model: add dots3-note (#27060 )

* text: conversion * init impl * address review comments * fix rope * move to a new llama_kv_cache_dsa_iswa
model: use ggml_rope_set_offset() (#27382 )
2026-08-21 20:22:34 +02:00 · 2026-08-21 19:52:34 +02:00 · 2026-08-21 18:54:29 +02:00 · 2026-08-21 18:56:04 +03:00 · 2026-08-21 15:01:24 +03:00 · 2026-08-21 14:51:26 +03:00
234 changed files with 15612 additions and 14058 deletions
@@ -1,3 +1,4 @@
+# note: place this as the last step of the job, so the new cache is saved by "Post ccache" right after the old one is cleared
 name: "ccache-clear"
 description: "Delete all GitHub Actions caches matching a key prefix"
 inputs:
@@ -27,30 +27,26 @@ jobs:
          cmake --install build --prefix "$PREFIX" --config Release

          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          build_commit=$(git rev-parse --short HEAD | xargs)
+          build_number=$(git rev-list --count HEAD | xargs)

-          set cmakelists [read [open "CMakeLists.txt" r]]
-          regexp {set\(LLAMA_VERSION_MAJOR\s+(\d+)\)} $cmakelists -> major
-          regexp {set\(LLAMA_VERSION_MINOR\s+(\d+)\)} $cmakelists -> minor
-          regexp {set\(LLAMA_VERSION_PATCH\s+(\d+)\)} $cmakelists -> patch
-          set build(version) "$major.$minor.$patch"
+          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          build_version="$major.$minor.$patch"

-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
+                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
+                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")

-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
+          for check in "${checks[@]}"; do
+              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
+                  echo "Checking llama-config.cmake version... \"$check\" failed!"
                  exit 1
-              }
-          }
-          puts "success."
-          EOF
+              fi
+          done
+
+          echo "Checking llama-config.cmake version... success."

          cd examples/simple-cmake-pkg
          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
@@ -97,8 +97,7 @@ jobs:
          cmake -B build \
            -DGGML_NATIVE=OFF \
            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON \
-            -DGGML_NATIVE=OFF
+            -DGGML_RPC=ON
          time cmake --build build --config Release -j $(nproc)

      - name: Test
@@ -119,6 +118,7 @@ jobs:
          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

  windows:
+    name: windows / ${{ matrix.build }}
    runs-on: windows-2025

    env:
@@ -130,13 +130,13 @@ jobs:
        include:
          - build: 'x64-cpu-static'
            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 -DBUILD_SHARED_LIBS=OFF'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
          - 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: 'arm64'
            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON'

    steps:
      - name: Clone
@@ -55,6 +55,20 @@ jobs:
        env:
          GITHUB_REPOSITORY: ${{ github.repository }}

+      - name: Create nightly-tag.txt
+        id: nightly_tag_file
+        run: |
+          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
+          if [[ -z "${NIGHTLY_TAG}" ]]; then
+            echo "Warning: no nightly tag found for the release commit - nightly-tag.txt will not be created"
+            echo "create=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          echo "${NIGHTLY_TAG}" > nightly-tag.txt
+          echo "create=true" >> "$GITHUB_OUTPUT"
+          echo "nightly-tag.txt:"
+          cat nightly-tag.txt
+
      - name: Create release
        if: ${{ github.event.inputs.dry_run == 'false' }}
        uses: ggml-org/action-create-release@v1
@@ -62,26 +76,49 @@ jobs:
          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
+          prerelease: false
+          # TODO: enrich the body of the release with more information
          body: |
-            > [!NOTE]
-            > Semantic versioning is still work in progress.
-            > More info can be found in https://github.com/ggml-org/ggml/discussions/1579
+            ## Overview
+
+            New version has been released.

            ${{ steps.desc.outputs.nightly }}

+            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+
+            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+
            ## ${{ steps.desc.outputs.changelog_title }}

            ${{ steps.desc.outputs.changelog }}

+      - name: Upload nightly-tag.txt
+        if: ${{ github.event.inputs.dry_run == 'false' && steps.nightly_tag_file.outputs.create == 'true' }}
+        uses: actions/github-script@v8
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            console.log('uploadReleaseAsset', 'nightly-tag.txt');
+            await github.rest.repos.uploadReleaseAsset({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              release_id: release_id,
+              name: 'nightly-tag.txt',
+              data: await fs.readFileSync('./nightly-tag.txt')
+            });
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
          if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
            echo "Dry run complete - all checks passed."
            echo "Would have created tag: ${{ steps.checks.outputs.version }}"
+            if [[ -n "${{ steps.desc.outputs.nightly_tag }}" ]]; then
+              echo "Would have uploaded nightly-tag.txt: ${{ steps.desc.outputs.nightly_tag }}"
+            fi
          else
            echo "::error::Dry run found release check failures. A release tag would not be created."
            exit 1
@@ -145,11 +145,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-${{ matrix.arch }}
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -166,6 +161,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-macos-${{ matrix.build }}.tar.gz
          name: llama-bin-macos-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-${{ matrix.arch }}
+
  ubuntu-cpu:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -231,12 +231,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        if: ${{ matrix.build != 's390x' }}
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-cpu
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -253,6 +247,12 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        if: ${{ matrix.build != 's390x' }}
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-cpu
+
  ubuntu-vulkan:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -318,11 +318,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-vulkan
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -339,6 +334,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-vulkan
+
  android-arm64:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -512,11 +512,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build/ReleaseOV --config Release --parallel

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -551,6 +546,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
          name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-openvino-release-no-preset-v1
+
  windows-openvino:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -637,11 +637,6 @@ jobs:

          cmake --build build\ReleaseOV --config Release -- /m

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-openvino
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -680,7 +675,13 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
          name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-openvino
+
  windows-cpu:
+    name: windows-cpu / ${{ matrix.arch }}
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

@@ -728,18 +729,13 @@ jobs:
            -DGGML_BACKEND_DL=ON ^
            -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
            -DGGML_OPENMP=ON ^
+            -DGGML_OPENMP_FETCH=ON ^
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-
      - name: Pack artifacts
        id: pack_artifacts
        run: |
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
          7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*

      - name: Upload artifacts
@@ -748,6 +744,11 @@ jobs:
          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+
  windows-rocm:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -840,11 +841,6 @@ jobs:
            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-
      - name: Verify HIP backend was built
        run: |
          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
@@ -877,6 +873,11 @@ jobs:
          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+
  windows:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1042,11 +1043,6 @@ jobs:
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
-
      - name: Pack artifacts
        id: pack_artifacts
        run: |
@@ -1082,6 +1078,11 @@ jobs:
          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+
  windows-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1141,11 +1142,6 @@ jobs:
            -DLLAMA_BUILD_BORINGSSL=ON
          cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-sycl
-
      - name: Build the release package
        id: pack_artifacts
        run: |
@@ -1192,6 +1188,11 @@ jobs:
          path: llama-bin-win-sycl-x64.zip
          name: llama-bin-win-sycl-x64.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
  ubuntu-24-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1264,11 +1265,6 @@ jobs:
            -DGGML_SYCL_F16=${{ matrix.fp16 }}
          time cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -1285,6 +1281,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
 #   ubuntu-22-rocm:
 #     needs: [check-release, get-version]
 #     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1379,11 +1380,6 @@ jobs:
 #             ${{ env.CMAKE_ARGS }}
 #           cmake --build build --config Release -j $(nproc)

-#       # - name: ccache-clear
-#       #   uses: ./.github/actions/ccache-clear
-#       #   with:
-#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
-
 #       - name: Determine tag name
 #         id: tag
 #         uses: ./.github/actions/get-tag-name
@@ -1403,6 +1399,11 @@ jobs:
 #           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
 #           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz

+#       # - name: ccache-clear
+#       #   uses: ./.github/actions/ccache-clear
+#       #   with:
+#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+
  ios-xcode:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1687,6 +1688,7 @@ jobs:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        with:
          tag_name: ${{ steps.tag.outputs.name }}
+          prerelease: true
          body: |
            <details open>

@@ -9,6 +9,7 @@ General:

 Coding:
 - When in doubt, always refer to the CONTRIBUTING.md file of the project
+- In `test-backend-ops.cpp`, do not mention specific backends (e.g. Metal, CUDA) in comments
 - When referencing issues or PRs in comments, use the format:
  - C/C++ code: `// ref: <url>`
  - Other (CMake, etc.): `# ref: <url>`
@@ -4,8 +4,8 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 2)
+set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

 # whether this is a development/nightly build
@@ -8,6 +8,7 @@ set( CMAKE_CXX_COMPILER  clang++ )

 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
+set( CMAKE_ASM_COMPILER_TARGET ${target} )

 set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
@@ -2595,6 +2595,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.mmproj_use_gpu = value;
        }
    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
+    add_opt(common_arg(
+        // note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet
+        {"-mmdev", "--mmproj-device"}, "DEVICE",
+        "device to use for multimodal projector (none = don't offload, default: auto)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            if (value == "none") {
+                params.mmproj_use_gpu = false;
+                params.mmproj_device  = nullptr;
+                return;
+            }
+            auto devices = parse_device_list(value);
+            // parse_device_list pushes nullptr at back so devices is length 2 for single device.
+            if (devices.size() > 2) {
+                throw std::invalid_argument("only one device may be specified for mmproj");
+            }
+            params.mmproj_use_gpu = true;
+            params.mmproj_device  = devices.front();
+        }
+    ).set_examples(mmproj_examples).set_env("MTMD_BACKEND_DEVICE")); // no LLAMA_ARG_ prefix for backward compatibility reason
    add_opt(common_arg(
        {"--image", "--audio", "--video"}, "FILE",
        "path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",
@@ -581,9 +581,10 @@ struct common_params {

    // multimodal models (see tools/mtmd)
    struct common_params_model mmproj;
-    bool mmproj_use_gpu = true;     // use GPU for multimodal model
-    bool no_mmproj = false;         // explicitly disable multimodal model
-    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
+    bool mmproj_use_gpu = true;                 // use GPU for multimodal model
+    ggml_backend_dev_t mmproj_device = nullptr; // GPU device to use for multimodal model
+    bool no_mmproj = false;                     // explicitly disable multimodal model
+    std::vector<std::string> image;             // path to image file(s) ; TODO: change the name to "media"
    int image_min_tokens = -1;
    int image_max_tokens = -1;
    int mtmd_batch_max_tokens = 1024;
@@ -278,7 +278,9 @@ static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };

-static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+static const int MAX_PATTERN_DEPTH = 100;
+
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?', '^', '$'};
 static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};

 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
@@ -309,6 +311,32 @@ static std::string format_literal(const std::string & literal) {

 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }

+static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
+    if (pos + 1 >= pattern.length() || pattern[pos] != '\\') {
+        return 0;
+    }
+    size_t n_hex = 0;
+    switch (pattern[pos + 1]) {
+        case 'x': n_hex = 2; break;
+        case 'u': n_hex = 4; break;
+        case 'U': n_hex = 8; break;
+        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']':
+            return 2;
+        default:
+            return 0;
+    }
+    if (pos + 2 + n_hex > pattern.length()) {
+        return 0;
+    }
+    for (size_t i = pos + 2; i < pos + 2 + n_hex; i++) {
+        char h = pattern[i];
+        if (!((h >= '0' && h <= '9') || (h >= 'a' && h <= 'f') || (h >= 'A' && h <= 'F'))) {
+            return 0;
+        }
+    }
+    return 2 + n_hex;
+}
+
 class common_schema_converter {
 private:
    friend class common_schema_info;
@@ -345,16 +373,42 @@ private:
        return string_join(rules, " | ");
    }

+    // thrown when the pattern is a valid regex with no grammar equivalent
+    struct unsupported_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
+    // thrown when the pattern is not a valid regex
+    struct invalid_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
-        if (!(pattern.front() == '^' && pattern.back() == '$')) {
-            _errors.push_back("Pattern must start with '^' and end with '$'");
+        auto rules_snapshot = _rules;
+        try {
+            return _pattern_to_rule(pattern, name);
+        } catch (const unsupported_pattern & err) {
+            // revert rules
+            _rules = std::move(rules_snapshot);
+            _warnings.push_back("pattern " + pattern + " is not supported (" + err.what() + "), accepting any string");
+            return _add_rule(name, _add_primitive("string", PRIMITIVE_RULES.at("string")));
+        } catch (const invalid_pattern & err) {
+            _rules = std::move(rules_snapshot);
+            _errors.push_back("Invalid pattern " + pattern + ": " + err.what());
            return "";
        }
+    }
+
+    std::string _pattern_to_rule(const std::string & pattern, const std::string & name) {
+        if (pattern.length() < 2 || pattern.front() != '^' || pattern.back() != '$') {
+            throw unsupported_pattern("not anchored with '^' and '$'");
+        }
        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
        std::unordered_map<std::string, std::string> sub_rule_ids;

        size_t i = 0;
        size_t length = sub_pattern.length();
+        int paren_depth = 0;

        using literal_or_rule = std::pair<std::string, bool>;
        auto to_rule = [&](const literal_or_rule & ls) {
@@ -363,7 +417,6 @@ private:
            return is_literal ? "\"" + s + "\"" : s;
        };
        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
-            size_t start = i;
            std::vector<literal_or_rule> seq;

            auto get_dot = [&]() {
@@ -420,43 +473,42 @@ private:
                        if (i + 1 < length && sub_pattern[i + 1] == ':') {
                            i += 2; // skip "?:" for non-capturing group, treat as regular group
                        } else {
-                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
-                            _warnings.push_back("Unsupported pattern syntax");
-                            // skip to matching ')' to avoid UB on empty seq
-                            int depth = 1;
-                            while (i < length && depth > 0) {
-                                if (sub_pattern[i] == '\\' && i + 1 < length) {
-                                    i += 2; // skip escaped character
-                                } else {
-                                    if (sub_pattern[i] == '(') depth++;
-                                    else if (sub_pattern[i] == ')') depth--;
-                                    i++;
-                                }
-                            }
-                            continue;
+                            // lookaround, named group, inline flags, ...
+                            throw unsupported_pattern("unsupported group syntax");
                        }
                    }
+                    paren_depth++;
+                    if (paren_depth > MAX_PATTERN_DEPTH) {
+                        throw unsupported_pattern("pattern nesting too deep");
+                    }
                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
                } else if (c == ')') {
                    i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
-                        _errors.push_back("Unbalanced parentheses");
+                    if (paren_depth == 0) {
+                        throw invalid_pattern("unbalanced parentheses");
                    }
+                    paren_depth--;
                    return join_seq();
+                } else if (c == '^' || c == '$') {
+                    throw unsupported_pattern("anchor inside the pattern");
                } else if (c == '[') {
                    std::string square_brackets = std::string(1, c);
                    i++;
                    while (i < length && sub_pattern[i] != ']') {
                        if (sub_pattern[i] == '\\') {
-                            square_brackets += sub_pattern.substr(i, 2);
-                            i += 2;
+                            auto escape_length = gbnf_escape_length(sub_pattern, i);
+                            if (escape_length == 0) {
+                                throw unsupported_pattern("unsupported escape in character class: " + sub_pattern.substr(i, 2));
+                            }
+                            square_brackets += sub_pattern.substr(i, escape_length);
+                            i += escape_length;
                        } else {
                            square_brackets += sub_pattern[i];
                            i++;
                        }
                    }
                    if (i >= length) {
-                        _errors.push_back("Unbalanced square brackets");
+                        throw invalid_pattern("unterminated character class");
                    }
                    square_brackets += ']';
                    i++;
@@ -465,6 +517,9 @@ private:
                    seq.emplace_back("|", false);
                    i++;
                } else if (c == '*' || c == '+' || c == '?') {
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
+                    }
                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
                    i++;
                } else if (c == '{') {
@@ -475,18 +530,19 @@ private:
                        i++;
                    }
                    if (i >= length) {
-                        _errors.push_back("Unbalanced curly brackets");
+                        throw unsupported_pattern("unterminated curly brackets");
                    }
                    curly_brackets += '}';
                    i++;
                    auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                    int min_times = 0;
                    int max_times = std::numeric_limits<int>::max();
+                    if (nums.size() != 1 && nums.size() != 2) {
+                        throw unsupported_pattern("wrong number of values in curly brackets");
+                    }
                    try {
                        if (nums.size() == 1) {
                            min_times = max_times = std::stoi(nums[0]);
-                        } else if (nums.size() != 2) {
-                            _errors.push_back("Wrong number of values in curly brackets");
                        } else {
                            if (!nums[0].empty()) {
                                min_times = std::stoi(nums[0]);
@@ -495,9 +551,11 @@ private:
                                max_times = std::stoi(nums[1]);
                            }
                        }
-                    } catch (const std::invalid_argument & e) {
-                        _errors.push_back("Invalid number in curly brackets");
-                        return std::make_pair("", false);
+                    } catch (const std::logic_error &) {
+                        throw unsupported_pattern("invalid number in curly brackets");
+                    }
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
                    }
                    auto &last = seq.back();
                    auto &sub = last.first;
@@ -523,15 +581,22 @@ private:
                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
                    };
                    while (i < length) {
-                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                        if (sub_pattern[i] == '\\') {
+                            if (i == length - 1) {
+                                throw invalid_pattern("trailing backslash");
+                            }
                            char next = sub_pattern[i + 1];
                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
                                i++;
                                literal += sub_pattern[i];
                                i++;
                            } else {
-                                literal += sub_pattern.substr(i, 2);
-                                i += 2;
+                                auto escape_length = gbnf_escape_length(sub_pattern, i);
+                                if (escape_length == 0) {
+                                    throw unsupported_pattern("unsupported escape: " + sub_pattern.substr(i, 2));
+                                }
+                                literal += sub_pattern.substr(i, escape_length);
+                                i += escape_length;
                            }
                        } else if (sub_pattern[i] == '"') {
                            literal += "\\\"";
@@ -544,14 +609,21 @@ private:
                            break;
                        }
                    }
-                    if (!literal.empty()) {
-                        seq.emplace_back(literal, true);
+                    if (literal.empty()) { // nothing was consumed, ex. a stray ']' or '}'
+                        throw unsupported_pattern(std::string("unsupported character: ") + c);
                    }
+                    seq.emplace_back(literal, true);
                }
            }
            return join_seq();
        };
-        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"");
+
+        auto rule = to_rule(transform());
+        if (paren_depth != 0) {
+            throw invalid_pattern("unbalanced parentheses");
+        }
+
+        return _add_rule(name, "\"\\\"\" (" + rule + ") \"\\\"\"");
    }

