u) / 4u];
+    return SRC0[(byte_offset & ~3u) / 4u];
 }

 fn load_u32_at_src0(byte_offset: u32) -> u32 {
    let word_idx = byte_offset / 4u;
    let shift = (byte_offset & 0x3u) * 8u;
-    let lo = src0[word_idx];
-    let hi = src0[word_idx + 1u];
+    let lo = SRC0[word_idx];
+    let hi = SRC0[word_idx + 1u];
    let shifted = (lo >> shift) | (hi << (32u - shift));
    return select(shifted, lo, shift == 0u);
 }
@@ -73,7 +77,7 @@ fn load_f16_at_src0(byte_offset: u32) -> f16 {
 }

 fn load_f16_as_f32_at_src0(byte_offset: u32) -> f32 {
-    let word = src0[byte_offset / 4u];
+    let word = SRC0[byte_offset / 4u];
    let shift = (byte_offset & 0x2u) * 8u;
    let d_bits = (word >> shift) & 0xFFFFu;
    return unpack2x16float(d_bits)[0];
@@ -1,3 +1,10 @@
+#ifndef SRC0
+#define SRC0 src0
+#endif
+#ifndef SRC1
+#define SRC1 src1
+#endif
+
 #ifdef VEC
 #define VEC_SIZE 4
 #define SHMEM_TYPE vec4<f16>
@@ -39,7 +46,7 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
        let src0_idx = batch_offset + global_m * params.stride_01 + global_k;
        let src0_val = select( // taking a slight performance hit to avoid oob
            SRC0_TYPE(0.0),
-            src0[src0_idx/VEC_SIZE],
+            SRC0[src0_idx/VEC_SIZE],
            global_m < params.m && global_k < params.k);
        store_shmem(SHMEM_TYPE(src0_val), elem_idx);
    }
@@ -57,7 +64,7 @@ fn init_shmem_src1(thread_id: u32, batch_offset: u32, offset_n: u32, k_outer: u3
        let src1_idx = batch_offset + global_n * params.stride_11 + global_k;
        let src1_val = select(
            SRC1_TYPE(0.0),
-            src1[src1_idx/VEC_SIZE],
+            SRC1[src1_idx/VEC_SIZE],
            global_n < params.n && global_k < params.k);
        store_shmem(SHMEM_TYPE(src1_val), TILE_SRC0_SHMEM + elem_idx);
    }
@@ -1,8 +1,12 @@
 enable f16;

 #define DECLARE_BYTE_LOADERS_SRC0
-#include "common_decls.tmpl"

+#ifdef SRC_OVERLAP
+#define SRC0 merged_src
+#define SRC1 merged_src
+#endif
+#include "common_decls.tmpl"
 #include "mul_mat_decls.tmpl"

 #ifdef VEC
@@ -36,11 +40,17 @@ struct MulMatParams {
    broadcast3: u32
 };

+#ifdef SRC_OVERLAP
+@group(0) @binding(0) var<storage, read_write> merged_src: array<SRC0_TYPE>;
+#define DST_BINDING 1
+#else
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+#define DST_BINDING 2
+#endif

-@group(0) @binding(3) var<uniform> params: MulMatParams;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+@group(0) @binding(DST_BINDING + 1) var<uniform> params: MulMatParams;

 fn get_local_n(thread_id: u32) -> u32 {
    return thread_id / WORKGROUP_SIZE_M;
@@ -4,6 +4,10 @@ enable subgroups;
 enable chromium_experimental_subgroup_matrix;

 #define DECLARE_BYTE_LOADERS_SRC0
+#ifdef SRC_OVERLAP
+#define SRC0 merged_src
+#define SRC1 merged_src
+#endif
 #include "common_decls.tmpl"

 #include "mul_mat_decls.tmpl"
@@ -48,11 +52,17 @@ struct MulMatParams {
 };

 // SRC0_TYPE and SRC1_TYPE are defined in mul_mat_decls, which is included
+#ifdef SRC_OVERLAP
+@group(0) @binding(0) var<storage, read_write> merged_src: array<SRC0_TYPE>;
+#define DST_BINDING 1
+#else
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+#define DST_BINDING 2
+#endif