    /*
@@ -2649,6 +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
@@ -57,12 +57,16 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3DSparkModel": "qwen",
    "DSparkDraftModel": "qwen",
    "DSparkSpeculator": "qwen",
+    "Lfm2DSparkDraftModel": "qwen",
    "DeepseekV4ForCausalLM": "deepseek",
    "DeepseekV4DSparkModel": "deepseek",
    "DistilBertForMaskedLM": "bert",
    "DistilBertForSequenceClassification": "bert",
    "DistilBertModel": "bert",
    "Dots1ForCausalLM": "dots1",
+    "Dots3NoteForCausalLM": "dots3",
+    "Dots3NoteForConditionalGeneration": "dots3",
+    "Dots3NoteTextForCausalLM": "dots3",
    "DotsOCRForCausalLM": "qwen",
    "DreamModel": "dream",
    "Ernie4_5ForCausalLM": "ernie",
@@ -0,0 +1,195 @@
+from __future__ import annotations
+
+import math
+import re
+
+from typing import TYPE_CHECKING, Callable, Iterable
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, gguf
+
+from .deepseek import DeepseekV2Model
+
+
+@ModelBase.register("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration", "Dots3NoteTextForCausalLM")
+class Dots3NoteModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DOTS3NOTE
+    skip_mtp = False
+    supports_mtp_export = True
+
+    # trunk layer count, stashed before indexing for filter_tensors (mirrors DeepseekV32Model)
+    _n_main_layers: int | None = None
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        hparams = self.hparams
+
+        # config file doesn't specify MTP block, detect it from model weight
+        self.n_nextn = 1 if "model.mtp.embed_tokens.weight" in self.model_tensors else 0
+        if self.n_nextn:
+            self.block_count += self.n_nextn
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self.layer_types = hparams["layer_types"]
+        if len(self.layer_types) < hparams["num_hidden_layers"]:
+            raise ValueError("layer_types is shorter than num_hidden_layers")
+
+        if hparams.get("use_dsa", True) is not True:
+            raise ValueError("dots3-note conversion requires use_dsa=true")
+        if hparams.get("normalization", "RMSNorm") != "RMSNorm" or hparams.get("final_norm", "RMSNorm") != "RMSNorm":
+            raise ValueError("dots3-note conversion only supports RMSNorm")
+        if hparams.get("k_rope_only_layernorm", True) is not True:
+            raise ValueError("dots3-note conversion requires k_rope_only_layernorm=true")
+        if hparams.get("topk_method", "noaux_tc") != "noaux_tc" or hparams.get("scoring_func") != "sigmoid":
+            raise ValueError("dots3-note conversion only supports noaux_tc/sigmoid expert gating")
+        if hparams.get("n_group", 1) != 1 or hparams.get("topk_group", 1) != 1:
+            raise ValueError("dots3-note conversion does not support grouped expert routing")
+        if hparams.get("use_dynamic_rsf", False) or hparams.get("moe_gating_fp32", False):
+            raise ValueError("dots3-note conversion does not support use_dynamic_rsf/moe_gating_fp32")
+        for key in ("attention_gate_type", "swa_attention_gate_type"):
+            if hparams.get(key, "headwise") != "headwise":
+                raise ValueError(f"dots3-note conversion only supports headwise attention gate, got {key}={hparams.get(key)!r}")
+        if hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"] != hparams.get("swa_head_dim", 256):
+            raise ValueError("swa_head_dim must equal swa_qk_nope_head_dim + swa_qk_rope_head_dim")
+        if hparams["swa_qk_rope_head_dim"] != hparams["qk_rope_head_dim"]:
+            # both layer kinds share a single rope_dimension_count
+            raise ValueError("swa_qk_rope_head_dim must match qk_rope_head_dim")
+
+        self.apply_lora_rescale = hparams.get("apply_mla_qkv_lora_rescale", False)
+
+    def _is_swa_layer(self, bid: int) -> bool:
+        if bid >= self.hparams["num_hidden_layers"]:
+            # note: the NextN/MTP block uses the sliding-attention MLA
+            return True
+        return self.layer_types[bid] == "sliding_attention"
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith(("vision_encoder.", "audio_encoder.")):
+            return None
+
+        assert cls._n_main_layers is not None
+        is_mtp = name.startswith("model.mtp.") or \
+            ((m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers)
+
+        # --no-mtp: drop the NextN/MTP block; --mtp: keep only that block plus the shared embeddings/norm/lm_head
+        if is_mtp and cls.no_mtp:
+            return None
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        # head_count is a per-layer array because the two layer kinds have different head counts
+        n_layer = hparams["num_hidden_layers"]
+        hparams["num_attention_heads"] = [
+            hparams["swa_num_attention_heads"] if self._is_swa_layer(il) else hparams["num_attention_heads"]
+            for il in range(self.block_count)
+        ]
+
+        # prevent the base class from emitting key/value_length from the unused head_dim
+        hparams.pop("head_dim", None)
+
+        super().set_gguf_parameters()
+
+        # MLA geometry of the sliding-window layers (rope.freq_base_swa is emitted by the base class)
+        swa_kv_lora_rank = hparams["swa_kv_lora_rank"]
+        self.gguf_writer.add_sliding_window(hparams["sliding_window_size"])
+        self.gguf_writer.add_sliding_window_pattern([self._is_swa_layer(il) for il in range(n_layer)])
+        self.gguf_writer.add_kv_lora_rank_swa(swa_kv_lora_rank)
+        self.gguf_writer.add_key_length_swa(swa_kv_lora_rank + hparams["swa_qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_swa(swa_kv_lora_rank)
+        self.gguf_writer.add_key_length_mla_swa(hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_mla_swa(hparams["swa_v_head_dim"])
+        if hparams["swa_q_lora_rank"] != hparams["q_lora_rank"]:
+            raise ValueError("dots3-note conversion assumes a shared q_lora_rank for both layer kinds")
+
+        if self.n_nextn:
+            self.gguf_writer.add_nextn_predict_layers(self.n_nextn)
+
+        # DSA indexer (full-attention layers only)
+        self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
+        self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
+        self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
+        self.gguf_writer.add_indexer_types([not self._is_swa_layer(il) for il in range(n_layer)])
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # move the MTP token embedding into the NextN block so the standard nextn mapping picks it up
+        if name == "model.mtp.embed_tokens.weight":
+            name = f"model.layers.{self.hparams['num_hidden_layers']}.embed_tokens.weight"
+            bid = self.hparams["num_hidden_layers"]
+
+        # fold the activation rescale sqrt(n_embd/lora_rank) into the preceding RMSNorm weight
+        # this also covers the indexer wq_b, which reads the same rescaled q_lora activation
+        if self.apply_lora_rescale and bid is not None:
+            if name.endswith("q_a_layernorm.weight"):
+                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / self.hparams["q_lora_rank"])
+            elif name.endswith("kv_a_layernorm.weight"):
+                rank = self.hparams["swa_kv_lora_rank"] if self._is_swa_layer(bid) else self.hparams["kv_lora_rank"]
+                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / rank)
+
+        # MLA absorption: split kv_b_proj into k_b (transposed) and v_b, per-layer-kind geometry
+        if name.endswith("kv_b_proj.weight"):
+            assert bid is not None
+            if self._is_swa_layer(bid):
+                n_head = self.hparams["swa_num_attention_heads"]
+                qk_nope_head_dim = self.hparams["swa_qk_nope_head_dim"]
+                v_head_dim = self.hparams["swa_v_head_dim"]
+            else:
+                n_head = self.hparams["num_attention_heads"]
+                qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+                v_head_dim = self.hparams["v_head_dim"]
+            if isinstance(n_head, list):  # set_gguf_parameters turns this into a per-layer array
+                n_head = n_head[bid]
+
+            assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)
+
+            kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])
+            k_b, v_b = kv_b.split([qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            yield from ModelBase.modify_tensors(self, k_b, name.replace("kv_b_proj", "k_b_proj"), bid)
+            yield from ModelBase.modify_tensors(self, v_b, name.replace("kv_b_proj", "v_b_proj"), bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -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"))
@@ -709,7 +709,7 @@ class DFlashModel(Qwen3Model):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator")
+@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator", "Lfm2DSparkDraftModel")
@ModelBase.example("satgeze/Qwen3.6-27B-DSpark")
 class DSparkModel(DFlashModel):
    # DSpark = DFlash + a semi-autoregressive Markov head.
@@ -759,6 +759,13 @@ class DSparkModel(DFlashModel):
            return None
        return super().filter_tensors(item)

+    _ROPE_PERMUTE_SUFFIXES = (
+        "self_attn.q_proj.weight",
+        "self_attn.k_proj.weight",
+        "self_attn.q_norm.weight",
+        "self_attn.k_norm.weight",
+    )
+
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if name == "model.d2t":
            self._d2t = data_torch
@@ -767,6 +774,12 @@ class DSparkModel(DFlashModel):
        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith(("embed_tokens.weight", "lm_head.weight")):
            return

+        # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd
+        if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES):
+            head_dim = self.hparams["head_dim"]
+            shape = data_torch.shape
+            data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape)
+
        yield from super().modify_tensors(data_torch, name, bid)

    def prepare_tensors(self):
@@ -116,7 +116,7 @@ in inline assembler.
 Most kernels are very naive with lots of low hanging fruits left:

 > [!IMPORTANT]
-> Several assembly instructions emmited by the compiler are not implemented
+> Several assembly instructions emitted by the compiler are not implemented
 > in hardware and software emulation in firmware is not ready yet.
 > Eventually firmware will transparently trap unimplemented instructions
 > and will emulate them inside exception handler. Until then, kernel
@@ -138,12 +138,12 @@ Most kernels are very naive with lots of low hanging fruits left:
 > kernel build process. Feel free to take ideas/code from there or try linking
 > it in.

-Before commiting any changes to operations and/or kernels, don't forget
+Before committing any changes to operations and/or kernels, don't forget
 to update supported ops reports (instructions at `docs/ops.md`).