-@group(0) @binding(3) var<uniform> params: MulMatParams;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+@group(0) @binding(DST_BINDING + 1) var<uniform> params: MulMatParams;

 const WG_M_SG_TILE_SIZE = SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
 const WG_N_SG_TILE_SIZE = SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
@@ -7,6 +7,11 @@ enable f16;
 requires packed_4x8_integer_dot_product;
 #endif

+#ifdef SRC_OVERLAP
+#define SRC0 merged_src
+#define SRC1 merged_src
+#endif
+
 #define DECLARE_BYTE_LOADERS_SRC0
 #include "common_decls.tmpl"

@@ -35,17 +40,22 @@ struct MulMatParams {
    broadcast3: u32
 };

+#if defined(MMVQ)
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>;
-
-#ifdef MMVQ
@group(0) @binding(1) var<storage, read_write> src1q: array<q8_1>;
+#define DST_BINDING 2
+#elif defined(SRC_OVERLAP)
+@group(0) @binding(0) var<storage, read_write> merged_src: array<SRC0_TYPE>;
+#define DST_BINDING 1
 #else
+@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>;
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>;
+#define DST_BINDING 2
 #endif

-@group(0) @binding(2) var<storage, read_write> dst: array<f32>;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<f32>;
 // "mul_mat_vec_acc.tmpl" requires params.k, params.m, params.stride_01
-@group(0) @binding(3) var<uniform> params: MulMatParams;
+@group(0) @binding(DST_BINDING + 1) var<uniform> params: MulMatParams;

 // Flattened as [row][thread] to keep each row's reduction contiguous in memory.
 var<workgroup> partial_sums: array<f32, OUTPUTS_PER_WG * WG_SIZE>;
@@ -1,3 +1,10 @@
+#ifndef SRC0
+#define SRC0 src0
+#endif
+#ifndef SRC1
+#define SRC1 src1
+#endif
+
 #ifdef U32_DEQUANT_HELPERS
 #define SRC0_TYPE u32

@@ -43,13 +50,13 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
    for (var k = thread_id; k < k_vec; k += WG_SIZE) {
        var x_vals: array<SRC1_TYPE, NUM_COLS>;
        for (var col = 0u;col < NUM_COLS;col += 1) {
-            x_vals[col] = src1[src1_idx_base_vec + col * (params.stride_11 / VEC_SIZE) + k];
+            x_vals[col] = SRC1[src1_idx_base_vec + col * (params.stride_11 / VEC_SIZE) + k];
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
            let output_row = row_base + row;
            if (output_row < params.m) {
                let src0_idx = (src0_batch_offset + output_row * params.stride_01) / VEC_SIZE + k;
-                let w = src0[src0_idx];
+                let w = SRC0[src0_idx];
                for (var col = 0u;col < NUM_COLS;col += 1) {
                    acc[col][row] += inner_dot(w, x_vals[col]);
                }
@@ -76,7 +83,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -116,8 +123,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -160,8 +167,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -205,8 +212,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -253,8 +260,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -302,7 +309,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -347,7 +354,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -409,10 +416,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 4u; i++) {
-                x_block[col][i]       = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4u]  = f32(src1[x_base + col * params.stride_11 + 32u + i]);
-                x_block[col][i + 8u]  = f32(src1[x_base + col * params.stride_11 + 64u + i]);
-                x_block[col][i + 12u] = f32(src1[x_base + col * params.stride_11 + 96u + i]);
+                x_block[col][i]       = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4u]  = f32(SRC1[x_base + col * params.stride_11 + 32u + i]);
+                x_block[col][i + 8u]  = f32(SRC1[x_base + col * params.stride_11 + 64u + i]);
+                x_block[col][i + 12u] = f32(SRC1[x_base + col * params.stride_11 + 96u + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -518,8 +525,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 8u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 8u] = f32(src1[x_base + col * params.stride_11 + 32u + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 8u] = f32(SRC1[x_base + col * params.stride_11 + 32u + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -610,10 +617,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        for (var col = 0u; col < NUM_COLS;col += 1) {
            let col_base = x_base + col * params.stride_11;
            for (var i = 0u; i < 4u; i++) {
-                x_block[col][i]       = f32(src1[col_base + i]);
-                x_block[col][i + 4u]  = f32(src1[col_base + 32u + i]);
-                x_block[col][i + 8u]  = f32(src1[col_base + 128u + i]);
-                x_block[col][i + 12u] = f32(src1[col_base + 160u + i]);
+                x_block[col][i]       = f32(SRC1[col_base + i]);
+                x_block[col][i + 4u]  = f32(SRC1[col_base + 32u + i]);
+                x_block[col][i + 8u]  = f32(SRC1[col_base + 128u + i]);
+                x_block[col][i + 12u] = f32(SRC1[col_base + 160u + i]);
            }
        }