 When logging is enabled (e.g. by setting `--log-file` cli param),
 each compute kernel run outputs a line with
-pipe-delimited key-value pairs containing kernel level performance infomation.
+pipe-delimited key-value pairs containing kernel level performance information.
 Line is prefixed with `ET_PERF`:

 ```
@@ -160,7 +160,7 @@ to `GGML_ET_PROFILE/et_runtime_trace.json` and `GGML_ET_PROFILE/kernel_map` on e

 ### Uberkernel

-The in-knernel implementaiton of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
+The in-kernel implementation of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
 dispatches multiple already existing kernel implementations with device side synchronization. Due to the processor's design, there is no natural memory visibility
 horizon between sub-kernel invocations. This makes uberkernel much more difficult to develop and debug. Currently Uberkerel is hidden begind the
 `GGML_ET_UBERKERNEL` environment variable and is disabled by default. Setting it to 1 enables it and provides significant performance improvements but is only
@@ -72,9 +72,10 @@ cmake --build build --config Release
    - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
    - For Windows on ARM (arm64, WoA) build with:
    ```bash
-    cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
+    cmake --preset arm64-windows-llvm-release -D GGML_OPENMP_FETCH=ON
    cmake --build build-arm64-windows-llvm-release
    ```
+    `GGML_OPENMP_FETCH` downloads the official LLVM OpenMP runtime and requires Clang, 7-Zip and network access during configuration. CMake selects the runtime from the target architecture, so this also works when cross-compiling for WoA from x64. The extracted header, import library, DLL and OpenMP license are placed under `build/_deps`. The build copies `libomp.dll` and `LICENSE-LLVM-OpenMP` to the runtime output directory and installs them together. Omit the option to use CMake's normal OpenMP detection, or pass `-D GGML_OPENMP=OFF` to disable OpenMP.
    For building with ninja generator and clang compiler as default:
      -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
      ```bash
@@ -166,6 +166,19 @@ Examples:
 - Some models require scaling the input position. For example, `[0, 1, 2, ...]` becomes `[0, 0.5, 1, ...]`. In this case, you can provide the scaling via `freq_scale = 0.5f`.
 - Some models use learned RoPE frequencies instead of relying on `powf(freq_base, -2.0 * i / n_dims)`. In this case, you can provide the learned frequencies via the `rope_freqs` tensor (corresponding to the `c` argument in `ggml_rope_ext`), then set `freq_base = 1.0f`. An important note is that `rope_freqs` in GGML is the **inverse** (`theta = pos[i] / rope_freqs`), so you may need to invert `rope_freqs` during conversion.

+### Rotating only a part of the head
+
+Many models rotate only a part of each head and leave the rest untouched (often called the "nope" part). Do not build this with views plus `ggml_concat`, it's not efficient. Both layouts can be done with a single RoPE op:
+
+- `[rope|nope]`, rotated dims first: pass `n_dims` smaller than the head size to `ggml_rope_ext`. Dims from `n_dims` to the end are copied as-is.
+- `[nope|rope]`, rotated dims last: call `ggml_rope_set_offset(cur, n_offs)` on the result of the RoPE, where `n_offs` is the size of the leading untouched part. Dims outside `[n_offs, n_offs + n_dims)` are copied as-is.
+
+`n_offs` must be even, `n_offs + n_dims` must fit in the row, and vision RoPE is not supported. Note that the frequencies are computed relative to the rotated window.
+
+Example: DeepSeek-V4 uses `[nope|rope]` for its query, key and compressed KV tensors, so `src/models/deepseek4.cpp` ropes the whole tensor and then calls `ggml_rope_set_offset(cur, n_embd_head_nope)`.
+
+Exception: some models apply an extra op to the `nope` part, for example `deepseek32.cpp`, and may not use this optimization. While RoPE can be applied selectively to a part of the head, the extra op may not, so these models still need views plus `ggml_concat`.
+
 ## GGUF specification

 https://github.com/ggml-org/ggml/blob/master/docs/gguf.md
@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 2)
+set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -243,6 +243,7 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                            "ggml: metal minimum macOS version")
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
+option(GGML_OPENMP_FETCH                    "ggml: fetch LLVM OpenMP"                         OFF)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
@@ -222,9 +222,123 @@ if (GGML_SCHED_NO_REALLOC)
    target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
 endif()

-if (GGML_OPENMP)
+if (GGML_OPENMP_FETCH)
+    if (NOT GGML_OPENMP)
+        message(FATAL_ERROR "GGML_OPENMP_FETCH requires GGML_OPENMP")
+    elseif (NOT WIN32 OR NOT (CMAKE_C_COMPILER_ID MATCHES "Clang"))
+        message(FATAL_ERROR "GGML_OPENMP_FETCH currently requires Clang on Windows")
+    endif()
+
+    set(GGML_OPENMP_LLVM_VERSION "20.1.8")
+    string(REGEX MATCH "^[0-9]+" GGML_OPENMP_LLVM_VERSION_MAJOR "${GGML_OPENMP_LLVM_VERSION}")
+    string(REGEX MATCH "^[0-9]+" GGML_OPENMP_COMPILER_VERSION_MAJOR "${CMAKE_C_COMPILER_VERSION}")
+    if (NOT GGML_OPENMP_COMPILER_VERSION_MAJOR STREQUAL GGML_OPENMP_LLVM_VERSION_MAJOR)
+        message(FATAL_ERROR "LLVM OpenMP ${GGML_OPENMP_LLVM_VERSION} requires Clang ${GGML_OPENMP_LLVM_VERSION_MAJOR}.x")
+    endif()
+
+    string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" GGML_OPENMP_SYSTEM_PROCESSOR)
+    if (GGML_OPENMP_SYSTEM_PROCESSOR MATCHES "^(amd64|x86_64)$")
+        set(GGML_OPENMP_ARCH "x64")
+        set(GGML_OPENMP_INSTALLER_SUFFIX "win64")
+        set(GGML_OPENMP_INSTALLER_SHA256 "3197846a2b19063687dd56e93e34cd941e3548d907f23a6131571321bdf9fe7b")
+    elseif (GGML_OPENMP_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64)$")
+        set(GGML_OPENMP_ARCH "arm64")
+        set(GGML_OPENMP_INSTALLER_SUFFIX "woa64")
+        set(GGML_OPENMP_INSTALLER_SHA256 "7c4ac97eb2ae6b960ca5f9caf3ff6124c8d2a18cc07a7840a4d2ea15537bad8e")
+    else()
+        message(FATAL_ERROR "GGML_OPENMP_FETCH does not support ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    set(GGML_OPENMP_CACHE_DIR "${CMAKE_BINARY_DIR}/_deps")
+    set(GGML_OPENMP_ROOT "${GGML_OPENMP_CACHE_DIR}/llvm-openmp-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_ARCH}")
+    set(GGML_OPENMP_LIBRARY "${GGML_OPENMP_ROOT}/lib/libomp.lib")
+    set(GGML_OPENMP_RUNTIME "${GGML_OPENMP_ROOT}/bin/libomp.dll")
+    set(GGML_OPENMP_HEADER "${GGML_OPENMP_ROOT}/include/omp.h")
+    set(GGML_OPENMP_LICENSE "${GGML_OPENMP_ROOT}/LICENSE.TXT")
+    set(GGML_OPENMP_LICENSE_SHA256 "fdad1758a9e1f9d5a81e18879b3406772115edc92c24bfa36b70c654f325e8e4")
+
+    if (NOT EXISTS "${GGML_OPENMP_LIBRARY}" OR NOT EXISTS "${GGML_OPENMP_RUNTIME}" OR NOT EXISTS "${GGML_OPENMP_HEADER}")
+        find_program(GGML_OPENMP_7Z NAMES 7z 7zz 7za)
+        if (NOT GGML_OPENMP_7Z)
+            message(FATAL_ERROR "GGML_OPENMP_FETCH requires 7-Zip to extract the LLVM installer")
+        endif()
+
+        set(GGML_OPENMP_INSTALLER "${GGML_OPENMP_ROOT}/LLVM-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_INSTALLER_SUFFIX}.exe")
+        set(GGML_OPENMP_EXTRACT_DIR "${GGML_OPENMP_ROOT}/extract")
+        set(GGML_OPENMP_INSTALLER_URL "https://github.com/llvm/llvm-project/releases/download/llvmorg-${GGML_OPENMP_LLVM_VERSION}/LLVM-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_INSTALLER_SUFFIX}.exe")
+
+        file(MAKE_DIRECTORY "${GGML_OPENMP_EXTRACT_DIR}")
+        file(DOWNLOAD "${GGML_OPENMP_INSTALLER_URL}" "${GGML_OPENMP_INSTALLER}"
+            EXPECTED_HASH "SHA256=${GGML_OPENMP_INSTALLER_SHA256}"
+            SHOW_PROGRESS
+            STATUS GGML_OPENMP_DOWNLOAD_STATUS)
+        list(GET GGML_OPENMP_DOWNLOAD_STATUS 0 GGML_OPENMP_DOWNLOAD_RESULT)
+        if (NOT GGML_OPENMP_DOWNLOAD_RESULT EQUAL 0)
+            list(GET GGML_OPENMP_DOWNLOAD_STATUS 1 GGML_OPENMP_DOWNLOAD_ERROR)
+            message(FATAL_ERROR "Failed to download LLVM OpenMP: ${GGML_OPENMP_DOWNLOAD_ERROR}")
+        endif()
+
+        execute_process(
+            COMMAND "${GGML_OPENMP_7Z}" e -y "-o${GGML_OPENMP_EXTRACT_DIR}" "${GGML_OPENMP_INSTALLER}" -r libomp.lib libomp.dll omp.h
+            RESULT_VARIABLE GGML_OPENMP_EXTRACT_RESULT
+            OUTPUT_QUIET)
+        if (NOT GGML_OPENMP_EXTRACT_RESULT EQUAL 0 OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/libomp.lib" OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/libomp.dll" OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/omp.h")
+            message(FATAL_ERROR "Failed to extract libomp from ${GGML_OPENMP_INSTALLER}")
+        endif()
+
+        file(MAKE_DIRECTORY "${GGML_OPENMP_ROOT}/lib" "${GGML_OPENMP_ROOT}/bin" "${GGML_OPENMP_ROOT}/include")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/libomp.lib" DESTINATION "${GGML_OPENMP_ROOT}/lib")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/libomp.dll" DESTINATION "${GGML_OPENMP_ROOT}/bin")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/omp.h" DESTINATION "${GGML_OPENMP_ROOT}/include")
+        file(REMOVE_RECURSE "${GGML_OPENMP_INSTALLER}" "${GGML_OPENMP_EXTRACT_DIR}")
+    endif()
+
+    # The NSIS installer embeds LLVM's general license in its UI but does not install it as a file; use OpenMP's license to include its additional notices.
+    if (EXISTS "${GGML_OPENMP_LICENSE}")
+        file(SHA256 "${GGML_OPENMP_LICENSE}" GGML_OPENMP_LICENSE_ACTUAL_SHA256)
+    endif()
+    if (NOT GGML_OPENMP_LICENSE_ACTUAL_SHA256 STREQUAL GGML_OPENMP_LICENSE_SHA256)
+        file(DOWNLOAD "https://raw.githubusercontent.com/llvm/llvm-project/llvmorg-${GGML_OPENMP_LLVM_VERSION}/openmp/LICENSE.TXT" "${GGML_OPENMP_LICENSE}"
+            EXPECTED_HASH "SHA256=${GGML_OPENMP_LICENSE_SHA256}")
+    endif()
+
+    if (COMMAND license_add_file)
+        license_add_file("LLVM OpenMP" "${GGML_OPENMP_LICENSE}")
+    endif()
+
+    add_library(ggml-openmp-c INTERFACE)
+    target_compile_options(ggml-openmp-c INTERFACE "$<$<COMPILE_LANGUAGE:C>:-fopenmp=libomp>")
+    target_include_directories(ggml-openmp-c SYSTEM INTERFACE "${GGML_OPENMP_ROOT}/include")
+    target_link_libraries(ggml-openmp-c INTERFACE "${GGML_OPENMP_LIBRARY}")
+
+    add_library(ggml-openmp-cxx INTERFACE)
+    target_compile_options(ggml-openmp-cxx INTERFACE "$<$<COMPILE_LANGUAGE:CXX>:-fopenmp=libomp>")
+    target_include_directories(ggml-openmp-cxx SYSTEM INTERFACE "${GGML_OPENMP_ROOT}/include")
+    target_link_libraries(ggml-openmp-cxx INTERFACE "${GGML_OPENMP_LIBRARY}")
+
+    set(GGML_OPENMP_RUNTIME_OUTPUT_DIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
+    if (CMAKE_CONFIGURATION_TYPES)
+        string(APPEND GGML_OPENMP_RUNTIME_OUTPUT_DIR "/$<CONFIG>")
+    endif()
+    add_custom_target(ggml-openmp-runtime ALL
+        COMMAND ${CMAKE_COMMAND} -E make_directory "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}"
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different "${GGML_OPENMP_RUNTIME}" "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}/libomp.dll"
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different "${GGML_OPENMP_LICENSE}" "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}/LICENSE-LLVM-OpenMP")
+    add_dependencies(ggml-base ggml-openmp-runtime)
+    install(FILES "${GGML_OPENMP_RUNTIME}" DESTINATION ${CMAKE_INSTALL_BINDIR})
+    install(FILES "${GGML_OPENMP_LICENSE}" DESTINATION ${CMAKE_INSTALL_BINDIR} RENAME LICENSE-LLVM-OpenMP)
+
+    set(GGML_OPENMP_TARGET_C ggml-openmp-c)
+    set(GGML_OPENMP_TARGET_CXX ggml-openmp-cxx)
+    set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
+elseif (GGML_OPENMP)
    find_package(OpenMP)
    if (OpenMP_FOUND)
+        set(GGML_OPENMP_TARGET_C OpenMP::OpenMP_C)
+        set(GGML_OPENMP_TARGET_CXX OpenMP::OpenMP_CXX)
        set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
    else()
        set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
@@ -236,7 +350,7 @@ endif()

 if (GGML_OPENMP_ENABLED)
    target_compile_definitions(ggml-base PRIVATE GGML_USE_OPENMP)
-    target_link_libraries(ggml-base PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+    target_link_libraries(ggml-base PRIVATE ${GGML_OPENMP_TARGET_C} ${GGML_OPENMP_TARGET_CXX})
 endif()

 add_library(ggml
@@ -1599,11 +1599,23 @@ 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 +1778,12 @@ 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;
@@ -74,7 +74,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)

    if (GGML_OPENMP_ENABLED)
        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
-        target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+        target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_OPENMP_TARGET_C} ${GGML_OPENMP_TARGET_CXX})
    endif()

    if (GGML_LLAMAFILE)
@@ -639,6 +639,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)

        set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -701,6 +702,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
@@ -737,8 +740,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
    endif()

-    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
-        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
-        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
-    endif()
+  if (CMAKE_C_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+      # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+      target_compile_options(${GGML_CPU_NAME} PRIVATE "$<$<OR:$<COMPILE_LANG_AND_ID:C,IntelLLVM>,$<COMPILE_LANG_AND_ID:CXX,IntelLLVM>>:$<$<BOOL:${WIN32}>:/clang:>-fno-associative-math>")
+  endif()
+
 endfunction()
@@ -23,6 +23,7 @@
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
 #include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.h"
 #include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.h"

 #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
@@ -76,6 +77,21 @@ static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
    Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
 }

+template <void (*Fn)(size_t, size_t, size_t, const void *, size_t, const void *, void *, size_t, size_t, float, float)>
+static inline void kernel_run_lhs_stride_fn10(size_t       m,
+                                              size_t       n,
+                                              size_t       k,
+                                              size_t       lhs_stride,
+                                              const void * lhs,
+                                              const void * rhs,
+                                              void *       dst,
+                                              size_t       dst_stride_row,
+                                              size_t       dst_stride_col,
+                                              float        clamp_min,
+                                              float        clamp_max) {
+    Fn(m, n, k, lhs, lhs_stride, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
 template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
 static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
                                         const void* lhs, const void* rhs, void* dst,
@@ -947,25 +963,25 @@ static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
        },
-        /* SME GEMV */
+        /* SME2 GEMV */
        {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ nullptr,
-            /* .get_rhs_packed_offset_ex = */ nullptr,
-            /* .run_kernel_ex         = */ nullptr,
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_mr                = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .run_kernel_ex         = */ &kernel_run_lhs_stride_fn10<kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
        },
        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
+            /* .get_offset            = */ nullptr,
+            /* .get_packed_offset_ex  = */ nullptr,
+            /* .packed_size_ex        = */ nullptr,
+            /* .pack_func_ex          = */ nullptr,
        },
        /* .rhs_info = */ {
            /* .packed_stride         = */ nullptr,
@@ -696,6 +696,15 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        }

        if (op->src[0]->type == GGML_TYPE_F32) {
+            ggml_kleidiai_kernels * primary     = kernel_chain[0];
+            kernel_info *           gemv_kernel = primary ? &primary->gemv : nullptr;
+            if (is_gemv && op->src[1]->nb[0] == (int64_t) sizeof(float) && gemv_kernel &&
+                gemv_kernel->get_lhs_offset_ex && gemv_kernel->get_rhs_packed_offset_ex &&
+                gemv_kernel->run_kernel_ex && gemv_kernel->get_dst_offset) {
+                size = 0;
+                return true;
+            }
+
            size_t cursor = 0;
            bool any_slot = false;

@@ -811,15 +820,28 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            return false;
        }

-        kernel_info * kernel        = &kernels->gemm;
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+        const bool use_gemv = m == 1 && src1->nb[0] == (int64_t) sizeof(float) &&
+                              kernels->gemv.get_lhs_offset_ex &&
+                              kernels->gemv.get_rhs_packed_offset_ex &&
+                              kernels->gemv.run_kernel_ex &&
+                              kernels->gemv.get_dst_offset;
+
+        kernel_info * kernel        = use_gemv ? &kernels->gemv : &kernels->gemm;
        lhs_packing_info * lhs_info = &kernels->gemm_lhs_info;

-        if (!kernel || !lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
-            !lhs_info->packed_size_ex || !lhs_info->pack_func_ex ||
+        if (!kernel || !kernel->get_lhs_offset_ex ||
            !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
            return false;
        }

+        if (!use_gemv && (!lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
+                          !lhs_info->packed_size_ex || !lhs_info->pack_func_ex)) {
+            return false;
+        }
+
        const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
        const bool has_header = kleidiai_is_weight_header_valid(header);

@@ -832,16 +854,14 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        const int nth = params->nth > 0 ? params->nth : 1;
        const int ith = params->ith;

-        const size_t k = ne00;
-        const size_t m = ne11;
-        const size_t n = ne01;
-
        const size_t mr = kernel->get_mr();
        const size_t kr = kernel->get_kr();
        const size_t sr = kernel->get_sr();

-        const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
-        GGML_ASSERT(lhs_packed_size <= params->wsize);
+        const size_t lhs_packed_size = use_gemv ? 0 : lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
+        if (!use_gemv) {
+            GGML_ASSERT(lhs_packed_size <= params->wsize);
+        }

        uint8_t * lhs_packed   = static_cast<uint8_t *>(params->wdata);
        const size_t dst_stride = dst->nb[1];
@@ -853,7 +873,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
            uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];

-            {
+            if (!use_gemv) {
                const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)mr);
                int64_t max_threads = mr ? (m_roundup_mr / (int64_t)mr) : nth;
                max_threads = std::max<int64_t>(1, max_threads);
@@ -903,15 +923,17 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                const size_t n_to_process = std::min(chunk_cols, n - n_start);

                if (n_to_process > 0) {
-                    const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
+                    const size_t lhs_offset = use_gemv ? kernel->get_lhs_offset_ex(0, k, 0)
+                                                       : lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
                    const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
                    const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);

-                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * lhs_ptr = use_gemv ? lhs_batch_base + lhs_offset
+                                                    : lhs_packed + lhs_offset;
                    const void * rhs_ptr = rhs_base + rhs_packed_offset;
                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);

-                    kernel->run_kernel_ex(m, n_to_process, k, 0,
+                    kernel->run_kernel_ex(m, n_to_process, k, use_gemv ? src1->nb[1] : 0,
                                          lhs_ptr,
                                          rhs_ptr,
                                          dst_ptr,
@@ -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,
@@ -290,6 +290,42 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
    if (!ggml_is_quantized(type)) {
        return false;
    }
+    // k-quants cost more to decode and mvq redoes that per column, so MMQ wins sooner.
+    // Only list quant-types MMQ supports, others would fall back to cuBLAS.
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_ADA_LOVELACE) {
+        switch (type) { // tuned on RTX 4090
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 4;
+            case GGML_TYPE_Q3_K:
+                return ne11 <= 6;
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_BLACKWELL) {
+        switch (type) { // tuned on RTX 5090
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 5;
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_DGX_SPARK) {
+        switch (type) { // tuned on DGX Spark GB10
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 6;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
    if (GGML_CUDA_CC_IS_CDNA(cc)) {
        if (GGML_CUDA_CC_IS_CDNA1(cc)) {
            switch (type) {
@@ -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);

@@ -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;
@@ -3180,8 +3180,9 @@ 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
+    // ggml_rope_set_offset: HVX kernels need a VLEN-aligned window start (32 f32 elems)
+    if (op_params[15] % 32 != 0) {
+        return false;
    }

    int mode = op_params[2];
@@ -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;
@@ -53,6 +53,7 @@

 struct htp_rope_context {
    int32_t n_dims;
+    int32_t n_offs;
    int32_t mode;
    int32_t n_ctx_orig;
    int32_t sections[4];
@@ -405,32 +406,40 @@ static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict

 static void inline rope_basic_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
    #pragma unroll(4)
    for (uint32_t i = 0; i < nr; i++) {
        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
        float * s = (float *) (src + i * rctx->src0_row_size_aligned);

-        hvx_rope_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);

        // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
        }
    }
 }

 static void inline rope_neox_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
    #pragma unroll(4)
    for (uint32_t i = 0; i < nr; i++) {
        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
        float * s = (float *) (src + i * rctx->src0_row_size_aligned);

-        hvx_rope_neox_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_neox_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);

        // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
        }
    }
 }
@@ -673,6 +682,7 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
    rctx.n_dims     = ((const int32_t *) op_params)[1];
    rctx.mode       = ((const int32_t *) op_params)[2];
    rctx.n_ctx_orig = ((const int32_t *) op_params)[4];
+    rctx.n_offs     = ((const int32_t *) op_params)[15];

    memcpy(&rctx.freq_base,   (int32_t *) op_params + 5,  sizeof(float));
    memcpy(&rctx.freq_scale,  (int32_t *) op_params + 6,  sizeof(float));
@@ -1409,6 +1409,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_p
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    char base[256];
+
+    snprintf(base, 256, "kernel_flash_attn_ext_kv_%s_f16", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, base);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, base, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
        ggml_metal_library_t lib,
        const struct ggml_tensor * op,
@@ -1460,7 +1477,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
-        int32_t nsg) {
+        int32_t nsg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);

    char base[256];
@@ -1469,15 +1489,14 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
    const int32_t dk = (int32_t) op->src[1]->ne[0];
    const int32_t dv = (int32_t) op->src[2]->ne[0];

-    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
-    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
+    const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);

    // do bounds checks for the mask?
    const bool bc_mask = op->src[3] && (op->src[3]->ne[1] % 8 != 0);

    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
            "flash_attn_ext",
-            ggml_type_name(op->src[1]->type),
+            type,
            dk,
            dv);

@@ -1526,7 +1545,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
        bool    has_scap,
        bool    has_kvpad,
        int32_t nsg,
-        int32_t nwg) {
+        int32_t nwg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);

    char base[256];
@@ -1535,12 +1557,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
    const int32_t dk = (int32_t) op->src[1]->ne[0];
    const int32_t dv = (int32_t) op->src[2]->ne[0];

-    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
-    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
+    const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);

    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
            "flash_attn_ext_vec",
-            ggml_type_name(op->src[1]->type),
+            type,
            dk,
            dv);

@@ -176,6 +176,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_mask,
        int32_t ncpsg);

+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op);
+
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
        ggml_metal_library_t lib,
        const struct ggml_tensor * op,
@@ -190,7 +194,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
-        int32_t nsg);
+        int32_t nsg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20);

 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec(
        ggml_metal_library_t lib,
@@ -201,7 +208,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_scap,
        bool    has_kvpad,
        int32_t nsg,
-        int32_t nwg);
+        int32_t nwg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20);

 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(
        ggml_metal_library_t lib,
@@ -345,6 +345,18 @@ typedef struct {
    bool     inplace;
 } ggml_metal_kargs_rope;

+typedef struct {
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  nblocks;
+} ggml_metal_kargs_flash_attn_ext_kv_f16;
+
 typedef struct {
    int32_t  ne11;
    int32_t  ne_12_2; // assume K and V are same shape
@@ -2801,6 +2801,51 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const ggml_tensor * op) {
    return (ne01 < 20) && (ne00 % 32 == 0);
 }

+// ref: https://github.com/ggml-org/llama.cpp/pull/27390
+// dequantize the quantized KV cache to F16 before running the F16 flash attention kernels
+static bool ggml_metal_op_flash_attn_ext_use_kv_f16(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    // depending on compute/bandwidth ratio, dequant to f16 kv is not always beneficial
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27390#issuecomment-5355152767
+    // TODO: tune per device
+    if (op->src[0]->ne[1] < 32) {
+        return false;
+    }
+
+    switch (op->src[1]->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:
+            return true;
+        default:
+            return false;
+    }
+}
+
+// in some models (e.g. MLA-based), V is a view of K (the first ne20 elements of each K row);
+// the dequantized V is then a view of the dequantized K and does not need its own dequant or scratch
+// - ref: https://github.com/ggml-org/llama.cpp/pull/13435
+static bool ggml_metal_op_flash_attn_ext_v_is_view_of_k(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    const ggml_tensor * K = op->src[1];
+    const ggml_tensor * V = op->src[2];
+
+    return V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));
+}
+
+// size of the F16 dequantized K tensor; the dequantized V tensor follows it in the same scratch buffer
+static size_t ggml_metal_op_flash_attn_ext_kv_f16_k_size(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+
+    return GGML_PAD(sizeof(ggml_fp16_t)*(size_t) ne10*ne11*ne12*ne13, 16);
+}
+
 size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);

@@ -2816,6 +2861,18 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
    size_t res = 0;

    const bool has_mask = op->src[3] != nullptr;
+    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    // when the KV is dequantized to F16, the pad kernel copies the tail chunk from the F16 scratch buffer
+    // note: when V is a view of K, the dequantized V is read from the dequantized K with K's row stride
+    const bool v_is_view_of_k = use_kv_f16 && ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+    uint64_t nb11_pad = nb11;
+    uint64_t nb21_pad = nb21;
+
+    if (use_kv_f16) {
+        nb11_pad = sizeof(ggml_fp16_t)*ne10;
+        nb21_pad = sizeof(ggml_fp16_t)*(v_is_view_of_k ? ne10 : ne20);
+    }

    // note: the non-vec kernel requires more extra memory, so always reserve for it
    GGML_ASSERT(OP_FLASH_ATTN_EXT_NCPSG >= OP_FLASH_ATTN_EXT_VEC_NCPSG);
@@ -2828,8 +2885,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {

        if (has_kvpad) {
            res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
-                nb11*ne12*ne13 +
-                nb21*ne22*ne23 +
+                nb11_pad*ne12*ne13 +
+                nb21_pad*ne22*ne23 +
                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
        }
    } else {
@@ -2838,8 +2895,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {

        if (has_kvpad) {
            res += OP_FLASH_ATTN_EXT_NCPSG*(
-                nb11*ne12*ne13 +
-                nb21*ne22*ne23 +
+                nb11_pad*ne12*ne13 +
+                nb21_pad*ne22*ne23 +
                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
        }
    }
@@ -2915,6 +2972,29 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
    return res;
 }

+size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    // note: always reserve the temp buffer to avoid graph reallocations
+    //if (!ggml_metal_op_flash_attn_ext_use_kv_f16(op)) {
+    //    return 0;
+    //}
+
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+
+    const size_t k_size = ggml_metal_op_flash_attn_ext_kv_f16_k_size(op);
+
+    // when V is a view of K, the dequantized V is a view of the dequantized K
+    const bool v_is_view_of_k = ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+    if (v_is_view_of_k) {
+        return k_size;
+    }
+
+    const size_t v_size = GGML_PAD(sizeof(ggml_fp16_t)*(size_t) ne20*ne21*ne22*ne23, 16);
+
+    return k_size + v_size;
+}
+
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -2989,6 +3069,111 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
    ggml_metal_buffer_id bid_tmp = bid_blk;
    bid_tmp.offs += ggml_metal_op_flash_attn_ext_extra_blk(op);

+    ggml_metal_buffer_id bid_kv_f16 = bid_tmp;
+    bid_kv_f16.offs += ggml_metal_op_flash_attn_ext_extra_tmp(op);
+
+    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    ggml_metal_buffer_id bid_k = bid_src1;
+    ggml_metal_buffer_id bid_v = bid_src2;
+
+    uint64_t nb10_attn = nb10;
+    uint64_t nb11_attn = nb11;
+    uint64_t nb12_attn = nb12;
+    uint64_t nb13_attn = nb13;
+    uint64_t nb20_attn = nb20;
+    uint64_t nb21_attn = nb21;
+    uint64_t nb22_attn = nb22;
+    uint64_t nb23_attn = nb23;
+
+    if (use_kv_f16) {
+        assert(ggml_metal_op_flash_attn_ext_extra_kv_f16(op) != 0);
+
+        const bool v_is_view_of_k = ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+
+        const int64_t nblocks1_64 = (ne10/ggml_blck_size(op->src[1]->type))*(int64_t) ne11*ne12*ne13;
+        GGML_ASSERT(nblocks1_64 <= INT32_MAX);
+        const int32_t nblocks1 = nblocks1_64;
+
+        ggml_metal_buffer_id bid_v_f16 = bid_kv_f16;
+        bid_v_f16.offs += ggml_metal_op_flash_attn_ext_kv_f16_k_size(op);
+
+        auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(lib, op);
+        const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline0), 256);
+
+        // K
+        ggml_metal_kargs_flash_attn_ext_kv_f16 args_k = {
+            /*.ne0    =*/ ne10,
+            /*.ne1    =*/ ne11,
+            /*.ne2    =*/ ne12,
+            /*.ne3    =*/ ne13,
+            /*.nb0    =*/ nb10,
+            /*.nb1    =*/ nb11,
+            /*.nb2    =*/ nb12,
+            /*.nb3    =*/ nb13,
+            /*.nblocks =*/ nblocks1,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline0);
+        ggml_metal_encoder_set_bytes   (enc, &args_k, sizeof(args_k), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1,        1);
+        ggml_metal_encoder_set_buffer  (enc, bid_kv_f16, 2);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, (nblocks1 + nth - 1)/nth, 1, 1, nth, 1, 1);
+
+        // V (skip when V is a view of K: the dequantized V is a view of the dequantized K)
+        if (!v_is_view_of_k) {
+            const int64_t nblocks2_64 = (ne20/ggml_blck_size(op->src[2]->type))*(int64_t) ne21*ne22*ne23;
+            GGML_ASSERT(nblocks2_64 <= INT32_MAX);
+            const int32_t nblocks2 = nblocks2_64;
+
+            ggml_metal_kargs_flash_attn_ext_kv_f16 args_v = {
+                /*.ne0    =*/ ne20,
+                /*.ne1    =*/ ne21,
+                /*.ne2    =*/ ne22,
+                /*.ne3    =*/ ne23,
+                /*.nb0    =*/ nb20,
+                /*.nb1    =*/ nb21,
+                /*.nb2    =*/ nb22,
+                /*.nb3    =*/ nb23,
+                /*.nblocks =*/ nblocks2,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args_v, sizeof(args_v), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2,        1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v_f16,       2);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, (nblocks2 + nth - 1)/nth, 1, 1, nth, 1, 1);
+        }
+
+        // the pad and attention kernels read the dequantized KV
+        ggml_metal_op_concurrency_reset(ctx);
+
+        bid_k = bid_kv_f16;
+        bid_v = v_is_view_of_k ? bid_k : bid_v_f16;
+
+        // contiguous F16 layout of the dequantized K
+        nb10_attn = sizeof(ggml_fp16_t);
+        nb11_attn = nb10_attn*ne10;
+        nb12_attn = nb11_attn*ne11;
+        nb13_attn = nb12_attn*ne12;
+
+        // if V is a view of K, the dequantized V is read from the dequantized K with K's strides
+        if (v_is_view_of_k) {
+            nb20_attn = nb10_attn;
+            nb21_attn = nb11_attn;
+            nb22_attn = nb12_attn;
+            nb23_attn = nb13_attn;
+        } else {
+            // contiguous F16 layout of the dequantized V
+            nb20_attn = sizeof(ggml_fp16_t);
+            nb21_attn = nb20_attn*ne20;
+            nb22_attn = nb21_attn*ne21;
+            nb23_attn = nb22_attn*ne22;
+        }
+    }
+
    if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
        // half8x8 kernel
        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
@@ -3009,12 +3194,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                /*.ne11    =*/ne11,
                /*.ne_12_2 =*/ne12,
                /*.ne_12_3 =*/ne13,
-                /*.nb11    =*/nb11,
-                /*.nb12    =*/nb12,
-                /*.nb13    =*/nb13,
-                /*.nb21    =*/nb21,
-                /*.nb22    =*/nb22,
-                /*.nb23    =*/nb23,
+                /*.nb11    =*/nb11_attn,
+                /*.nb12    =*/nb12_attn,
+                /*.nb13    =*/nb13_attn,
+                /*.nb21    =*/nb21_attn,
+                /*.nb22    =*/nb22_attn,
+                /*.nb23    =*/nb23_attn,
                /*.ne31    =*/ne31,
                /*.ne32    =*/ne32,
                /*.ne33    =*/ne33,
@@ -3027,8 +3212,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

            ggml_metal_encoder_set_pipeline(enc, pipeline0);
            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
-            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
-            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_k,    1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v,    2);
            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);

@@ -3073,7 +3258,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            ggml_metal_op_concurrency_reset(ctx);
        }

-        const int is_q = ggml_is_quantized(op->src[1]->type) ? 1 : 0;
+        const int is_q = !use_kv_f16 && ggml_is_quantized(op->src[1]->type) ? 1 : 0;

        // 2*(2*ncpsg)
        // ncpsg soft_max values + ncpsg mask values
@@ -3104,6 +3289,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

        const size_t smem = FATTN_SMEM(nsg);

+        const int32_t ns10 = nb11_attn/nb10_attn;
+        const int32_t ns20 = nb21_attn/nb20_attn;
+
        ggml_metal_kargs_flash_attn_ext args = {
            /*.ne01          =*/ ne01,
            /*.ne02          =*/ ne02,
@@ -3114,14 +3302,14 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.ne11          =*/ ne11,
            /*.ne_12_2       =*/ ne12,
            /*.ne_12_3       =*/ ne13,
-            /*.ns10          =*/ int32_t(nb11/nb10),
-            /*.nb11          =*/ nb11,
-            /*.nb12          =*/ nb12,
-            /*.nb13          =*/ nb13,
-            /*.ns20          =*/ int32_t(nb21/nb20),
-            /*.nb21          =*/ nb21,
-            /*.nb22          =*/ nb22,
-            /*.nb23          =*/ nb23,
+            /*.ns10          =*/ ns10,
+            /*.nb11          =*/ nb11_attn,
+            /*.nb12          =*/ nb12_attn,
+            /*.nb13          =*/ nb13_attn,
+            /*.ns20          =*/ ns20,
+            /*.nb21          =*/ nb21_attn,
+            /*.nb22          =*/ nb22_attn,
+            /*.nb23          =*/ nb23_attn,
            /*.ne31          =*/ ne31,
            /*.ne32          =*/ ne32,
            /*.ne33          =*/ ne33,
@@ -3139,13 +3327,13 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.logit_softcap =*/ logit_softcap,
        };

-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, use_kv_f16, ns10, ns20);

        ggml_metal_encoder_set_pipeline(enc, pipeline);
        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
-        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
-        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_k,    2);
+        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
        ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
@@ -3177,12 +3365,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                /*.ne11    =*/ne11,
                /*.ne_12_2 =*/ne12,
                /*.ne_12_3 =*/ne13,
-                /*.nb11    =*/nb11,
-                /*.nb12    =*/nb12,
-                /*.nb13    =*/nb13,
-                /*.nb21    =*/nb21,
-                /*.nb22    =*/nb22,
-                /*.nb23    =*/nb23,
+                /*.nb11    =*/nb11_attn,
+                /*.nb12    =*/nb12_attn,
+                /*.nb13    =*/nb13_attn,
+                /*.nb21    =*/nb21_attn,
+                /*.nb22    =*/nb22_attn,
+                /*.nb23    =*/nb23_attn,
                /*.ne31    =*/ne31,
                /*.ne32    =*/ne32,
                /*.ne33    =*/ne33,
@@ -3195,8 +3383,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

            ggml_metal_encoder_set_pipeline(enc, pipeline0);
            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
-            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
-            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_k,    1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v,    2);
            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);

@@ -3242,6 +3430,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            }
        }

+        const int32_t ns10 = nb11_attn/nb10_attn;
+        const int32_t ns20 = nb21_attn/nb20_attn;
+
        ggml_metal_kargs_flash_attn_ext_vec args = {
            /*.ne01          =*/ ne01,
            /*.ne02          =*/ ne02,
@@ -3252,14 +3443,14 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.ne11          =*/ ne11,
            /*.ne_12_2       =*/ ne12,
            /*.ne_12_3       =*/ ne13,
-            /*.ns10          =*/ int32_t(nb11/nb10),
-            /*.nb11          =*/ nb11,
-            /*.nb12          =*/ nb12,
-            /*.nb13          =*/ nb13,
-            /*.ns20          =*/ int32_t(nb21/nb20),
-            /*.nb21          =*/ nb21,
-            /*.nb22          =*/ nb22,
-            /*.nb23          =*/ nb23,
+            /*.ns10          =*/ ns10,
+            /*.nb11          =*/ nb11_attn,
+            /*.nb12          =*/ nb12_attn,
+            /*.nb13          =*/ nb13_attn,
+            /*.ns20          =*/ ns20,
+            /*.nb21          =*/ nb21_attn,
+            /*.nb22          =*/ nb22_attn,
+            /*.nb23          =*/ nb23_attn,
            /*.ne31          =*/ ne31,
            /*.ne32          =*/ ne32,
            /*.ne33          =*/ ne33,
@@ -3277,15 +3468,15 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.logit_softcap =*/ logit_softcap,
        };

-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg, use_kv_f16, ns10, ns20);

        GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));

        ggml_metal_encoder_set_pipeline(enc, pipeline);
        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
-        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
-        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_k,    2);
+        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);

@@ -42,6 +42,7 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_pad(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_blk(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
+size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const struct ggml_tensor * op);

 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_repeat            (ggml_metal_op_t ctx, int idx);
@@ -225,6 +225,7 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
                res += ggml_metal_op_flash_attn_ext_extra_pad(tensor);
                res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
                res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_kv_f16(tensor);
            } break;
        case GGML_OP_CUMSUM:
        case GGML_OP_ARGSORT:
@@ -6318,6 +6318,53 @@ template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f
 template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
 template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;

+// dequantize a quantized KV cache tensor to contiguous F16 before running the F16 flash attention kernels
+// - one thread per block; dispatched separately for K and V
+// - ref: https://github.com/ggml-org/llama.cpp/pull/27390
+template <
+    typename block_t,
+    short QK,
+    void (*deq_t4x4)(device const block_t *, short, thread float4x4 &)>
+kernel void kernel_flash_attn_ext_kv_f16(
+        constant ggml_metal_kargs_flash_attn_ext_kv_f16 & args,
+        device const char * x,
+        device       half * x_dst,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= (uint) args.nblocks) {
+        return;
+    }
+
+    const uint nb = args.ne0/QK;
+    const uint i0 = gid%nb;
+    uint ib       = gid/nb;
+    const uint i1 = ib%args.ne1;
+    ib /= args.ne1;
+    const uint i2 = ib%args.ne2;
+    const uint i3 = ib/args.ne2;
+
+    const uint64_t offs = i0*args.nb0 + i1*args.nb1 + i2*args.nb2 + i3*args.nb3;
+
+    device const block_t * src = (device const block_t *) (x + offs);
+    device half4 * dst = (device half4 *) x_dst + (QK/4)*gid;
+
+    for (short i = 0; i < QK/16; ++i) {
+        float4x4 reg;
+        deq_t4x4(src, i, reg);
+        dst[4*i + 0] = (half4) reg[0];
+        dst[4*i + 1] = (half4) reg[1];
+        dst[4*i + 2] = (half4) reg[2];
+        dst[4*i + 3] = (half4) reg[3];
+    }
+}
+
+typedef decltype(kernel_flash_attn_ext_kv_f16<block_q8_0, 32, dequantize_q8_0>) kernel_flash_attn_ext_kv_f16_t;
+
+template [[host_name("kernel_flash_attn_ext_kv_q4_0_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q4_0, 32, dequantize_q4_0>;
+template [[host_name("kernel_flash_attn_ext_kv_q4_1_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q4_1, 32, dequantize_q4_1>;
+template [[host_name("kernel_flash_attn_ext_kv_q5_0_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q5_0, 32, dequantize_q5_0>;
+template [[host_name("kernel_flash_attn_ext_kv_q5_1_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q5_1, 32, dequantize_q5_1>;
+template [[host_name("kernel_flash_attn_ext_kv_q8_0_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q8_0, 32, dequantize_q8_0>;
+
 constant bool FC_flash_attn_ext_pad_has_mask [[function_constant(FC_FLASH_ATTN_EXT_PAD + 0)]];

 constant int32_t FC_flash_attn_ext_pad_ncpsg [[function_constant(FC_FLASH_ATTN_EXT_PAD + 25)]];
@@ -10318,9 +10365,12 @@ kernel void kernel_mul_mm(
    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));

    // Configure matmul operation
+    // note: K is dynamic_extent (clamped to the valid range in PHASE 2), since a static
+    //       N_MM_NK_TOTAL K tile would read src1 out of bounds when K % N_MM_NK_TOTAL != 0
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27064
    mpp::tensor_ops::matmul2d<
        mpp::tensor_ops::matmul2d_descriptor(
-            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            NRB, NRA, static_cast<int>(dynamic_extent), false, true, true,
            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;

@@ -10372,10 +10422,14 @@ kernel void kernel_mul_mm(
        threadgroup_barrier(mem_flags::mem_threadgroup);

        // === PHASE 2: Tensor matmul ===
-        auto mA = tA.slice(0, 0);
-        auto mB = tB.slice(loop_k, rb);
+        // Clamp the K extent of both operand tensors to the remaining valid K range so
+        // the dynamic-K op never reads past the K extent of src1 (or the staged A tile).
+        const int kExt = min(N_MM_NK_TOTAL, K - loop_k);

-        mm.run(mB, mA, cT);
+        auto tAv = tensor(sa, dextents<int32_t, 2>(kExt, NRA), array<int, 2>({1, N_MM_NK_TOTAL}));
+        auto tBv = tensor(ptrB + loop_k + rb * strideB, dextents<int32_t, 2>(kExt, N - rb), array<int, 2>({1, strideB}));
+
+        mm.run(tBv, tAv, cT);

        threadgroup_barrier(mem_flags::mem_threadgroup);
    }
@@ -202,6 +202,7 @@ set(GGML_OPENCL_KERNELS
    sqr
    sqrt
    ssm_conv
+    ssm_scan
    gated_delta_net
    sub
    sum_rows
@@ -582,6 +582,8 @@ struct ggml_backend_opencl_context {
    bool adreno_use_bin_kernels;
    get_adreno_bin_kernel_func_t get_adreno_bin_kernel_func = nullptr;
    ggml_cl_compiler_version adreno_cl_compiler_version;
+    // The q6_K flat mul_mat codegen workarounds are needed by old E031 compilers only.
+    bool q6_k_flat_old_compiler;

    std::string kernel_compile_opts;  // cached for lazy-compiled kernels.

@@ -866,6 +868,9 @@ 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 +897,7 @@ 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;
@@ -1927,8 +1933,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
 #else
        const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
 #endif
+        // The codegen workarounds in this kernel are a measured 13-20% loss on