@@ -713,10 +720,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        for (var col = 0u; col < NUM_COLS;col += 1) {
            let col_base = x_base + col * params.stride_11;
            for (var i = 0u; i < 4u; i++) {
-                x_block[col][i]       = f32(src1[col_base + i]);
-                x_block[col][i + 4u]  = f32(src1[col_base + 32u + i]);
-                x_block[col][i + 8u]  = f32(src1[col_base + 128u + i]);
-                x_block[col][i + 12u] = f32(src1[col_base + 160u + i]);
+                x_block[col][i]       = f32(SRC1[col_base + i]);
+                x_block[col][i + 4u]  = f32(SRC1[col_base + 32u + i]);
+                x_block[col][i + 8u]  = f32(SRC1[col_base + 128u + i]);
+                x_block[col][i + 12u] = f32(SRC1[col_base + 160u + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -823,10 +830,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        for (var col = 0u; col < NUM_COLS;col += 1) {
            let col_base = x_base + col * params.stride_11;
            for (var l = 0u; l < 4u; l++) {
-                x_block[col][l]       = f32(src1[col_base + l]);
-                x_block[col][l + 4u]  = f32(src1[col_base + 32u + l]);
-                x_block[col][l + 8u]  = f32(src1[col_base + 64u + l]);
-                x_block[col][l + 12u] = f32(src1[col_base + 96u + l]);
+                x_block[col][l]       = f32(SRC1[col_base + l]);
+                x_block[col][l + 4u]  = f32(SRC1[col_base + 32u + l]);
+                x_block[col][l + 8u]  = f32(SRC1[col_base + 64u + l]);
+                x_block[col][l + 12u] = f32(SRC1[col_base + 96u + l]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -899,7 +906,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -960,7 +967,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1039,7 +1046,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1101,7 +1108,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1168,7 +1175,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1234,7 +1241,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1302,7 +1309,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1367,8 +1374,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2u; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4u] = f32(src1[x_base + col * params.stride_11 + i + 16u]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4u] = f32(SRC1[x_base + col * params.stride_11 + i + 16u]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1418,7 +1425,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1476,8 +1483,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1521,8 +1528,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 8] = f32(src1[x_base + col * params.stride_11 + i + 8]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 8] = f32(SRC1[x_base + col * params.stride_11 + i + 8]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -4200,7 +4200,7 @@ static struct ggml_tensor * ggml_rope_impl(

    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

-    int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
+    int32_t params[16] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
    memcpy(params +  5, &freq_base,    sizeof(float));
    memcpy(params +  6, &freq_scale,   sizeof(float));
    memcpy(params +  7, &ext_factor,   sizeof(float));
@@ -4212,6 +4212,8 @@ static struct ggml_tensor * ggml_rope_impl(
    } else {
        memset(params + 11, 0,         sizeof(int32_t) * GGML_MROPE_SECTIONS);
    }
+    params[15] = 0; // n_offs, set via ggml_rope_set_offset()
+
    ggml_set_op_params(result, params, sizeof(params));

    result->op     = GGML_OP_ROPE;
@@ -4422,6 +4424,20 @@ struct ggml_tensor * ggml_rope_multi_back(
    result->op = GGML_OP_ROPE_BACK;
    return result;
 }
+
+struct ggml_tensor * ggml_rope_set_offset(
+        struct ggml_tensor  * a,
+        int                   n_offs) {
+    GGML_ASSERT(a->op == GGML_OP_ROPE || a->op == GGML_OP_ROPE_BACK);
+    GGML_ASSERT(n_offs >= 0);
+
+    const int32_t mode = ggml_get_op_params_i32(a, 2);
+    GGML_ASSERT(mode != GGML_ROPE_TYPE_VISION);
+
+    ggml_set_op_params_i32(a, 15, n_offs);
+    return a;
+}
+
 // ggml_clamp

 struct ggml_tensor * ggml_clamp(
@@ -162,6 +162,7 @@ class Keys:
        TARGET_LAYERS                     = "{arch}.target_layers"
        TARGET_HIDDEN_SIZE                = "{arch}.target_hidden_size"
        BLOCK_SIZE                        = "{arch}.block_size"
+        SAMPLE_FROM_ANCHOR                = "{arch}.sample_from_anchor"
        NORM_BEFORE_RESIDUAL              = "{arch}.norm_before_residual"
        NORM_BEFORE_FC                    = "{arch}.norm_before_fc"

@@ -207,6 +208,7 @@ class Keys:
        SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
        SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
        TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
+        ROPE_PATTERN                 = "{arch}.attention.rope_pattern"

        class Indexer:
            HEAD_COUNT = "{arch}.attention.indexer.head_count"
@@ -548,6 +550,7 @@ class MODEL_ARCH(IntEnum):
    GRANITE_MOE      = auto()
    GRANITE_HYBRID   = auto()
    GRANITE_SWITCH   = auto()
+    GRANITE_SWA      = auto()
    CHAMELEON        = auto()
    WAVTOKENIZER_DEC = auto()
    PLM              = auto()
@@ -1264,6 +1267,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.GRANITE_MOE:      "granitemoe",
    MODEL_ARCH.GRANITE_HYBRID:   "granitehybrid",
    MODEL_ARCH.GRANITE_SWITCH:   "graniteswitch",
+    MODEL_ARCH.GRANITE_SWA:      "granite_swa",
    MODEL_ARCH.CHAMELEON:        "chameleon",
    MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
    MODEL_ARCH.PLM:              "plm",
@@ -4151,6 +4155,31 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.GRANITE_SWA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_SINKS,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        # MoE (GraniteMoeSWA)
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        # Shared expert - gate+up kept fused in FFN_UP_SHEXP (LLM_FFN_SWIGLU)
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+    ],
    MODEL_ARCH.CHAMELEON: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -4819,6 +4848,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
    ],
    MODEL_ARCH.DFLASH: [
        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_NORM,
        MODEL_TENSOR.ATTN_Q,
@@ -4858,6 +4888,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_UP_SHEXP,
        MODEL_TENSOR.FC,
        MODEL_TENSOR.ENC_OUTPUT_NORM,
+        MODEL_TENSOR.D2T,
        # optional DSpark heads
        MODEL_TENSOR.DSPARK_MARKOV_W1,
        MODEL_TENSOR.DSPARK_MARKOV_W2,
@@ -32,6 +32,10 @@ from gguf.constants import (
    GGUFEndian,
 )

+# limits mirroring ggml/src/gguf.cpp (not part of gguf.h)
+GGUF_MAX_STRING_LENGTH  = 1024 * 1024 * 1024
+GGUF_MAX_ARRAY_ELEMENTS = 1024 * 1024 * 1024
+
 logger = logging.getLogger(__name__)

 READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION]
@@ -167,6 +171,10 @@ class GGUFReader:
        offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
        offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
        tensor_count, kv_count = temp_counts
+        if tensor_count > GGUF_MAX_ARRAY_ELEMENTS:
+            raise ValueError(f'Tensor count {tensor_count} exceeds maximum {GGUF_MAX_ARRAY_ELEMENTS}')
+        if kv_count > GGUF_MAX_ARRAY_ELEMENTS:
+            raise ValueError(f'KV count {kv_count} exceeds maximum {GGUF_MAX_ARRAY_ELEMENTS}')
        offs = self._build_fields(offs, kv_count)