+        // compilers that do not need them, so only the affected ones build them;
+        // everyone else gets the original source.
+        const std::string q6k_opts = backend_ctx->q6_k_flat_old_compiler
+            ? compile_opts + " -DADRENO_OLD_COMPILER=1"
+            : compile_opts;
        cl_program prog =
-            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+            build_program_from_source(backend_ctx, kernel_src.c_str(), q6k_opts);

        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
        CL_CHECK(clReleaseProgram(prog));
@@ -3154,6 +3166,24 @@ 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 +4472,7 @@ 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(".");
    }
@@ -5894,6 +5925,16 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);

+    // The q6_K flat mul_mat miscompile is a defect of the older E031 compilers, not a
+    // property of any GPU generation: it reproduces on E031.38 (Adreno 642L) and E031.41
+    // (Adreno 740) and is fixed by E031.45 (Adreno 619). Gate on the compiler so parts
+    // that do not need the workarounds do not pay for them. The explicit type check is
+    // required: newer_than_or_same() is false for every non-E031 compiler, so negating it
+    // alone would enable the workarounds on E17/DX.
+    backend_ctx->q6_k_flat_old_compiler =
+        backend_ctx->adreno_cl_compiler_version.type == E031 &&
+        !backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 45, 0, 0);
+
    size_t ext_str_size;
    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
    char *ext_buffer = (char *)alloca(ext_str_size + 1);
@@ -7301,6 +7342,23 @@ 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}.
@@ -7335,6 +7393,19 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                       op->src[0]->type == GGML_TYPE_Q4_K  ||
                       op->src[0]->type == GGML_TYPE_Q5_K  ||
                       op->src[0]->type == GGML_TYPE_Q6_K) {
+                // The E031.41 compiler (usually with A7x) miscompiles the flat K-quant
+                // GEMV kernels (kernel_mul_mv_q*_K_f32_flat) and makes lm_head run much
+                // slower than it should. So, make it fallback to CPU to preserve performance
+                // for this compiler series.
+                static const char * a7x_lmhead_env = getenv("GGML_OPENCL_A7X_LMHEAD_CPU");
+                static const bool   a7x_lmhead_cpu = (a7x_lmhead_env == nullptr || a7x_lmhead_env[0] != '0');
+                if (a7x_lmhead_cpu &&
+                    backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                    (op->src[0]->type == GGML_TYPE_Q4_K || op->src[0]->type == GGML_TYPE_Q5_K ||
+                     op->src[0]->type == GGML_TYPE_Q6_K) &&
+                    op->src[0]->ne[1] >= 32768) {   // vocab-scale weight; no FFN/attn weight is this tall
+                    return false;
+                }
                return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
            } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
                return op->src[1]->type == GGML_TYPE_F32;
@@ -7376,9 +7447,6 @@ 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;
@@ -7456,6 +7524,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                     v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
            const bool is_f32_f16  = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
                                     v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
            const bool is_f32_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 &&
                                     v->type == GGML_TYPE_Q8_0 && op->type == GGML_TYPE_F32 &&
                                     dk % 32 == 0 && dv % 32 == 0;
@@ -7463,6 +7532,21 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                     v->type == GGML_TYPE_Q4_0 && op->type == GGML_TYPE_F32 &&
                                     dk % 32 == 0 && dv % 32 == 0;

+            // A7X (Adreno 740, compiler E031.41) SIGSEGVs inside clBuildProgram
+            // building the flash_attn programs whose KV path is mixed-type or
+            // dequantized — f32_f16, q8_0, q4_0 (reproduced at DK=40 and DK=64; it
+            // is DK-independent). It is a driver crash, not codegen-wrong-output, so
+            // it cannot be caught in-process (fatal=false only handles clean compile
+            // errors). The uniform f16_f16 / f32_f32 programs compile fine on this
+            // compiler, so decline only the KV-convert variants; ggml then runs
+            // those (f16-KV / quant-KV) attention layers on the CPU backend.
+            // Negative compiler carve-out, same idiom as the Intel DK=512 decline
+            // below and the X1E driver-quirk guards.
+            if (backend_ctx && backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                (is_f32_f16 || is_f32_q8_0 || is_f32_q4_0)) {
+                return false;
+            }
+
            // Asymmetric KV: host-dequants both sides to F32, uses f32 kernel.
            auto is_kv_type_ok = [](ggml_type t) {
                return t == GGML_TYPE_F16 || t == GGML_TYPE_F32 ||
@@ -12260,6 +12344,103 @@ 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);
@@ -12693,7 +12874,10 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);

-    const int nth = MIN(64, ne00);
+    int nth = 1;
+    while (nth < ne00 && nth < 64) {
+        nth *= 2;
+    }

    cl_kernel kernel = backend_ctx->kernel_norm;

@@ -20443,6 +20627,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+            // The optimizer-barrier arg exists only in the ADRENO_OLD_COMPILER build of
+            // this kernel; conformant compilers get the original 17-arg signature.
+            if (backend_ctx->q6_k_flat_old_compiler) {
+                cl_uchar q6k_mask = 0xFF;   // never 0xFE in prod; see the kernel note
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_uchar), &q6k_mask));
+            }
 #else
            kernel = backend_ctx->kernel_mul_mv_q6_K_f32;

@@ -20728,18 +20918,42 @@ 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
@@ -23706,6 +23920,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    const int n_dims     = ((int *) dst->op_params)[1];
    const int mode       = ((int *) dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];

    float freq_base;
    float freq_scale;
@@ -23734,6 +23949,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const

    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
    }

    cl_kernel kernel;
@@ -23825,6 +24041,12 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    if (is_mrope && !is_vision) {
        CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
    }
+    // norm and neox have n_offs after beta_slow, mrope has it after is_imrope
+    if (!is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int), &n_offs));
+    } else if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 35, sizeof(int), &n_offs));
+    }

    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
    size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -24746,6 +24968,14 @@ 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;
@@ -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,
@@ -28,6 +28,13 @@

 #define QK_K       256

+// ADRENO_OLD_COMPILER is defined by the host (-D) only for the Adreno E031
+// compilers older than E031.45, which miscompile several constructs this kernel
+// used (confirmed on E031.38 and E031.41; E031.45 is clean). Every other
+// compiler -- newer E031, E17, DX, Intel, and every non-Adreno device that
+// builds this program -- takes the #else branches, which are the original
+// source: the workarounds below cost ~13% on the q6_K flat n=1 GEMV where they
+// are not needed.
 inline float block_q_6_K_dot_y_flat(
    global uchar * blk_ql,
    global uchar * blk_qh,
@@ -37,6 +44,9 @@ inline float block_q_6_K_dot_y_flat(
    int ip,
    int is,
    int l0,
+#if defined(ADRENO_OLD_COMPILER)
+    int dbg,
+#endif
    float4 y0,
    float4 y1,
    float4 y2,
@@ -48,10 +58,40 @@ inline float block_q_6_K_dot_y_flat(
    global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
    global uchar * q2 = q1         + QK_K/8;
    global uchar * qh = blk_qh     + ib*64 + q_offset_h;
-    global char  * sc = blk_scales + ib*16 + is;

    float dall = blk_d[ib];

+#if defined(ADRENO_OLD_COMPILER)
+    // The vectorized dequant (int4/float4 bit-ops, convert_*4, dot()) and vload4
+    // are miscompiled here -> garbage weights. Reconstruct the 6-bit weights and
+    // take the dot product scalar. q4_K/q5_K flat already use scalar paths, which
+    // is why q6_K was the only flat GEMV that failed.
+    // Scales are SIGNED int8; read as uchar and sign-extend arithmetically so the
+    // result does not depend on whether the compiler treats `char` as signed.
+    global uchar * sc = (global uchar *)(blk_scales + ib*16 + is);
+
+    int s0 = (int)sc[0] - 256*(sc[0] >> 7);
+    int s2 = (int)sc[2] - 256*(sc[2] >> 7);
+    int s4 = (int)sc[4] - 256*(sc[4] >> 7);
+    int s6 = (int)sc[6] - 256*(sc[6] >> 7);
+
+    // one 6-bit weight: low/high nibble of a ql byte OR'd with a 2-bit qh plane
+    // (plane p in {0,1,2,3} selects qh bits 2p..2p+1) placed at bits 4-5, minus 32.
+    #define Q6W(qb, sh, hb, p) ((float)((((int)(qb) >> (sh)) & 15) | ((((int)(hb) >> (2*(p))) & 3) << 4)) - 32.f)
+
+    float d0 = y0.s0*Q6W(q1[0],0,qh[0],0) + y0.s1*Q6W(q1[1],0,qh[1],0) + y0.s2*Q6W(q1[2],0,qh[2],0) + y0.s3*Q6W(q1[3],0,qh[3],0);
+    float d1 = y1.s0*Q6W(q2[0],0,qh[0],1) + y1.s1*Q6W(q2[1],0,qh[1],1) + y1.s2*Q6W(q2[2],0,qh[2],1) + y1.s3*Q6W(q2[3],0,qh[3],1);
+    float d2 = y2.s0*Q6W(q1[0],4,qh[0],2) + y2.s1*Q6W(q1[1],4,qh[1],2) + y2.s2*Q6W(q1[2],4,qh[2],2) + y2.s3*Q6W(q1[3],4,qh[3],2);
+    float d3 = y3.s0*Q6W(q2[0],4,qh[0],3) + y3.s1*Q6W(q2[1],4,qh[1],3) + y3.s2*Q6W(q2[2],4,qh[2],3) + y3.s3*Q6W(q2[3],4,qh[3],3);
+    #undef Q6W
+
+    if (dbg) printf("HELPER dall=%f s=[%d %d %d %d] d=[%f %f %f %f] ql0=%d qh0=%d y00=%f\n",
+        dall, s0, s2, s4, s6, d0, d1, d2, d3, (int)q1[0], (int)qh[0], y0.s0);
+
+    return dall * (d0 * s0 + d1 * s2 + d2 * s4 + d3 * s6);
+#else
+    global char * sc = blk_scales + ib*16 + is;
+
    // Vectorized loads: 3 uchar4 weight loads instead of 12 scalar byte reads.
    // q_offset_l/h are 4-aligned, so these are aligned vector loads.
    uchar4 q1v = vload4(0, q1);
@@ -72,6 +112,7 @@ inline float block_q_6_K_dot_y_flat(

    return dall * (dot(y0, w0) * sc[0] + dot(y1, w1) * sc[2] +
                   dot(y2, w2) * sc[4] + dot(y3, w3) * sc[6]);
+#endif
 }

 #undef N_DST
@@ -113,6 +154,11 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
        int ne1,
        int r2,
        int r3
+#if defined(ADRENO_OLD_COMPILER)
+        ,
+        uchar q6k_mask   // runtime 0xFF; the host passes it so the compiler cannot
+                         // constant-fold the printf guards below into nothing
+#endif
 ) {
    src1 = (global float*)((global char*)src1 + offset1);
    dst = (global float*)((global char*)dst + offsetd);
@@ -128,6 +174,22 @@ kernel void kernel_mul_mv_q6_K_f32_flat(

    int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;

+#if defined(ADRENO_OLD_COMPILER)
+    // 64-bit `ulong` integer arithmetic is miscompiled here -> the base-pointer byte
+    // offsets came out wrong, so EVERY weight/scale read hit the wrong address. This
+    // was the primary cause of the q6_K flat failure (q5_K uses int offsets and is
+    // unaffected). Compute the block index in `int` and widen to `ulong` only inside
+    // the pointer expression: the byte offset stays 64-bit, but there is no ulong
+    // arithmetic chain to miscompile. The int index would overflow past ~2^31 blocks,
+    // which no realistic weight reaches -- but that is a narrowing, so keep it off the
+    // conformant path, which retains full ulong arithmetic.
+    int offset_src0 = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + (ulong)offset_src0 * 128;
+    global uchar * blk_qh     = (global uchar *) src0_qh + (ulong)offset_src0 * 64;
+    global char  * blk_scales = (global char  *) src0_s  + (ulong)offset_src0 * 16;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0;
+#else
    ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
    ulong offset_src0_ql = offset_src0 * 128;
    ulong offset_src0_qh = offset_src0 * 64;
@@ -138,6 +200,7 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
    global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
    global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
    global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+#endif
    global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;

    int tid = get_sub_group_local_id()%(N_SIMDWIDTH/BLOCK_STRIDE); // within-super-block part, 0..15
@@ -155,24 +218,55 @@ kernel void kernel_mul_mv_q6_K_f32_flat(

    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
        global float * y = yy + ib * QK_K + 128*ip + l0;
+#if defined(ADRENO_OLD_COMPILER)
+        // vload4 of f32 is miscompiled here; index the lanes scalar instead.
+        float4 y0 = (float4)(y[ 0], y[ 1], y[ 2], y[ 3]);
+        float4 y1 = (float4)(y[32], y[33], y[34], y[35]);
+        float4 y2 = (float4)(y[64], y[65], y[66], y[67]);
+        float4 y3 = (float4)(y[96], y[97], y[98], y[99]);
+#else
        float4 y0 = vload4(0, y +  0);
        float4 y1 = vload4(0, y + 32);
        float4 y2 = vload4(0, y + 64);
        float4 y3 = vload4(0, y + 96);
+#endif

        for (int row = 0; row < N_DST; row++) {
            if (first_row + row < ne01) {
+#if defined(ADRENO_OLD_COMPILER)
+                int dbg = (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && row==0 && ib==0 &&
+                           ne00==256 && ne01==16 && get_sub_group_local_id()==0) ? 1 : 0;
+                sumf[row] += block_q_6_K_dot_y_flat(
+                    blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
+                    ib, ip, is, l0, dbg, y0, y1, y2, y3);
+#else
                sumf[row] += block_q_6_K_dot_y_flat(
                    blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
                    ib, ip, is, l0, y0, y1, y2, y3);
+#endif
            }
        }
    }

+#if defined(ADRENO_OLD_COMPILER)
+    // Optimizer barrier. This compiler drops the sumf partials unless a side effect
+    // forces them to materialize. q6k_mask is a kernel arg the compiler cannot prove
+    // is never 0xFE (the host always passes 0xFF), so the printf survives compilation
+    // but never executes. FRAGILE: the exact set and placement of these guarded
+    // printfs is load-bearing on E031.41 -- removing any one re-breaks q6_K.
+    if (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && ne00==256 && ne01==16 && get_sub_group_local_id()<16) {
+        printf("Q6KLANE lane=%d ip=%d il=%d is=%d l0=%d sumf0=%f\n",
+            get_sub_group_local_id(), ip, il, is, l0, sumf[0]);
+    }
+#endif
    for (int row = 0; row < N_DST; row++) {
        float tot = sub_group_reduce_add(sumf[row]);
        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+#if defined(ADRENO_OLD_COMPILER)
+            if (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && row==0 && ne00==256 && ne01==16)
+                printf("Q6KTOT tot=%f\n", tot);
+#endif
        }
    }
 }
@@ -75,7 +75,8 @@ kernel void kernel_rope_norm_f32(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -94,14 +95,15 @@ kernel void kernel_rope_norm_f32(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

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

            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

            global float * src       = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -154,7 +156,8 @@ kernel void kernel_rope_norm_f16(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -173,14 +176,15 @@ kernel void kernel_rope_norm_f16(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

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

            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -233,7 +237,8 @@ kernel void kernel_rope_neox_f32(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -252,17 +257,18 @@ kernel void kernel_rope_neox_f32(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

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

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -312,7 +318,8 @@ kernel void kernel_rope_neox_f16(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -331,17 +338,18 @@ kernel void kernel_rope_neox_f16(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

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

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -393,7 +401,8 @@ kernel void kernel_rope_multi_f32(
        float beta_fast,
        float beta_slow,
        int4 sections,
-        int  is_imrope
+        int  is_imrope,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -414,10 +423,11 @@ kernel void kernel_rope_multi_f32(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            const int sector = (i0 / 2) % sect_dims;
+            const int sector = ic % sect_dims;
            float theta_base = 0.0f;

            if (is_imrope) {
@@ -445,14 +455,14 @@ kernel void kernel_rope_multi_f32(
                }
            }

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

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -504,7 +514,8 @@ kernel void kernel_rope_multi_f16(
        float beta_fast,
        float beta_slow,
        int4 sections,
-        int  is_imrope
+        int  is_imrope,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -525,10 +536,11 @@ kernel void kernel_rope_multi_f16(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            const int sector = (i0 / 2) % sect_dims;
+            const int sector = ic % sect_dims;
            float theta_base = 0.0f;

            if (is_imrope) {
@@ -556,14 +568,14 @@ kernel void kernel_rope_multi_f16(
                }
            }

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

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -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;
+}
@@ -1955,6 +1955,23 @@ static void dequantize_mul_mat_vec_q4_K_sycl_reorder_esimd(const void *vx, const
    });
 }

+static void dequantize_mul_mat_vec_q5_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q5_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
 static void dequantize_mul_mat_vec_q6_K_sycl_reorder_esimd(const void *vx, const float *y,
                                                           float *dst, const int ncols,
                                                           const int nrows,
@@ -2134,7 +2151,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
        case GGML_TYPE_Q5_K:
            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q5_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q5_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q5_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
            } else {
                dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
            }
@@ -62,7 +62,7 @@

 #define DPCT_UNUSED(x) (void)(x)

-inline void _abort(const char * str) {
+[[noreturn]] inline void _abort(const char * str) {
    std::cerr << str << std::endl;
    std::abort();
 }
@@ -10,7 +10,7 @@
    (ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))

 static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2,
    const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
    const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
@@ -455,7 +455,7 @@ static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, c
 namespace ggml_sycl_detail {
 static void acc_f32_sycl(const char *x, const char *y, float *dst,
                         const int64_t n_elements,
-                         const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+                         const int64_t ne0, const int64_t ne1, const int64_t ne2,
                         const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
                         const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
                         const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
@@ -466,7 +466,7 @@ static void acc_f32_sycl(const char *x, const char *y, float *dst,
                                           sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
                         [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                            acc_f32(x, y, dst, n_elements,
-                                ne0, ne1, ne2, ne3,
+                                ne0, ne1, ne2,
                                nb00, nb01, nb02, nb03,
                                ne10, ne11, ne12, ne13,
                                nb10, nb11, nb12, nb13,
@@ -970,7 +970,7 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
    const int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);

    ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),
-        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2],
        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
        src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
        src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
@@ -1,15 +1,3 @@
-//
-// MIT license
-// Copyright (C) 2026 Intel Corporation
-// SPDX-License-Identifier: MIT
-//
-
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-
 #ifndef GGML_SYCL_ESIMD_HPP
 #define GGML_SYCL_ESIMD_HPP

@@ -287,6 +275,128 @@ template <> struct esimd_reorder_q_traits<GGML_TYPE_Q4_K> {
    }
 };

+// ---------------------------------------------------------------------------
+// Q5_K, SOA reorder layout produced by reorder_qw_q5_k:
+//   [qs: nb*(QK_K/2)] [qh: nb*(QK_K/8)] [scales: nb*K_SCALE_SIZE] [dm: nb*sizeof(half2)]
+// with nb = nrows*num_blocks_per_row.
+//
+// Identical to Q4_K except each 4-bit quant gains a 5th (high) bit from qh:
+// output chunk c (0..7) adds 16 when bit c of qh[l] is set, where qh[l] indexes
+// the same 32 bytes for every chunk (matches dequantize_row_q5_K).
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q5_K> {
+    struct ptrs {
+        const uint8_t *    qs;
+        const uint8_t *    qh;
+        const uint8_t *    scales;
+        const sycl::half * dm;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t * qs     = (const uint8_t *) vx;
+        const uint8_t * qh     = qs + nb * (QK_K / 2);
+        const uint8_t * scales = qh + nb * (QK_K / 8);
+        const sycl::half * dm  = (const sycl::half *) (scales + nb * K_SCALE_SIZE);
+        return { qs, qh, scales, dm };
+    }
+
+    // extract bit `bit` (0..7) of each lane and move it to bit position 4,
+    // e.g. for the 4-bit base quant's 5th (high) bit. `bit` is always a
+    // compile-time-known unrolled loop constant at call sites, so this folds
+    // to a single mask (bit==4), mask+left-shift (bit<4), or mask+right-shift
+    // (bit>4) instead of the shift+mask+shift a naive `(qh>>bit & 1) << 4` emits.
+    static ESIMD_INLINE sycl::ext::intel::esimd::simd<uint16_t, 32> extract_bit_to_pos4(
+            sycl::ext::intel::esimd::simd<uint8_t, 32> qh, int bit) {
+        using namespace sycl::ext::intel::esimd;
+        simd<uint16_t, 32> masked = convert<uint16_t>(qh & simd<uint8_t, 32>((uint8_t) (1u << bit)));
+        if (bit < 4) {
+            return masked << simd<uint16_t, 32>((uint16_t) (4 - bit));
+        } else if (bit > 4) {
+            return masked >> simd<uint16_t, 32>((uint16_t) (bit - 4));
+        }
+        return masked;
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 128> qs_a     = block_load<uint8_t, 128>(pa.qs + bia * (QK_K / 2));
+        simd<uint8_t, 128> qs_b     = 0;
+        simd<uint8_t, 32>  qh_a     = block_load<uint8_t, 32>(pa.qh + bia * (QK_K / 8));
+        simd<uint8_t, 32>  qh_b     = 0;
+        simd<uint8_t, 12>  scales_a = block_load<uint8_t, 12>(pa.scales + bia * K_SCALE_SIZE);