        # Build Tensor Info Fields
@@ -217,6 +225,10 @@ class GGUFReader:

    def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]:
        slen = self._get(offset, np.uint64)
+        if int(slen[0]) > GGUF_MAX_STRING_LENGTH:
+            raise ValueError(f'String length {int(slen[0])} exceeds maximum {GGUF_MAX_STRING_LENGTH}')
+        if offset + 8 + int(slen[0]) > self.data.nbytes:
+            raise ValueError(f'String length {int(slen[0])} exceeds remaining file size {self.data.nbytes - offset - 8}')
        return slen, self._get(offset + 8, np.uint8, slen[0])

    def _get_field_parts(
@@ -241,6 +253,8 @@ class GGUFReader:
            raw_itype = self._get(offs, np.uint32)
            offs += int(raw_itype.nbytes)
            alen = self._get(offs, np.uint64)
+            if int(alen[0]) > GGUF_MAX_ARRAY_ELEMENTS:
+                raise ValueError(f'Array length {int(alen[0])} exceeds maximum {GGUF_MAX_ARRAY_ELEMENTS}')
            offs += int(alen.nbytes)
            aparts: list[npt.NDArray[Any]] = [raw_itype, alen]
            data_idxs: list[int] = []
@@ -824,6 +824,9 @@ class GGUFWriter:
        else:
            self.add_array(key, value)

+    def add_rope_pattern(self, value: Sequence[bool]) -> None:
+        self.add_array(Keys.Attention.ROPE_PATTERN.format(arch=self.arch), value)
+
    def add_dense_features_dims(self, dense:str, in_f:int, out_f:int) -> None:
        self.add_uint32(Keys.LLM.DENSE_FEAT_IN_SIZE.format(arch=self.arch, dense=dense), in_f)
        self.add_uint32(Keys.LLM.DENSE_FEAT_OUT_SIZE.format(arch=self.arch, dense=dense), out_f)
@@ -981,6 +984,9 @@ class GGUFWriter:
    def add_block_size(self, value: int) -> None:
        self.add_uint32(Keys.LLM.BLOCK_SIZE.format(arch=self.arch), value)

+    def add_sample_from_anchor(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.SAMPLE_FROM_ANCHOR.format(arch=self.arch), value)
+
    def add_target_layers(self, value: Sequence[int]) -> None:
        self.add_array(Keys.LLM.TARGET_LAYERS.format(arch=self.arch), value)

@@ -76,14 +76,14 @@ class TensorNameMap:
        # Output
        MODEL_TENSOR.OUTPUT: (
            "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2 phimoe plamo2
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2 phimoe plamo2 llama4
            "output",                    # llama-pth bloom internlm2
            "word_embeddings_for_head",  # persimmon
            "lm_head.linear",            # phi2
            "output_layer",              # chatglm
            "head",                      # rwkv
            "head.out",                  # wavtokenizer
-            "lm_head",                   # llama4
+            "model.lm_head",             # dflash
            "model.transformer.ff_out",  # llada
            "head.decoder",              # modern-bert
        ),
@@ -458,6 +458,7 @@ class TensorNameMap:
            "transformer.decoder_layer.{bid}.router",           # Grok
            "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
+            "model.layers.{bid}.block_sparse_moe.router",       # granite_swa
            "model.layers.{bid}.feed_forward.router",           # llama4 jamba
            "encoder.layers.{bid}.mlp.router.layer",            # nomic-bert-moe
            "model.layers.{bid}.mlp.router",                    # openai-moe
@@ -71,6 +71,26 @@ if git ls-remote --tags origin "${VERSION}" | grep -q "${VERSION}"; then
 fi
 echo "Tag ${VERSION} does not exist on remote - OK"