+        simd<uint8_t, 12>  scales_b = 0;
+
+        const float dall_a = (float) pa.dm[bia * 2 + 0];
+        const float dmin_a = (float) pa.dm[bia * 2 + 1];
+        float dall_b = 0.0f;
+        float dmin_b = 0.0f;
+        if (has_b) {
+            qs_b     = block_load<uint8_t, 128>(pb.qs + bib * (QK_K / 2));
+            qh_b     = block_load<uint8_t, 32>(pb.qh + bib * (QK_K / 8));
+            scales_b = block_load<uint8_t, 12>(pb.scales + bib * K_SCALE_SIZE);
+            dall_b = (float) pb.dm[bib * 2 + 0];
+            dmin_b = (float) pb.dm[bib * 2 + 1];
+        }
+
+        simd<float, 8> scale_f_a, min_f_a, scale_f_b, min_f_b;
+        unpack_scale_min_k4(scales_a, dall_a, dmin_a, scale_f_a, min_f_a);
+        unpack_scale_min_k4(scales_b, dall_b, dmin_b, scale_f_b, min_f_b);
+
+        simd<uint8_t, 128> qs_lo_a = qs_a & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_a = qs_a >> simd<uint8_t, 128>(4);
+        simd<uint8_t, 128> qs_lo_b = qs_b & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_b = qs_b >> simd<uint8_t, 128>(4);
+
+#pragma unroll
+        for (int sb = 0; sb < 8; sb += 2) {
+            const int q_offset = sb * 16;
+            simd<float, 32> y_lo = y_vec.select<32, 1>(sb * 32);
+            simd<float, 32> y_hi = y_vec.select<32, 1>((sb + 1) * 32);
+
+            const float scale_a_lo = scale_f_a[sb];
+            const float scale_a_hi = scale_f_a[sb + 1];
+            const float min_a_lo   = min_f_a[sb];
+            const float min_a_hi   = min_f_a[sb + 1];
+            const float scale_b_lo = scale_f_b[sb];
+            const float scale_b_hi = scale_f_b[sb + 1];
+            const float min_b_lo   = min_f_b[sb];
+            const float min_b_hi   = min_f_b[sb + 1];
+
+            simd<uint8_t, 32> qa_lo_u8 = qs_lo_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qa_hi_u8 = qs_hi_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_lo_u8 = qs_lo_b.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_hi_u8 = qs_hi_b.select<32, 1>(q_offset);
+            simd<uint16_t, 32> qa_lo = convert<uint16_t>(qa_lo_u8);
+            simd<uint16_t, 32> qa_hi = convert<uint16_t>(qa_hi_u8);
+            simd<uint16_t, 32> qb_lo = convert<uint16_t>(qb_lo_u8);
+            simd<uint16_t, 32> qb_hi = convert<uint16_t>(qb_hi_u8);
+
+            // add the 5th bit: chunk sb uses qh bit sb, chunk sb+1 uses qh bit sb+1;
+            // qh always indexes the same 32 bytes regardless of chunk
+            qa_lo += extract_bit_to_pos4(qh_a, sb);
+            qa_hi += extract_bit_to_pos4(qh_a, sb + 1);
+            qb_lo += extract_bit_to_pos4(qh_b, sb);
+            qb_hi += extract_bit_to_pos4(qh_b, sb + 1);
+
+            simd<float, 32> deq_a_lo = convert<float>(qa_lo) * scale_a_lo + min_a_lo;
+            simd<float, 32> deq_a_hi = convert<float>(qa_hi) * scale_a_hi + min_a_hi;
+            simd<float, 32> deq_b_lo = convert<float>(qb_lo) * scale_b_lo + min_b_lo;
+            simd<float, 32> deq_b_hi = convert<float>(qb_hi) * scale_b_hi + min_b_hi;
+
+            acc_a += y_lo * deq_a_lo;
+            acc_b += y_lo * deq_b_lo;
+            acc_a += y_hi * deq_a_hi;
+            acc_b += y_hi * deq_b_hi;
+        }
+    }
+};
+
 // ---------------------------------------------------------------------------
 // Q6_K, SOA reorder layout:
 //   [ql: nb*(QK_K/2)] [qh: nb*(QK_K/4)] [scales(int8): nb*(QK_K/16)] [d: nb*half]
@@ -43,7 +43,7 @@ static void mkl_fa_pack_q_fp16(
    dpct::queue_ptr stream,
    sycl::half * __restrict dst,
    const float * __restrict q_src,
-    int n_queries, int n_query_rows, int DKQ,
+    int n_queries, int DKQ,
    int gqa_ratio, int kvh_base_head,
    float q_scale, int64_t q_row_stride, int64_t q_head_stride,
    int64_t wg_size) {
@@ -121,7 +121,7 @@ static void mkl_fa_online_softmax_chunk(
    float * __restrict VKQ_accum,
    int q0, int q_rows, int n_queries, int DV,
    int chunk_size, int chunk_start,
-    int kvh_head, int gqa_ratio,
+    int kvh_head,
    const sycl::half * mask_data, int64_t mask_head_stride,
    int64_t mask_row_stride, int mask_n_heads,
    float logit_softcap, int64_t wg_size) {
@@ -473,7 +473,6 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
    MKL_ACCUM(dequant_time_us, t_deq);

    // --- Resolve mask pointers ---
-    const sycl::half * mask_data = nullptr;
    int64_t mask_head_stride = 0;
    int64_t mask_row_stride  = 0;
    int     mask_n_heads     = 0;
@@ -547,7 +546,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
            // 1. Pack all GQA Q heads into fp16 (full n_query_rows)
            mkl_fa_pack_q_fp16(stream,
                Q_head_f16_ptr, Q_batch,
-                n_queries, n_query_rows, DKQ,
+                n_queries, DKQ,
                gqa_ratio, kvh_base_head,
                q_scale, q_row_stride, q_head_stride, wg_size);

@@ -605,7 +604,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
                            KQ_max_ptr, KQ_sum_ptr, VKQ_accum_ptr,
                            q0, q_rows, n_queries, DV,
                            this_chunk, chunk_start,
-                            kvh_base_head, gqa_ratio,
+                            kvh_base_head,
                            mask_batch, mask_head_stride,
                            mask_row_stride, mask_n_heads,
                            logit_softcap, wg_size);
@@ -21,14 +21,6 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
    if (!g_ggml_sycl_fa_onednn) {
        return false;
    }
-    // Battlemage (Xe2) only, for now. On other Intel archs oneDNN's fused SDPA returns wrong results
-    // for some shapes (e.g. head_dim=64 on Arc / xe_hpg) -- an oneDNN bug tracked upstream at
-    // https://github.com/uxlfoundation/oneDNN/issues/5510. Remove this hardware limitation once that
-    // is fixed; until then non-BMG archs fall back to the existing FA kernel.
-    const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
-    if (arch != gpu_arch::intel_gpu_bmg_g21 && arch != gpu_arch::intel_gpu_bmg_g31) {
-        return false;
-    }
    const ggml_tensor * Q     = dst->src[0];
    const ggml_tensor * K     = dst->src[1];
    const ggml_tensor * V     = dst->src[2];
@@ -60,6 +52,17 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
            }
        }
    }
+    // This is the improved SPDA gate. Rather than gating Alchemist GPUs from all SPDA features, we instead target only the failing shapes.
+    // If the GPU being assessed isn't in the grouping below, it has full access to all SPDA shapes. Otherwise, if it's an Alchemist GPU, we block only the shapes with head sizes that fail.
+    // It is much easier to compare the device to a small list of failing cases than to define all the passing ones.
+    const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
+    bool support_spda = !(arch == gpu_arch::intel_gpu_dg2_g10 ||
+                         arch == gpu_arch::intel_gpu_dg2_g11 ||
+                         arch == gpu_arch::intel_gpu_dg2_g12);
+
+    if (!support_spda && K->ne[0] == 64) {
+        return false;
+    }
    // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
    // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
    // some stacks; past the cap we fall back to the native FA kernel instead.
@@ -921,16 +921,16 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,

    void * dev_ptr;
    if (use_usm_system) {
-        GGML_SYCL_DEBUG("[SYCL] allocating %lu Bytes with USM system\n", size);
+        GGML_SYCL_DEBUG("[SYCL] allocating %zu Bytes with USM system\n", size);
        dev_ptr = (void *)aligned_malloc_host(alignment, aligned_size);
        if (!dev_ptr) {
-            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);
+            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);
            return nullptr;
        }
    } else {
        SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)ggml_sycl_malloc_device(size, *stream)));
        if (!dev_ptr) {
-          GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+          GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
          return nullptr;
        }
    }
@@ -1177,7 +1177,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
        SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)ggml_sycl_malloc_device(size, *stream)));
        if (!buf) {
            char err_buf[1024];
-            snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+            snprintf(err_buf, 1023, "%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
            throw std::runtime_error(err_buf);
        }
        // set padding to 0 to avoid possible NaN values
@@ -1517,8 +1517,13 @@ static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggm
 }

 static size_t ggml_backend_sycl_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    ggml_backend_sycl_device_context * dev_ctx = (ggml_backend_sycl_device_context *) buft->device->context;
-    return dpct::dev_mgr::instance().get_device(dev_ctx->device).get_max_mem_alloc_size();
+
+    if (g_ggml_sycl_enable_host_pinned_mem) {
+        ggml_backend_sycl_device_context * dev_ctx = (ggml_backend_sycl_device_context *) buft->device->context;
+        return dpct::dev_mgr::instance().get_device(dev_ctx->device).get_max_mem_alloc_size();
+    } else {
+        return SIZE_MAX;
+    }
 }

 ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
@@ -1646,7 +1651,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {

        SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *)ggml_sycl_malloc_device(look_ahead_size, *qptr)));
        if (!ptr) {
-            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
+            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
            return nullptr;
        }

@@ -1658,7 +1663,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
                (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
 #endif

-        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);
+        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%zu, return %p\n", look_ahead_size, ptr);
        return ptr;
    }

@@ -1838,7 +1843,7 @@ struct ggml_sycl_pool_host : public ggml_sycl_pool {

            SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *) sycl::malloc_host(size, *qptr)));
            if (!ptr) {
-                GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);
+                GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);
                return nullptr;
            }
            pool_size += size;
@@ -2774,9 +2779,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();

        {
+#if GGML_SYCL_DNNL
            const int64_t gemm_flops = (int64_t)row_diff * src1_ncols * ne10;
            const bool use_mkl_direct = gemm_flops < 256 * 256 * 256;
-#if GGML_SYCL_DNNL
            if (g_ggml_sycl_enable_dnn && !use_mkl_direct) {
                DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
                                          DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
@@ -3513,7 +3518,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
    float *            dst_ddf  = static_cast<float *>(dst->data);

    const sycl::half * src1_f16       = static_cast<const sycl::half *>(src1->data);
+#if GGML_SYCL_DNNL
    const size_t       type_size_src0 = ggml_type_size(src0->type);
+#endif
    const size_t       type_size_src1 = ggml_type_size(src1->type);

    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
@@ -3530,6 +3537,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
        scope_op_debug_print    scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2,
                                                " : converting src1 to fp16");

+#if GGML_SYCL_DNNL
        // iterate tensor dims and find the slowest moving dim and stride
        int last_dim=0;
        int last_str=0;
@@ -3549,7 +3557,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
            }

        }
-#if GGML_SYCL_DNNL
        // oneDNN handles strided data and does not need overhead of ggml_get_to_fp16_nc_sycl
        const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
        src1_f16_alloc.alloc(ne_src1);
@@ -3804,6 +3811,7 @@ static bool ggml_sycl_supports_reorder_esimd(enum ggml_type type) {
    switch (type) {
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
        case GGML_TYPE_Q6_K:
            return true;
        default:
@@ -6242,8 +6250,6 @@ 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:
@@ -85,7 +85,7 @@ static void im2col_sycl(const float *   x,
    */
    stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
                                           sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3>) {
                             im2col_kernel(x, dst, IC, IW, IH, OH, OW, KW, KH, IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,
                                           s0, s1, p0, p1, d0, d1);
                         });
@@ -271,7 +271,7 @@ static void im2col_3d_sycl(const float *   src,
    */
    stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
                                           sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3>) {
                             im2col_3d_kernel(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, OH_OW, KD_KH_KW,
                                              ID_IH_IW, KH_KW, IH_IW, IC_ID_IH_IW, IC_KD_KH_KW, OW_KD_KH_KW,
                                              OD_OH_OW_IC_KD_KH_KW, OH_OW_IC_KD_KH_KW, OW_IC_KD_KH_KW, N_OD_OH, OD_OH,
@@ -7,9 +7,6 @@ static void norm_f32(const float* x, float* dst, const int ncols,
    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
    const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {

-    const int nrows = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
-
    const int nthreads = item_ct1.get_local_range(2);
    const int sample  = item_ct1.get_group(0);
    const int channel = item_ct1.get_group(1);
@@ -155,9 +152,6 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols,
    const float* mul = nullptr, const int64_t mul_stride_row = 0, const int64_t mul_stride_channel = 0,
    const int64_t mul_stride_sample = 0, const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {

-    const int nrows = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
-
    const int sample  = item_ct1.get_group(0);
    const int channel = item_ct1.get_group(1);
    const int row     = item_ct1.get_group(2);
@@ -225,8 +219,6 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
    const int64_t src_stride_sample, const int64_t dst_stride_col, const int64_t dst_stride_row,
    const int64_t dst_stride_channel, const int64_t dst_stride_sample, const float eps,
    const sycl::nd_item<3>& item_ct1, float* s_sum, const int block_size) {
-    const int nrows     = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);

    const int row     = item_ct1.get_group(2);
    const int channel = item_ct1.get_group(1);
@@ -41,7 +41,7 @@ template <bool forward, bool has_ff, typename T, typename D>
 static void rope_norm(const T *x, D *dst, const int ne00, const int ne01,
                      const int ne02, const int s01, const int s02,
                      const int s03, const int s1, const int s2, const int s3,
-                      const int n_dims, const int32_t *pos,
+                      const int n_dims, const int n_offs, const int32_t *pos,
                      const float freq_scale, const float ext_factor,
                      const float attn_factor, const rope_corr_dims corr_dims,
                      const float theta_scale, const float *freq_factors,
@@ -78,19 +78,21 @@ static void rope_norm(const T *x, D *dst, const int ne00, const int ne01,
            ggml_sycl_memcpy_1<4>(dst + idst, &v);
        }
    };
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
        store_coaelsced(x[ix + 0], x[ix + 1]);
        return;
    }

-    const float theta_base = pos[i2] * dpct::pow(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] * dpct::pow(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,
+    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];
@@ -104,7 +106,7 @@ template <bool forward, bool has_ff, typename T, typename D>
 static void rope_neox(const T *x, D *dst, const int ne00, const int ne01,
                      const int ne02, const int s01, const int s02,
                      const int s03, const int s1, const int s2, const int s3,
-                      const int n_dims, const int32_t *pos,
+                      const int n_dims, const int n_offs, const int32_t *pos,
                      const float freq_scale, const float ext_factor,
                      const float attn_factor, const rope_corr_dims corr_dims,
                      const float theta_scale, const float *freq_factors,
@@ -132,35 +134,38 @@ static void rope_neox(const T *x, D *dst, const int ne00, const int ne01,
        idst += row_indices[i2] * set_rows_stride;
    }

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

        return;
    }

-    const float theta_base = pos[i2] * dpct::pow(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] * dpct::pow(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,
+    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_sycl_cast<D>(x0 * cos_theta - x1 * sin_theta);
-    dst[idst + n_dims / 2] = ggml_sycl_cast<D>(x0 * sin_theta + x1 * cos_theta);
+    dst[idst + n_offs / 2 + 0] = ggml_sycl_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_offs / 2 + n_dims / 2] = ggml_sycl_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

 template <bool forward, bool has_ff, typename T>
 static void rope_multi(const T *x, T *dst, const int ne00, const int ne01,
                       const int ne02, const int s01, const int s02,
                       const int s03, const int s1, const int s2, const int s3,
-                       const int n_dims, const int32_t *pos,
+                       const int n_dims, const int n_offs, const int32_t *pos,
                       const float freq_scale, const float ext_factor,
                       const float attn_factor, const rope_corr_dims corr_dims,
                       const float theta_scale, const float *freq_factors,
@@ -183,54 +188,57 @@ static void rope_multi(const T *x, T *dst, const int ne00, const int ne01,
    int idst = i0 / 2 + i1 * s1 + i2 * s2 + i3 * s3;
    const int ix = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
        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] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
-            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
-            theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
        } else {
-            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, iw / 2.0f);
        }
    } else {
        if (sector < sections.v[0]) {
-            theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * dpct::pow(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,
+    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>
@@ -293,7 +301,7 @@ static void
 rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
               const int ne02, const int s01, const int s02, const int s03,
               const int s1, const int s2, const int s3, const int n_dims,
-               const int nr, const int32_t *pos, const float freq_scale,
+               const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
               const float freq_base, const float ext_factor,
               const float attn_factor, const rope_corr_dims corr_dims,
               const float *freq_factors, const int64_t *row_indices,
@@ -313,7 +321,7 @@ rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_norm<forward, false>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    } else {
@@ -323,7 +331,7 @@ rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_norm<forward, true>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    }
@@ -334,7 +342,7 @@ static void
 rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
               const int ne02, const int s01, const int s02, const int s03,
               const int s1, const int s2, const int s3, const int n_dims,
-               const int nr, const int32_t *pos, const float freq_scale,
+               const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
               const float freq_base, const float ext_factor,
               const float attn_factor, const rope_corr_dims corr_dims,
               const float *freq_factors, const int64_t *row_indices,
@@ -354,7 +362,7 @@ rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_neox<forward, false>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    } else {
@@ -364,7 +372,7 @@ rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_neox<forward, true>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    }
@@ -375,7 +383,7 @@ static void
 rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                const int ne02, const int s01, const int s02, const int s03,
                const int s1, const int s2, const int s3, const int n_dims,
-                const int nr, const int32_t *pos, const float freq_scale,
+                const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
                const float freq_base, const float ext_factor,
                const float attn_factor, const rope_corr_dims corr_dims,
                const float *freq_factors, const mrope_sections sections,
@@ -395,7 +403,7 @@ rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_multi<forward, false, T>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, sections, is_imrope);
            });
    } else {
@@ -405,7 +413,7 @@ rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_multi<forward, true, T>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, sections, is_imrope);
            });
    }
@@ -497,6 +505,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
    const int n_dims = ((int32_t *)dst->op_params)[1];
    const int mode = ((int32_t *)dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *)dst->op_params)[4];
+    const int n_offs = ((int32_t *)dst->op_params)[15];
    mrope_sections sections;

    float freq_base;
@@ -526,6 +535,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,

    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;
@@ -545,19 +555,19 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_neox_sycl<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,
+                s02, 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);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_neox_sycl<forward, float, sycl::half>(
                (const float *)src0_d, (sycl::half *)dst_d, ne00, ne01, ne02,
-                s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                s01, s02, 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);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_neox_sycl<forward, sycl::half, sycl::half>(
                (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, 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);
        } else {
@@ -568,13 +578,13 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
        if (src0->type == GGML_TYPE_F32) {
            rope_multi_sycl<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,
+                                     s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                     ext_factor, attn_factor, corr_dims,
                                     freq_factors, sections, is_imrope, stream);
        } else if (src0->type == GGML_TYPE_F16) {
            rope_multi_sycl<forward>(
                (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                sections, is_imrope, stream);
        } else {
@@ -602,19 +612,19 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_norm_sycl<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,
+                s02, 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);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_norm_sycl<forward, float, sycl::half>(