+echo "Checking release.yml status for commit ${SHA}..."
+if [[ -z "${GITHUB_REPOSITORY:-}" ]]; then
+    echo "Warning: GITHUB_REPOSITORY not set - skipping CI check (local run)"
+else
+    RUNS=$(gh api "repos/${GITHUB_REPOSITORY}/actions/workflows/release.yml/runs?per_page=100" \
+        --jq "[.workflow_runs[] | select(.head_sha == \"${SHA}\" and .conclusion == \"success\")] | length")
+    if [[ "$RUNS" -eq 0 ]]; then
+        if [[ "$DRY_RUN" == "true" ]]; then
+            echo "Warning: no successful release.yml run found for HEAD (${SHA}) (dry run, continuing)."
+            CHECKS_PASSED=false
+        else
+            echo "Error: no successful release.yml run found for HEAD (${SHA})"
+            echo "The nightly build must complete successfully before making a release."
+            exit 1
+        fi
+    else
+        echo "Found successful release.yml run for HEAD."
+    fi
+fi
+
 MAJOR=$(grep "set(GGML_VERSION_MAJOR" "$REPO_ROOT/ggml/CMakeLists.txt" | grep -oP '\d+')
 MINOR=$(grep "set(GGML_VERSION_MINOR" "$REPO_ROOT/ggml/CMakeLists.txt" | grep -oP '\d+')
 PATCH=$(grep "set(GGML_VERSION_PATCH" "$REPO_ROOT/ggml/CMakeLists.txt" | grep -oP '\d+')
@@ -0,0 +1,87 @@
+#!/bin/bash
+# Generate the description of a release: the previous release version, the
+# change log and the link to the nightly release corresponding to the commit being released.
+#
+# Usage: make-release-desc.sh <version>
+#   <version>: current release version (v<maj>.<min>.<pat>, the leading v is optional)
+#
+# The previous version is the highest plain semver tag (v<maj>.<min>.<pat>)
+# strictly below <version>. The change log lists all commits between the
+# previous version tag and the release commit, one line per commit.
+#
+# The release commit is the commit <version> points at when the tag exists,
+# HEAD otherwise. The nightly release is the b* tag pointing at that commit
+# (release.yml tags the same commit); the link is only generated when that
+# tag exists.
+#
+# Env (when running in GitHub Actions):
+#   GITHUB_OUTPUT: previous_tag, changelog_title, changelog and nightly are written here
+#   GITHUB_REPOSITORY: owner/repo, used to build the nightly release URL (skipped when unset)
+set -euo pipefail
+
+if [[ $# -ne 1 ]]; then
+    echo "Usage: $(basename "$0") <version>"
+    exit 1
+fi
+VERSION="$1"
+
+# Accept the version with or without the leading v, reject anything else
+if [[ "${VERSION}" =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
+    VERSION="v${VERSION}"
+elif [[ ! "${VERSION}" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
+    echo "Error: invalid version '${VERSION}' (expected v<maj>.<min>.<pat>)"
+    exit 1
+fi
+
+# Make sure all remote tags are available locally (skipped on local runs without origin)
+if ! git fetch --tags origin 2>/dev/null; then
+    echo "Warning: could not fetch tags from origin (local run?)"
+fi
+
+# Release commit: the commit <version> points at when the tag exists, HEAD otherwise.
+if ! RELEASE_COMMIT="$(git rev-parse -q --verify "refs/tags/${VERSION}^{commit}" 2>/dev/null)"; then
+    RELEASE_COMMIT="$(git rev-parse HEAD)"
+fi
+
+echo "Release commit: $(git rev-parse --short "${RELEASE_COMMIT}")"
+
+PREV="$( { git tag --list; echo "${VERSION}"; } \
+    | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' \
+    | sort -V \
+    | awk -v cur="${VERSION}" '$0 == cur { exit } { prev = $0 } END { print prev }')"
+
+if [[ -n "${PREV}" ]]; then
+    CHANGELOG="$(git log --oneline "${PREV}..${RELEASE_COMMIT}")"
+    CHANGELOG_TITLE="Change log since ${PREV}"
+else
+    CHANGELOG="(no previous release tag found)"
+    CHANGELOG_TITLE="Change log"
+fi
+
+# Nightly release: the b* tag pointing at the release commit (|| true: no match is not an error)
+NIGHTLY_TAG="$(git tag --points-at "${RELEASE_COMMIT}" | grep -E '(^|-)b[0-9]+(-[0-9a-f]{7})?$' | head -n 1 || true)"
+
+NIGHTLY=""
+if [[ -n "${NIGHTLY_TAG}" ]]; then
+    if [[ -n "${GITHUB_REPOSITORY:-}" ]]; then
+        NIGHTLY_URL="https://github.com/${GITHUB_REPOSITORY}/releases/tag/${NIGHTLY_TAG}"
+        NIGHTLY="**Nightly build:** [${NIGHTLY_TAG}](${NIGHTLY_URL})"
+        echo "Nightly release: ${NIGHTLY_URL}"
+    fi
+else
+    echo "No nightly release found for commit $(git rev-parse --short "${RELEASE_COMMIT}")"
+fi
+
+echo "Previous version: ${PREV:-none}"
+echo "${CHANGELOG}"
+
+if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
+    {
+        echo "previous_tag=${PREV}"
+        echo "changelog_title=${CHANGELOG_TITLE}"
+        echo "nightly=${NIGHTLY}"
+        echo "changelog<<CHANGELOG_EOF"
+        echo "${CHANGELOG}"
+        echo "CHANGELOG_EOF"
+    } >> "${GITHUB_OUTPUT}"
+fi
@@ -1 +1 @@
-3834fd814e74e8af277939dabd69ecc780affd21
+8c63e70982c95ceb862e3a1073a2c1beef75d60a
@@ -7,6 +7,12 @@ import subprocess