                (const float *)src0_d, (sycl::half *)dst_d, ne00, ne01, ne02,
-                s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                s01, s02, 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);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_norm_sycl<forward, sycl::half, sycl::half>(
                (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, 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);
        } else {
@@ -291,7 +291,7 @@ static void set_rows_sycl(

    stream->parallel_for(
        sycl::nd_range<1>(grid_size * block_size, block_size),
-        [=](sycl::nd_item<1> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
            k_set_rows<TIn, TIdx, TOut>(
                src0_d, src1_d, dst_d,
                ne00, ne01, ne02,
@@ -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()
@@ -121,13 +121,13 @@ void main() {
        const uint buf_ib = r * qf_stride + d / 8;
        const uint buf_iqs = d % 8;

-        FLOAT_TYPEV4 vals = is_in_bounds ? FLOAT_TYPEV4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale) : FLOAT_TYPEV4(0.0f);
-        const FLOAT_TYPEV4 abs_vals = abs(vals);
+        vec4 vals = is_in_bounds ? data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale : vec4(0.0f);
+        const vec4 abs_vals = abs(vals);

-        const FLOAT_TYPE thread_max = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
-        const FLOAT_TYPE amax = subgroupClusteredMax(thread_max, 8);
-        const FLOAT_TYPE qd = amax / FLOAT_TYPE(127.0);
-        const FLOAT_TYPE qd_inv = qd != FLOAT_TYPE(0.0) ? FLOAT_TYPE(1.0) / qd : FLOAT_TYPE(0.0);
+        const float thread_max = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
+        const float amax = subgroupClusteredMax(thread_max, 8);
+        const float qd = amax / 127.0f;
+        const float qd_inv = qd != 0.0f ? 1.0f / qd : 0.0f;
        vals = round(vals * qd_inv);

        Qf[buf_ib].qs[buf_iqs] = pack32(i8vec4(vals));
@@ -136,11 +136,11 @@ void main() {
        // the row-sum scaled by qd, used in k_dot_correction.
        if (FaTypeK == FA_TYPE_Q8_0) {
            if (buf_iqs == 0) {
-                Qf[buf_ib].ds = FLOAT_TYPEV2(qd, 0.0);
+                Qf[buf_ib].ds = FLOAT_TYPEV2(qd, 0.0f);
            }
        } else {
-            const FLOAT_TYPE thread_sum = vals.x + vals.y + vals.z + vals.w;
-            const FLOAT_TYPE sum = subgroupClusteredAdd(thread_sum, 8);
+            const float thread_sum = vals.x + vals.y + vals.z + vals.w;
+            const float sum = subgroupClusteredAdd(thread_sum, 8);

            if (buf_iqs == 0) {
                Qf[buf_ib].ds = FLOAT_TYPEV2(qd, sum * qd);
@@ -2714,6 +2714,7 @@ static webgpu_encoded_op ggml_webgpu_rope(webgpu_context & ctx,
    const int n_dims     = ((int32_t *) dst->op_params)[1];
    const int mode       = ((int32_t *) dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];

    float freq_base;
    float freq_scale;
@@ -2762,7 +2763,8 @@ static webgpu_encoded_op ggml_webgpu_rope(webgpu_context & ctx,
        (uint32_t) sections[0],
        (uint32_t) sections[1],
        (uint32_t) sections[2],
-        (uint32_t) sections[3]
+        (uint32_t) sections[3],
+        (uint32_t) n_offs
    };

    std::vector<wgpu::BindGroupEntry> entries     = { ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0),
@@ -4472,9 +4474,7 @@ 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:
-            // 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;
+            supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
            break;
        case GGML_OP_GLU:
            switch (ggml_get_glu_op(op)) {
@@ -38,7 +38,8 @@ struct Params {
    sections0: u32,
    sections1: u32,
    sections2: u32,
-    sections3: u32
+    sections3: u32,
+    n_offs: u32
 };

@group(0) @binding(0)
@@ -126,7 +127,8 @@ fn rope_yarn(theta_extrap: f32, i: u32) -> vec2<f32> {

 fn pair_base(i0: u32, div_2: bool) -> u32 {
    if (div_2) {
-        return i0 / 2;
+        // first channel of the rotated pair: n_offs + (i0 - n_offs)/2
+        return i0 / 2 + params.n_offs / 2;
    } else {
        return i0;
    }
@@ -165,20 +167,22 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i_src_row = params.offset_src0 + i3 * params.stride_src03 + i2 * params.stride_src02 + i1 * params.stride_src01;
    let i_dst_row = params.offset_dst + i3 * params.stride_dst3 + i2 * params.stride_dst2 + i1 * params.stride_dst1;

-    if (i0 >= params.n_dims && !is_vision) {
+    if ((i0 < params.n_offs || i0 >= params.n_offs + params.n_dims) && !is_vision) {
        let i_src = i_src_row + i0;
        let i_dst = i_dst_row + i0;
        rotate(i_dst, i_dst + 1, f32(src0[i_src]), f32(src0[i_src + 1]));
        return;
    }

+    let iw = i0 - params.n_offs; // relative idx
+
    var theta_base_mult: u32 = 0;
-    var theta_scale_pwr: u32 = i0 / 2;
+    var theta_scale_pwr: u32 = iw / 2;
    if (is_mrope) {
        let sect_dims = params.sections0 + params.sections1 + params.sections2 + params.sections3;
        let sec_w = params.sections1 + params.sections0;
        let sec_e = params.sections2 + sec_w;
-        let sector = (i0 / 2) % sect_dims;
+        let sector = (iw / 2) % sect_dims;
        if (is_imrope) {
          if (sector % 3 == 1 && sector < 3 * params.sections1) {
              theta_base_mult = 1;
@@ -203,7 +207,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
          } else if (sector >= sec_e) {
              if (is_vision) {
                  theta_scale_pwr = sector - sec_e;
-                  theta_scale_pwr = (i0 / 2) % sec_e;
+                  theta_scale_pwr = (iw / 2) % sec_e;
              }
              theta_base_mult = 3;
          } else if (is_vision) {
@@ -212,7 +216,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
        }
    }
    let theta_base = f32(src1[params.offset_src1 + i2 + params.ne2 * theta_base_mult]) * pow(params.theta_scale, f32(theta_scale_pwr));
-    let thetas = rope_yarn(theta_base/freq_factor(i0), i0);
+    let thetas = rope_yarn(theta_base/freq_factor(iw), iw);

    let i_src = i_src_row + pair_base(i0, is_neox || is_mrope || is_vision);
    let i_dst = i_dst_row + pair_base(i0, is_neox || is_mrope || is_vision);
@@ -86,6 +86,6 @@ endif()

 target_link_libraries(ggml-zendnn PRIVATE m pthread)

-if (GGML_OPENMP)
-    target_link_libraries(ggml-zendnn PRIVATE OpenMP::OpenMP_CXX)
+if (GGML_OPENMP_ENABLED)
+    target_link_libraries(ggml-zendnn PRIVATE ${GGML_OPENMP_TARGET_CXX})
 endif()
@@ -205,6 +205,9 @@ class Keys:
        VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
        KEY_LENGTH_SWA               = "{arch}.attention.key_length_swa"
        VALUE_LENGTH_SWA             = "{arch}.attention.value_length_swa"
+        KEY_LENGTH_MLA_SWA           = "{arch}.attention.key_length_mla_swa"
+        VALUE_LENGTH_MLA_SWA         = "{arch}.attention.value_length_mla_swa"
+        KV_LORA_RANK_SWA             = "{arch}.attention.kv_lora_rank_swa"
        SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
        SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
        TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
@@ -558,6 +561,7 @@ class MODEL_ARCH(IntEnum):
    BAILINGMOE2      = auto()
    BAILINGMOE3      = auto()
    DOTS1            = auto()
+    DOTS3NOTE        = auto()
    ARCEE            = auto()
    AFMOE            = auto()
    LAGUNA           = auto()
@@ -1275,6 +1279,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.BAILINGMOE2:      "bailingmoe2",
    MODEL_ARCH.BAILINGMOE3:      "bailingmoe3",
    MODEL_ARCH.DOTS1:            "dots1",
+    MODEL_ARCH.DOTS3NOTE:        "dots3note",
    MODEL_ARCH.ARCEE:            "arcee",
    MODEL_ARCH.AFMOE:            "afmoe",
    MODEL_ARCH.LAGUNA:           "laguna",
@@ -4334,6 +4339,44 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_UP_EXP,
        MODEL_TENSOR.FFN_UP_SHEXP,
    ],
+    MODEL_ARCH.DOTS3NOTE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.INDEXER_K_NORM,
+        MODEL_TENSOR.INDEXER_PROJ,
+        MODEL_TENSOR.INDEXER_ATTN_K,
+        MODEL_TENSOR.INDEXER_ATTN_Q_B,
+        # NextN/MTP tensors - preserved but unused
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+    ],
    MODEL_ARCH.ARCEE: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -785,6 +785,15 @@ class GGUFWriter:
    def add_key_length_swa(self, length: int) -> None:
        self.add_uint32(Keys.Attention.KEY_LENGTH_SWA.format(arch=self.arch), length)

+    def add_key_length_mla_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KEY_LENGTH_MLA_SWA.format(arch=self.arch), length)
+
+    def add_value_length_mla_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA_SWA.format(arch=self.arch), length)
+
+    def add_kv_lora_rank_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KV_LORA_RANK_SWA.format(arch=self.arch), length)
+
    def add_value_length_swa(self, length: int) -> None:
        self.add_uint32(Keys.Attention.VALUE_LENGTH_SWA.format(arch=self.arch), length)

@@ -723,6 +723,7 @@ class TensorNameMap:
            "model.layers.layers.{bid}.mixer.k",                              # plamo2
            "model.layers.layers.{bid}.mixer.k_norm",                         # plamo3
            "layers.{bid}.self_attn.k_norm",                                  # qwen3-embedding
+            "model.layers.{bid}.self_attn.k_rope_only_layernorm",             # dots3note
            "model.layers.{bid}.attention.key_layernorm",                     # apertus
        ),

@@ -15,7 +15,8 @@
 # tag exists.
 #
 # Env (when running in GitHub Actions):
-#   GITHUB_OUTPUT: previous_tag, changelog_title, changelog and nightly are written here
+#   GITHUB_OUTPUT: previous_tag, changelog_title, changelog, nightly and nightly_tag
+#     are written here
 #   GITHUB_REPOSITORY: owner/repo, used to build the nightly release URL (skipped when unset)
 set -euo pipefail

@@ -52,10 +53,10 @@ PREV="$( { git tag --list; echo "${VERSION}"; } \

 if [[ -n "${PREV}" ]]; then
    CHANGELOG="$(git log --oneline "${PREV}..${RELEASE_COMMIT}")"
-    CHANGELOG_TITLE="Change log since ${PREV}"
+    CHANGELOG_TITLE="Changelog since ${PREV}"
 else
    CHANGELOG="(no previous release tag found)"
-    CHANGELOG_TITLE="Change log"
+    CHANGELOG_TITLE="Changelog"
 fi

 # Nightly release: the b* tag pointing at the release commit (|| true: no match is not an error)
@@ -80,6 +81,7 @@ if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
        echo "previous_tag=${PREV}"
        echo "changelog_title=${CHANGELOG_TITLE}"
        echo "nightly=${NIGHTLY}"
+        echo "nightly_tag=${NIGHTLY_TAG}"
        echo "changelog<<CHANGELOG_EOF"
        echo "${CHANGELOG}"
        echo "CHANGELOG_EOF"
@@ -0,0 +1,204 @@
+#!/bin/bash
+#
+# Release preparation script for llama.cpp.
+#
+# Bumps the version in CMakeLists.txt on a release candidate branch.
+# The branch should then be pushed and a PR created, reviewed, and
+# merged. After the PR is merged and the build-cpu workflow has
+# completed successfully, the release is finalized by the make-release
+# workflow (.github/workflows/make-release.yml), which creates the tag.
+#
+# Usage:
+#   ./scripts/release.sh [major|minor|patch] [--dry-run]
+#
+# Example:
+#   $ ./scripts/release.sh minor
+#
+# The script:
+# 1. Creates a release candidate branch (llama-rc-v<major>.<minor>.<patch>)
+# 2. Bumps the version in CMakeLists.txt
+# 3. Commits the version bump
+#
+
+set -e
+
+if [ ! -f "CMakeLists.txt" ] || [ ! -d "scripts" ]; then
+    echo "Error: Must be run from llama.cpp root directory"
+    exit 1
+fi
+
+# Parse command line arguments
+VERSION_TYPE=""
+DRY_RUN=false
+
+for arg in "$@"; do
+    case $arg in
+        --dry-run)
+            DRY_RUN=true
+            ;;
+        major|minor|patch)
+            VERSION_TYPE="$arg"
+            ;;
+        *)
+            echo "Error: Unknown argument '$arg'"
+            echo "Usage: $0 [major|minor|patch] [--dry-run]"
+            exit 1
+            ;;
+    esac
+done
+
+# Default to patch if no version type specified
+VERSION_TYPE="${VERSION_TYPE:-patch}"
+
+# Common validation functions
+check_git_status() {
+    # Check for uncommitted changes (skip in dry-run)
+    if [ "$DRY_RUN" = false ] && ! git diff-index --quiet HEAD --; then
+        echo "Error: You have uncommitted changes. Please commit or stash them first."
+        exit 1
+    fi
+}
+
+check_master_branch() {
+    # Ensure we're on master branch
+    CURRENT_BRANCH=$(git branch --show-current)
+    if [ "$CURRENT_BRANCH" != "master" ]; then
+        if [ "$DRY_RUN" = true ]; then
+            echo "[dry run] Warning: Not on master branch (currently on: $CURRENT_BRANCH). Continuing with dry-run..."
+            echo ""
+        else
+            echo "Error: Must be on master branch. Currently on: $CURRENT_BRANCH"
+            exit 1
+        fi
+    fi
+}
+
+check_master_up_to_date() {
+    # Check if we have the latest from master (skip in dry-run)
+    if [ "$DRY_RUN" = false ]; then
+        echo "Checking if local master is up-to-date with remote..."
+        git fetch origin master
+        LOCAL=$(git rev-parse HEAD)
+        REMOTE=$(git rev-parse origin/master)
+
+        if [ "$LOCAL" != "$REMOTE" ]; then
+            echo "Error: Your local master branch is not up-to-date with origin/master."
+            echo "Please run 'git pull origin master' first."
+            exit 1
+        fi
+        echo "✓ Local master is up-to-date with remote"
+        echo ""
+    elif [ "$(git branch --show-current)" = "master" ]; then
+        echo "[dry run] Warning: Dry-run mode - not checking if master is up-to-date with remote"
+        echo ""
+    fi
+}
+
+# In-place sed that works on both GNU (Linux) and BSD (macOS) sed
+sed_inplace() {
+    if sed --version >/dev/null 2>&1; then
+        sed -i "$@"
+    else
+        sed -i '' "$@"
+    fi
+}
+
+prepare_release() {
+    if [ "$DRY_RUN" = true ]; then
+        echo "[dry-run] Preparing release (no changes will be made)"
+    else
+        echo "Starting release preparation..."
+    fi
+    echo ""
+
+    check_git_status
+    check_master_branch
+    check_master_up_to_date
+
+    # Extract current version from CMakeLists.txt
+    echo "Step 1: Reading current version..."
+    MAJOR=$(grep "set(LLAMA_VERSION_MAJOR" CMakeLists.txt | sed 's/.*MAJOR \([0-9]*\).*/\1/')
+    MINOR=$(grep "set(LLAMA_VERSION_MINOR" CMakeLists.txt | sed 's/.*MINOR \([0-9]*\).*/\1/')
+    PATCH=$(grep "set(LLAMA_VERSION_PATCH" CMakeLists.txt | sed 's/.*PATCH \([0-9]*\).*/\1/')
+
+    echo "Current version: $MAJOR.$MINOR.$PATCH"
+
+    # Calculate new version
+    case $VERSION_TYPE in
+        major)
+            NEW_MAJOR=$((MAJOR + 1))
+            NEW_MINOR=0
+            NEW_PATCH=0
+            ;;
+        minor)
+            NEW_MAJOR=$MAJOR
+            NEW_MINOR=$((MINOR + 1))
+            NEW_PATCH=0
+            ;;
+        patch)
+            NEW_MAJOR=$MAJOR
+            NEW_MINOR=$MINOR
+            NEW_PATCH=$((PATCH + 1))
+            ;;
+    esac
+
+    NEW_VERSION="$NEW_MAJOR.$NEW_MINOR.$NEW_PATCH"
+    RC_BRANCH="llama-rc-v$NEW_VERSION"
+    echo "New release version: $NEW_VERSION"
+    echo "Release candidate branch: $RC_BRANCH"
+    echo ""
+
+    # Create release candidate branch
+    echo "Step 2: Creating release candidate branch..."
+    if [ "$DRY_RUN" = true ]; then
+        echo "  [dry-run] Would create branch: $RC_BRANCH"
+    else
+        git checkout -b "$RC_BRANCH"
+        echo "✓ Created and switched to branch: $RC_BRANCH"
+    fi
+    echo ""
+
+    # Update CMakeLists.txt for release
+    echo "Step 3: Updating version in CMakeLists.txt..."
+    if [ "$DRY_RUN" = true ]; then
+        echo "  [dry-run] Would update LLAMA_VERSION_MAJOR to $NEW_MAJOR"
+        echo "  [dry-run] Would update LLAMA_VERSION_MINOR to $NEW_MINOR"
+        echo "  [dry-run] Would update LLAMA_VERSION_PATCH to $NEW_PATCH"
+    else
+        sed_inplace -e "s/set(LLAMA_VERSION_MAJOR [0-9]*)/set(LLAMA_VERSION_MAJOR $NEW_MAJOR)/" CMakeLists.txt
+        sed_inplace -e "s/set(LLAMA_VERSION_MINOR [0-9]*)/set(LLAMA_VERSION_MINOR $NEW_MINOR)/" CMakeLists.txt
+        sed_inplace -e "s/set(LLAMA_VERSION_PATCH [0-9]*)/set(LLAMA_VERSION_PATCH $NEW_PATCH)/" CMakeLists.txt
+    fi
+    echo ""
+
+    # Commit version bump
+    echo "Step 4: Committing version bump..."
+    if [ "$DRY_RUN" = true ]; then
+        echo "  [dry-run] Would commit: 'llama.cpp : bump version to $NEW_VERSION'"
+    else
+        git add CMakeLists.txt
+        git commit -m "llama.cpp : bump version to $NEW_VERSION"
+    fi
+    echo ""
+
+    echo ""
+    if [ "$DRY_RUN" = true ]; then
+        echo "[dry-run] Summary (no changes were made):"
+        echo "  • Would have created branch: $RC_BRANCH"
+        echo "  • Would have updated version to: $NEW_VERSION"
+    else
+        echo "Release preparation completed!"
+        echo "Summary:"
+        echo "  • Created branch: $RC_BRANCH"
+        echo "  • Updated version to: $NEW_VERSION"
+        echo ""
+        echo "Next steps:"
+        echo "  • Push branch to remote: git push origin $RC_BRANCH"
+        echo "  • Create a Pull Request from $RC_BRANCH to master"
+        echo "  • After the PR is merged and the build-cpu workflow has passed,"
+        echo "    create the release with the make-release workflow"
+        echo "    (.github/workflows/make-release.yml)"
+    fi
+}
+
+prepare_release
@@ -1 +1 @@
-8c63e70982c95ceb862e3a1073a2c1beef75d60a
+8599e0ea3756c4bac4ef813af2241cb1a8bbfb0b
@@ -25,6 +25,7 @@ add_library(llama
            llama-kv-cache.cpp
            llama-kv-cache-iswa.cpp
            llama-kv-cache-dsa.cpp
+            llama-kv-cache-dsa-iswa.cpp
            llama-kv-cache-msa.cpp
            llama-kv-cache-dsv4.cpp
            llama-memory.cpp
@@ -110,6 +110,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
    { LLM_ARCH_BAILINGMOE3,      "bailingmoe3"      },
    { LLM_ARCH_DOTS1,            "dots1"            },
+    { LLM_ARCH_DOTS3NOTE,        "dots3note"        },
    { LLM_ARCH_ARCEE,            "arcee"            },
    { LLM_ARCH_AFMOE,            "afmoe"            },
    { LLM_ARCH_LAGUNA,           "laguna"           },
@@ -273,6 +274,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
    { LLM_KV_ATTENTION_KEY_LENGTH_SWA,               "%s.attention.key_length_swa"               },
    { LLM_KV_ATTENTION_VALUE_LENGTH_SWA,             "%s.attention.value_length_swa"             },
+    { LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,           "%s.attention.key_length_mla_swa"           },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA,         "%s.attention.value_length_mla_swa"         },
+    { LLM_KV_ATTENTION_KV_LORA_RANK_SWA,             "%s.attention.kv_lora_rank_swa"             },
    { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,           "%s.attention.indexer.head_count"           },
    { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
    { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
@@ -1032,6 +1036,8 @@ bool llm_arch_supports_rs_rollback(const llm_arch & arch) {
        case LLM_ARCH_DEEPSEEK4:
        case LLM_ARCH_NEMOTRON_H:
        case LLM_ARCH_NEMOTRON_H_MOE:
+        case LLM_ARCH_LFM2:
+        case LLM_ARCH_LFM2MOE:
            return true;
        default:
            return false;
@@ -1054,14 +1060,13 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
        case LLM_ARCH_DEEPSEEK2:
        case LLM_ARCH_DEEPSEEK32:
        case LLM_ARCH_DEEPSEEK4:
+        case LLM_ARCH_DOTS3NOTE:
        case LLM_ARCH_GLM_DSA:
        case LLM_ARCH_BITNET:
        case LLM_ARCH_T5:
        case LLM_ARCH_NEMOTRON_H:
        case LLM_ARCH_NEMOTRON_H_MOE:
        case LLM_ARCH_GRANITE_HYBRID:
-        case LLM_ARCH_LFM2:
-        case LLM_ARCH_LFM2MOE:
        case LLM_ARCH_MINIMAX_01:
        case LLM_ARCH_MINIMAX_M2:
        case LLM_ARCH_MINIMAX_M3:
@@ -115,6 +115,7 @@ enum llm_arch {
    LLM_ARCH_BAILINGMOE2,
    LLM_ARCH_BAILINGMOE3,
    LLM_ARCH_DOTS1,
+    LLM_ARCH_DOTS3NOTE,
    LLM_ARCH_ARCEE,
    LLM_ARCH_AFMOE,
    LLM_ARCH_LAGUNA,
@@ -278,6 +279,9 @@ enum llm_kv {
    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
    LLM_KV_ATTENTION_KEY_LENGTH_SWA,
    LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA,
+    LLM_KV_ATTENTION_KV_LORA_RANK_SWA,
    LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
    LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
    LLM_KV_ATTENTION_INDEXER_TOP_K,
@@ -9,6 +9,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-kv-cache-dsa.h"
+#include "llama-kv-cache-dsa-iswa.h"
 #include "llama-kv-cache-msa.h"
 #include "llama-kv-cache-dsv4.h"
 #include "llama-memory-hybrid.h"
@@ -507,10 +508,12 @@ void llm_graph_input_attn_k::set_input(const llama_ubatch * ubatch) {
 }

 bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
+    mctx = static_cast<const llama_kv_cache_context *>(params.mctx);

-    this->mctx = mctx;
+    return can_reuse_impl(params);
+}

+bool llm_graph_input_attn_k::can_reuse_impl(const llm_graph_params & params) {
    bool res = true;

    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
@@ -567,10 +570,12 @@ void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
 }

 bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_dsa_context *>(params.mctx);
+    mctx = static_cast<const llama_kv_cache_dsa_context *>(params.mctx);

-    this->mctx = mctx;
+    return can_reuse_impl(params);
+}

+bool llm_graph_input_attn_k_dsa::can_reuse_impl(const llm_graph_params & params) {
    bool res = true;

    res &= self_k_idxs_mla->ne[0] == params.ubatch.n_tokens;
@@ -582,6 +587,25 @@ bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
    return res;
 }

+void llm_graph_input_attn_k_dsa_iswa::set_input(const llama_ubatch * ubatch) {
+    inp_dsa->set_input(ubatch);
+    inp_swa->set_input(ubatch);
+}
+
+bool llm_graph_input_attn_k_dsa_iswa::can_reuse(const llm_graph_params & params) {
+    mctx = static_cast<const llama_kv_cache_dsa_iswa_context *>(params.mctx);
+
+    inp_dsa->mctx = mctx->get_dsa();
+    inp_swa->mctx = mctx->get_swa();
+
+    bool res = true;
+
+    res &= inp_dsa->can_reuse_impl(params);
+    res &= inp_swa->can_reuse_impl(params);
+
+    return res;
+}
+
 void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
    // base tensors may not be allocated if there are no non-SWA attention layers
    if (self_k_idxs && self_k_idxs->buffer) {
@@ -3099,8 +3123,6 @@ 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 +3155,7 @@ 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);
@@ -3212,8 +3234,12 @@ ggml_tensor * llm_graph_context::build_attn(
    return cur;
 }

-llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_dsa_context *>(mctx);
+static std::unique_ptr<llm_graph_input_attn_k_dsa> build_attn_inp_k_dsa_impl(
+           ggml_context * ctx0,
+     const llama_ubatch & ubatch,
+    const llama_hparams & hparams,
+    const llama_cparams & cparams,
+    const llama_kv_cache_dsa_context * mctx_cur) {

    auto inp = std::make_unique<llm_graph_input_attn_k_dsa>(hparams, cparams, mctx_cur);

@@ -3237,9 +3263,35 @@ llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
        inp->self_k_rot_lid = mctx_cur->get_lid()->build_input_k_rot(ctx0);
    }

+    return inp;
+}
+
+llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_dsa_context *>(mctx);
+
+    auto inp = build_attn_inp_k_dsa_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+
    return (llm_graph_input_attn_k_dsa *) res->add_input(std::move(inp));
 }

+llm_graph_input_attn_k_dsa_iswa * llm_graph_context::build_attn_inp_k_dsa_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_dsa_iswa_context *>(mctx);
+
+    auto inp_dsa = build_attn_inp_k_dsa_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_dsa());
+
+    // build_attn_inp_k_impl rejects SWA caches, so construct the input directly
+    auto inp_swa = std::make_unique<llm_graph_input_attn_k>(hparams, cparams, mctx_cur->get_swa());