 HTTPLIB_VERSION = "refs/tags/v0.53.1"

+# used by examples/gguf-hash, these repos have no release tag, so we pin a commit
+XXHASH_COMMIT      = "9f465f1ea932d6ad9a26cd77496311ffa544cd68"
+SHA1_COMMIT        = "e1e2536fcf6a8f9703be8c85d58724b408552287"
+SHA256_COMMIT      = "5e637272c13f200872d55ff579f7e2ab6c3f252f"
+ROTATE_BITS_COMMIT = "27e784942f67db44abf2115c6638e735b579acd1"
+
 vendor = {
    "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
    "https://github.com/nlohmann/json/releases/latest/download/json_fwd.hpp": "vendor/nlohmann/json_fwd.hpp",
@@ -22,12 +28,95 @@ vendor = {
    f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/LICENSE":   "vendor/cpp-httplib/LICENSE",

    "https://raw.githubusercontent.com/sheredom/subprocess.h/9ce0d701b6fb10f8f8c4445edd31e7c60a1237e3/subprocess.h": "vendor/sheredom/subprocess.h",
+
+    f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/xxhash.c":      "vendor/hash/xxhash/xxhash.c",
+    f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/xxhash.h":      "vendor/hash/xxhash/xxhash.h",
+    f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/LICENSE":       "vendor/hash/xxhash/LICENSE",
+
+    # clibs/sha1 ships no license file, the source header says public domain
+    f"https://raw.githubusercontent.com/clibs/sha1/{SHA1_COMMIT}/sha1.c": "vendor/hash/sha1/sha1.c",
+    f"https://raw.githubusercontent.com/clibs/sha1/{SHA1_COMMIT}/sha1.h": "vendor/hash/sha1/sha1.h",
+
+    f"https://raw.githubusercontent.com/jb55/sha256.c/{SHA256_COMMIT}/sha256.c": "vendor/hash/sha256/sha256.c",
+    f"https://raw.githubusercontent.com/jb55/sha256.c/{SHA256_COMMIT}/sha256.h": "vendor/hash/sha256/sha256.h",
+    f"https://raw.githubusercontent.com/jb55/sha256.c/{SHA256_COMMIT}/LICENSE":  "vendor/hash/sha256/LICENSE",
+
+    f"https://raw.githubusercontent.com/jb55/rotate-bits.h/{ROTATE_BITS_COMMIT}/rotate-bits.h": "vendor/hash/rotate-bits/rotate-bits.h",
+    f"https://raw.githubusercontent.com/jb55/rotate-bits.h/{ROTATE_BITS_COMMIT}/LICENSE.md":   "vendor/hash/rotate-bits/LICENSE.md",
+}
+
+# local changes kept on top of the upstream sources
+patches = {
+    "vendor/hash/xxhash/xxhash.h": [(
+        '#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */\n',
+        '/* Windows SDK under 10.0.22000 is missing stdalign.h so we add a check\n'
+        '   before allowing the windows compiler to use the C11 form.\n'
+        '   Reference: https://github.com/Cyan4973/xxHash/issues/955 */\n'
+        '#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) \\\n'
+        '    && (defined(_MSC_VER) && (_MSC_VER >= 1000) || !defined(_MSC_VER)) /* >= C11 */\n'
+    )],
+
+    # sha1 exports a bare "SHA1" symbol, which clashes with the boringssl one at link time.
+    # we compile it as C++ (see vendor/hash/CMakeLists.txt) and put it in a namespace.
+    "vendor/hash/sha1/sha1.h": [
+        (
+            '#if defined(__cplusplus)\n'
+            'extern "C" {\n'
+            '#endif\n',
+
+            'namespace vendor_hash {\n'
+        ),
+        (
+            '#if defined(__cplusplus)\n'
+            '}\n'
+            '#endif\n',
+
+            '} // namespace vendor_hash\n'
+        ),
+    ],
+
+    "vendor/hash/sha1/sha1.c": [
+        (
+            '#include "sha1.h"\n',
+
+            '#include "sha1.h"\n'
+            '\n'
+            'namespace vendor_hash {\n'
+        ),
+        (
+            '    SHA1Final((unsigned char *)hash_out, &ctx);\n'
+            '}\n',
+
+            '    SHA1Final((unsigned char *)hash_out, &ctx);\n'
+            '}\n'
+            '\n'
+            '} // namespace vendor_hash\n'
+        ),
+    ],
+
+    # silence a maybe-uninitialized warning
+    "vendor/hash/sha256/sha256.c": [(
+        "  uint32_t W[16];\n",
+        "  uint32_t W[16] = {0};\n"
+    )],
 }

 for url, filename in vendor.items():
    print(f"downloading {url} to {filename}") # noqa: NP100
    urllib.request.urlretrieve(url, filename)

+for filename, replacements in patches.items():
+    print(f"patching {filename}") # noqa: NP100
+    with open(filename, "r", encoding="utf-8", newline="") as f:
+        content = f.read()
+    for old, new in replacements:
+        if content.count(old) != 1:
+            print(f"Error: cannot apply patch on {filename}, upstream code has changed") # noqa: NP100
+            sys.exit(1)
+        content = content.replace(old, new)
+    with open(filename, "w", encoding="utf-8", newline="") as f:
+        f.write(content)
+
 print("Splitting httplib.h...") # noqa: NP100
 try:
    subprocess.check_call([
@@ -102,6 +102,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
    { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
    { LLM_ARCH_GRANITE_SWITCH,   "graniteswitch"    },
+    { LLM_ARCH_GRANITE_SWA,      "granite_swa"      },
    { LLM_ARCH_CHAMELEON,        "chameleon"        },
    { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
    { LLM_ARCH_PLM,              "plm"              },
@@ -261,6 +262,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
    { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
    { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,       "%s.attention.sliding_window_pattern"       },
+    { LLM_KV_ATTENTION_ROPE_PATTERN,                 "%s.attention.rope_pattern"                 },
+
    { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
    { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
    { LLM_KV_ATTENTION_VALUE_SCALE,                  "%s.attention.value_scale"                  },
@@ -107,6 +107,7 @@ enum llm_arch {
    LLM_ARCH_GRANITE_MOE,
    LLM_ARCH_GRANITE_HYBRID,
    LLM_ARCH_GRANITE_SWITCH,
+    LLM_ARCH_GRANITE_SWA,
    LLM_ARCH_CHAMELEON,
    LLM_ARCH_WAVTOKENIZER_DEC,
    LLM_ARCH_PLM,
@@ -267,6 +268,8 @@ enum llm_kv {
    LLM_KV_ATTENTION_SLIDING_WINDOW,
    LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
    LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_ROPE_PATTERN,
+
    LLM_KV_ATTENTION_OUTPUT_SCALE,
    LLM_KV_ATTENTION_VALUE_SCALE,
    LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
@@ -3099,8 +3099,6 @@ ggml_tensor * llm_graph_context::build_attn(
            int       il) const {
    const bool is_swa = hparams.is_swa(il);

-    GGML_UNUSED(v_cur);
-
    auto * k_rot = is_swa ? inp->self_k_rot_swa : inp->self_k_rot;

    if (k_rot) {
@@ -3133,7 +3131,7 @@ ggml_tensor * llm_graph_context::build_attn(
    // MLA-style attention: the cached K is used as V
    ggml_tensor * q = q_cur;
    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
-    ggml_tensor * v = k;
+    ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);

    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);
@@ -291,7 +291,11 @@ bool llama_hparams::has_rope(uint32_t il) const {
        return false;
    }

-    return true;
+    if (il < n_layer_all) {
+        return rope_pattern[il] != 0;
+    }
+
+    GGML_ABORT("%s: il (%u) out of bounds (n_layer_all: %u)\n", __func__, il, n_layer_all);
 }

 uint32_t llama_hparams::n_layer() const {
--- a/Show More
+++ b/Show More