+
+    inp_swa->self_k_idxs = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
+
+    inp_swa->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur->get_swa(), ubatch, cparams);
+    inp_swa->self_kq_mask_cnv = inp_swa->self_kq_mask;
+
+    auto inp = std::make_unique<llm_graph_input_attn_k_dsa_iswa>(std::move(inp_dsa), std::move(inp_swa), mctx_cur);
+
+    return (llm_graph_input_attn_k_dsa_iswa *) res->add_input(std::move(inp));
+}
+
 llm_graph_input_attn_kv_msa * llm_graph_context::build_attn_inp_kv_msa(bool msa_enabled) const {
    const auto * mctx_cur = static_cast<const llama_kv_cache_msa_context *>(mctx);

@@ -23,6 +23,7 @@ struct llama_memory_context_i;

 class llama_kv_cache_context;
 class llama_kv_cache_dsa_context;
+class llama_kv_cache_dsa_iswa_context;
 class llama_kv_cache_msa_context;
 class llama_kv_cache_dsv4_raw_context;
 class llama_kv_cache_dsv4_context;
@@ -374,6 +375,9 @@ public:

    bool can_reuse(const llm_graph_params & params) override;

+    // like can_reuse, but does not re-bind mctx
+    bool can_reuse_impl(const llm_graph_params & params);
+
    ggml_tensor * get_k_idxs() const { return self_k_idxs; }

    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
@@ -405,6 +409,9 @@ public:

    bool can_reuse(const llm_graph_params & params) override;

+    // like can_reuse, but does not re-bind mctx
+    bool can_reuse_impl(const llm_graph_params & params);
+
    ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }
    ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }

@@ -427,6 +434,32 @@ public:
    const llama_kv_cache_dsa_context * mctx;
 };

+// DSA input (full-attention layers + indexer) with K-only input for the SWA layers
+class llm_graph_input_attn_k_dsa_iswa : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_k_dsa_iswa(
+            std::unique_ptr<llm_graph_input_attn_k_dsa> inp_dsa,
+            std::unique_ptr<llm_graph_input_attn_k>     inp_swa,
+            const llama_kv_cache_dsa_iswa_context *     mctx) :
+        inp_dsa(std::move(inp_dsa)),
+        inp_swa(std::move(inp_swa)),
+        mctx(mctx) {
+    }
+    ~llm_graph_input_attn_k_dsa_iswa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    llm_graph_input_attn_k_dsa * get_dsa() const { return inp_dsa.get(); }
+    llm_graph_input_attn_k     * get_swa() const { return inp_swa.get(); }
+
+    std::unique_ptr<llm_graph_input_attn_k_dsa> inp_dsa;
+    std::unique_ptr<llm_graph_input_attn_k>     inp_swa;
+
+    const llama_kv_cache_dsa_iswa_context * mctx;
+};
+
 // standard K/V attention input against the base cache, plus destination indices for the indexer key cache
 class llm_graph_input_attn_kv_msa : public llm_graph_input_attn_kv {
 public:
@@ -1191,6 +1224,8 @@ struct llm_graph_context {

    llm_graph_input_attn_k_dsa * build_attn_inp_k_dsa() const;

+    llm_graph_input_attn_k_dsa_iswa * build_attn_inp_k_dsa_iswa() const;
+
    llm_graph_input_attn_kv_msa * build_attn_inp_kv_msa(bool msa_enabled) const;

    ggml_tensor * build_attn(
@@ -101,6 +101,11 @@ struct llama_hparams {
    uint32_t n_group_used       = 0;
    uint32_t n_group_experts    = 0;

+    // MLA + SWA (i.e. dots3note)
+    uint32_t n_lora_kv_swa           = 0;
+    uint32_t n_embd_head_k_mla_swa   = 0;
+    uint32_t n_embd_head_v_mla_swa   = 0;
+
    float    expert_group_scale   = 0.05f;
    float    expert_weights_scale = 0.0f;
    bool     expert_weights_norm  = false;
@@ -0,0 +1,341 @@
+#include "llama-kv-cache-dsa-iswa.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <cassert>
+
+//
+// llama_kv_cache_dsa_iswa
+//
+
+llama_kv_cache_dsa_iswa::llama_kv_cache_dsa_iswa(
+        const llama_model & model,
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   offload,
+                     bool   swa_full,
+                     bool   unified,
+                 uint32_t   kv_size,
+                 uint32_t   n_seq_max,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad,
+    const layer_filter_cb & filter_mla,
+    const layer_filter_cb & filter_lid,
+    const  layer_reuse_cb & reuse) : unified(unified) {
+
+    const auto & hparams = model.hparams;
+
+    // chain filters
+    const layer_filter_cb filter_dsa = [&](int32_t il) {
+        if (filter_mla && !filter_mla(il)) {
+            return false;
+        }
+
+        return !hparams.is_swa(il);
+    };
+
+    const layer_filter_cb filter_swa = [&](int32_t il) {
+        if (filter_mla && !filter_mla(il)) {
+            return false;
+        }
+
+        return hparams.is_swa(il);
+    };
+
+    const uint32_t size_dsa = kv_size;
+
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_dsa, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
+
+    // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
+    if (swa_full) {
+        LLAMA_LOG_WARN("%s: using full-size SWA cache (ref: %s)\n",
+                __func__, "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
+
+        size_swa = size_dsa;
+    }
+
+    LLAMA_LOG_INFO("%s: creating DSA KV cache, size = %u cells\n", __func__, size_dsa);
+
+    kv_dsa = std::make_unique<llama_kv_cache_dsa>(
+            model, type_k, type_v,
+            v_trans, offload, unified, size_dsa, n_seq_max, n_pad,
+            0, LLAMA_SWA_TYPE_NONE, filter_dsa, filter_lid, reuse);
+
+    LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
+
+    kv_swa = std::make_unique<llama_kv_cache>(
+            model, hparams, type_k, type_v,
+            v_trans, offload, unified, size_swa, n_seq_max, n_pad,
+            hparams.n_swa, hparams.swa_type, nullptr, filter_swa, reuse, nullptr);
+}
+
+void llama_kv_cache_dsa_iswa::clear(bool data) {
+    kv_dsa->clear(data);
+    kv_swa->clear(data);
+}
+
+bool llama_kv_cache_dsa_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    bool res = true;
+
+    res = res & kv_dsa->seq_rm(seq_id, p0, p1);
+    res = res & kv_swa->seq_rm(seq_id, p0, p1);
+
+    return res;
+}
+
+void llama_kv_cache_dsa_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    kv_dsa->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    kv_swa->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_dsa_iswa::seq_keep(llama_seq_id seq_id) {
+    kv_dsa->seq_keep(seq_id);
+    kv_swa->seq_keep(seq_id);
+}
+
+void llama_kv_cache_dsa_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    kv_dsa->seq_add(seq_id, p0, p1, shift);
+    kv_swa->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_kv_cache_dsa_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    kv_dsa->seq_div(seq_id, p0, p1, d);
+    kv_swa->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_kv_cache_dsa_iswa::seq_pos_min(llama_seq_id seq_id) const {
+    // the DSA cache is a superset of the SWA cache, so we can just check the SWA cache
+    return kv_swa->seq_pos_min(seq_id);
+}
+
+llama_pos llama_kv_cache_dsa_iswa::seq_pos_max(llama_seq_id seq_id) const {
+    return kv_swa->seq_pos_max(seq_id);
+}
+
+std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_dsa_iswa::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = kv_dsa->memory_breakdown();
+    for (const auto & buft_size : kv_swa->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+llama_memory_context_ptr llama_kv_cache_dsa_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
+
+    // first try simple split
+    do {
+        if (!unified) {
+            // requires equal splits, so we skip the simple split
+            break;
+        }
+
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_simple(n_ubatch);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_mla = kv_dsa->get_mla()->prepare(ubatches);
+        if (sinfos_mla.empty()) {
+            break;
+        }
+
+        auto sinfos_lid = kv_dsa->get_lid()->prepare(ubatches);
+        if (sinfos_lid.empty()) {
+            break;
+        }
+
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_mla.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_dsa_iswa_context>(
+                this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(sinfos_swa), std::move(ubatches));
+    } while (false);
+
+    // if it fails, try equal split
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_equal(n_ubatch, !unified, 0);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_mla = kv_dsa->get_mla()->prepare(ubatches);
+        if (sinfos_mla.empty()) {
+            break;
+        }
+
+        auto sinfos_lid = kv_dsa->get_lid()->prepare(ubatches);
+        if (sinfos_lid.empty()) {
+            break;
+        }
+
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_mla.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_dsa_iswa_context>(
+                this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(sinfos_swa), std::move(ubatches));
+    } while (false);
+
+    return std::make_unique<llama_kv_cache_dsa_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_kv_cache_dsa_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_dsa_iswa_context>(this);
+}
+
+llama_memory_context_ptr llama_kv_cache_dsa_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_dsa_iswa_context>(this, lctx, optimize);
+}
+
+bool llama_kv_cache_dsa_iswa::get_can_shift() const {
+    return kv_dsa->get_can_shift() &&
+           kv_swa->get_can_shift() &&
+           kv_dsa->get_mla()->get_size() == kv_swa->get_size();
+}
+
+void llama_kv_cache_dsa_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
+        kv_dsa->state_write(io, seq_id, flags);
+    }
+
+    kv_swa->state_write(io, seq_id, flags);
+}
+
+void llama_kv_cache_dsa_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
+        kv_dsa->state_read(io, seq_id, flags);
+    }
+
+    kv_swa->state_read(io, seq_id, flags);
+}
+
+llama_kv_cache_dsa * llama_kv_cache_dsa_iswa::get_dsa() const {
+    return kv_dsa.get();
+}
+
+llama_kv_cache * llama_kv_cache_dsa_iswa::get_swa() const {
+    return kv_swa.get();
+}
+
+//
+// llama_kv_cache_dsa_iswa_context
+//
+
+llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(llama_memory_status status) : status(status) {}
+
+llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(
+        llama_kv_cache_dsa_iswa * kv) :
+    ctx_dsa(kv->get_dsa()->init_full()),
+    ctx_swa(kv->get_swa()->init_full()),
+    status(llama_memory_status_combine(ctx_dsa->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(
+        llama_kv_cache_dsa_iswa * kv,
+        llama_context * lctx,
+        bool optimize) :
+    ctx_dsa(kv->get_dsa()->init_update(lctx, optimize)),
+    ctx_swa(kv->get_swa()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_dsa->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(
+        llama_kv_cache_dsa_iswa * kv,
+        slot_info_vec_t sinfos_mla,
+        slot_info_vec_t sinfos_lid,
+        slot_info_vec_t sinfos_swa,
+        std::vector<llama_ubatch> ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_dsa(new llama_kv_cache_dsa_context(kv->get_dsa(), std::move(sinfos_mla), std::move(sinfos_lid), this->ubatches)),
+    ctx_swa(new llama_kv_cache_context(kv->get_swa(), std::move(sinfos_swa), this->ubatches)),
+    status(llama_memory_status_combine(ctx_dsa->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_dsa_iswa_context:: ~llama_kv_cache_dsa_iswa_context() = default;
+
+bool llama_kv_cache_dsa_iswa_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_dsa->next();
+    ctx_swa->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_kv_cache_dsa_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_dsa->apply();
+    res = res & ctx_swa->apply();
+
+    return res;
+}
+
+llama_memory_status llama_kv_cache_dsa_iswa_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_kv_cache_dsa_iswa_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_dsa_context * llama_kv_cache_dsa_iswa_context::get_dsa() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_dsa_context *>(ctx_dsa.get());
+}
+
+const llama_kv_cache_context * llama_kv_cache_dsa_iswa_context::get_swa() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_context *>(ctx_swa.get());
+}
@@ -0,0 +1,134 @@
+#pragma once
+
+#include "llama-kv-cache-dsa.h"
+
+#include <vector>
+
+//
+// llama_kv_cache_dsa_iswa
+//
+
+// utilizes two child memories: llama_kv_cache_dsa for the full-attention (DSA) layers and llama_kv_cache for the SWA layers
+
+class llama_kv_cache_dsa_iswa : public llama_memory_i {
+public:
+    llama_kv_cache_dsa_iswa(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
+                         bool   swa_full,
+                         bool   unified,
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max,
+                     uint32_t   n_ubatch,
+                     uint32_t   n_pad,
+        const layer_filter_cb & filter_mla,
+        const layer_filter_cb & filter_lid,
+        const  layer_reuse_cb & reuse);
+
+    ~llama_kv_cache_dsa_iswa() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
+
+    //
+    // llama_kv_cache_dsa_iswa specific API
+    //
+
+    llama_kv_cache_dsa * get_dsa() const;
+    llama_kv_cache     * get_swa() const;
+
+private:
+    const bool unified;
+
+    std::unique_ptr<llama_kv_cache_dsa> kv_dsa;
+    std::unique_ptr<llama_kv_cache>     kv_swa;
+};
+
+class llama_kv_cache_dsa_iswa_context : public llama_memory_context_i {
+public:
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
+
+    // used for errors
+    llama_kv_cache_dsa_iswa_context(llama_memory_status status);
+
+    // used to create a full-cache context
+    llama_kv_cache_dsa_iswa_context(
+            llama_kv_cache_dsa_iswa * kv);
+
+    // used to create an update context
+    llama_kv_cache_dsa_iswa_context(
+            llama_kv_cache_dsa_iswa * kv,
+            llama_context * lctx,
+            bool optimize);
+
+    // used to create a batch processing context from a batch
+    llama_kv_cache_dsa_iswa_context(
+            llama_kv_cache_dsa_iswa * kv,
+            slot_info_vec_t sinfos_mla,
+            slot_info_vec_t sinfos_lid,
+            slot_info_vec_t sinfos_swa,
+            std::vector<llama_ubatch> ubatches);
+
+    virtual ~llama_kv_cache_dsa_iswa_context();
+
+    //
+    // llama_memory_context_i
+    //
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_kv_cache_dsa_iswa_context specific API
+    //
+
+    const llama_kv_cache_dsa_context * get_dsa() const;
+    const llama_kv_cache_context     * get_swa() const;
+
+private:
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    const llama_memory_context_ptr ctx_dsa;
+    const llama_memory_context_ptr ctx_swa;
+
+    const llama_memory_status status;
+};
@@ -323,7 +323,8 @@ llama_kv_cache::llama_kv_cache(
            hparams.n_embd_head_k() % 64 == 0;

        // always create Hadamard rotation tensors for DeepSeek lightning indexers
-        if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 || model.arch == LLM_ARCH_GLM_DSA) &&
+        if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 ||
+                model.arch == LLM_ARCH_GLM_DSA || model.arch == LLM_ARCH_DOTS3NOTE) &&
                hparams.n_embd_head_k_full == hparams.indexer_head_size) {
            attn_rot_k = true;
        }
@@ -31,6 +31,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
        case LLM_ARCH_MELLUM:
        case LLM_ARCH_LAGUNA:
        case LLM_ARCH_GRANITE_SWA:
+        case LLM_ARCH_DOTS3NOTE: // TODO: need to handle SWA pattern and MLA+SWA config
            return false;
        default:
            return true;
@@ -11,6 +11,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-kv-cache-dsa.h"
+#include "llama-kv-cache-dsa-iswa.h"
 #include "llama-kv-cache-msa.h"
 #include "llama-kv-cache-dsv4.h"
 #include "llama-memory-hybrid.h"
@@ -194,6 +195,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_deepseek2ocr(params);
        case LLM_ARCH_DEEPSEEK32:
            return new llama_model_deepseek32(params);
+        case LLM_ARCH_DOTS3NOTE:
+            return new llama_model_dots3note(params);
        case LLM_ARCH_DEEPSEEK4:
            return new llama_model_deepseek4(params);
        case LLM_ARCH_GLM_DSA:
@@ -487,6 +490,10 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "ssm_out.weight");
        }
        if (std::regex_match(tensor_name, pattern_r_cache) || std::regex_match(tensor_name, pattern_s_cache)) {
+            if (ud->model->arch == LLM_ARCH_LFM2 || ud->model->arch == LLM_ARCH_LFM2MOE) {
+                // the LFM2 shortconv block runs fully mirrored, so its conv state must be mirrored too
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED, "");
+            }
            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "ssm_out.weight");
        }
        if (std::regex_match(tensor_name, pattern_ssm_conv1d)) {
@@ -847,6 +854,7 @@ const char * llm_type_name(llm_type type) {
        case LLM_TYPE_230B_A10B:     return "230B.A10B";
        case LLM_TYPE_428B_A23B:     return "428B.A23B";
        case LLM_TYPE_235B_A22B:     return "235B.A22B";
+        case LLM_TYPE_288B_A19B:     return "288B.A19B";
        case LLM_TYPE_300B_A47B:     return "300B.A47B";
        case LLM_TYPE_310B_A15B:     return "310B.A15B";
        case LLM_TYPE_355B_A32B:     return "355B.A32B";
@@ -1920,7 +1928,9 @@ void llama_model::print_info() const {
            LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
        }

-        if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR || arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
+        if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR ||
+                arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA ||
+                arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4) {
            LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
            LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
            LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
@@ -2189,6 +2199,57 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                            nullptr);
                }
            } break;
+        case LLM_ARCH_DOTS3NOTE:
+            {
+                GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
+
+                if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && hparams.n_layer_nextn > 0) {
+                    // MTP draft context: plain attention KV cache holding only the nextn layer
+                    llama_kv_cache::layer_filter_cb filter =
+                        [&](uint32_t il) { return il >= hparams.n_layer(); };
+
+                    res = new llama_kv_cache(
+                            *this,
+                            hparams,
+                            params.type_k,
+                            params.type_v,
+                            !cparams.flash_attn,
+                            cparams.offload_kqv,
+                            cparams.kv_unified,
+                            cparams.n_ctx_seq,
+                            cparams.n_seq_max,
+                            1,
+                            hparams.n_swa,
+                            hparams.swa_type,
+                            nullptr,
+                            filter,
+                            nullptr,
+                            nullptr);
+                } else {
+                    // main context: DSA cache for the trunk full-attention layers plus a window-sized SWA cache
+                    llama_kv_cache::layer_filter_cb filter_mla = nullptr;
+                    if (hparams.n_layer_nextn > 0) {
+                        filter_mla = [&](uint32_t il) { return il < hparams.n_layer(); };
+                    }
+                    llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return il < hparams.n_layer() && hparams.is_indexer_full(il); };
+
+                    res = new llama_kv_cache_dsa_iswa(
+                            *this,
+                            params.type_k,
+                            params.type_v,
+                            !cparams.flash_attn,
+                            cparams.offload_kqv,
+                            params.swa_full,
+                            cparams.kv_unified,
+                            cparams.n_ctx_seq,
+                            cparams.n_seq_max,
+                            cparams.n_ubatch,
+                            1,
+                            filter_mla,
+                            filter_lid,
+                            nullptr);
+                }
+            } break;
        case LLM_ARCH_DEEPSEEK4:
            {
                GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
@@ -2657,6 +2718,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_LLAMA_EMBED:
        case LLM_ARCH_MAINCODER:
        case LLM_ARCH_GLM_DSA:
+        case LLM_ARCH_DOTS3NOTE:
        case LLM_ARCH_NANBEIGE:
        case LLM_ARCH_POCKETTTS:
            return LLAMA_ROPE_TYPE_NORM;
@@ -140,6 +140,7 @@ enum llm_type {
    LLM_TYPE_230B_A10B, // Minimax M2
    LLM_TYPE_428B_A23B, // Minimax M3
    LLM_TYPE_235B_A22B,
+    LLM_TYPE_288B_A19B, // dots3-note
    LLM_TYPE_300B_A47B, // Ernie MoE big
    LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
    LLM_TYPE_355B_A32B, // GLM-4.5
@@ -524,17 +524,9 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
                cb(q, "q", il);
            }
-            // split into {n_embd_head_qk_nope, n_head, n_tokens}
-            ggml_tensor * q_nope =
-                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
-                             ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
-            cb(q_nope, "q_nope", il);
-
-            // and {n_embd_head_qk_rope, n_head, n_tokens}
-            ggml_tensor * q_pe = ggml_view_3d(
-                ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
-                ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
-            cb(q_pe, "q_pe", il);
+            // {n_embd_head_k, n_head, n_tokens}
+            q = ggml_reshape_3d(ctx0, q, n_embd_head_k, n_head, n_tokens);
+            cb(q, "q", il);

            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
@@ -552,10 +544,6 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
            cb(k_pe, "k_pe", il);

-            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                 ext_factor, attn_factor, beta_fast, beta_slow);
-            cb(q_pe, "q_pe", il);
-
            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                                 ext_factor, attn_factor, beta_fast, beta_slow);
            cb(k_pe, "k_pe", il);
@@ -564,6 +552,20 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
            cb(kv_cmpr, "kv_cmpr", il);

            if (is_mla) {
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                                                    q->nb[1], q->nb[2], 0);
+                cb(q_nope, "q_nope", il);
+
+                // and {n_embd_head_qk_rope, n_head, n_tokens}
+                ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                                                  q->nb[1], q->nb[2], ggml_row_size(q->type, n_embd_head_qk_nope));
+                cb(q_pe, "q_pe", il);
+
+                q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(q_pe, "q_pe", il);
+
                // {n_embd_head_qk_nope, n_tokens, n_head}
                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
                cb(q_nope, "q_nope_perm", il);
@@ -623,10 +625,14 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
                Vcur = ggml_cont(ctx0, Vcur);
                cb(Vcur, "Vcur_cont", il);

-                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);
+                // RoPE is applied to the trailing dims only
+                ggml_tensor * Qcur = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                                                   freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+                Qcur = ggml_rope_set_offset(Qcur, n_embd_head_qk_nope);
                cb(Qcur, "Qcur", il);

-                ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_nope,
+                        ggml_repeat_4d(ctx0, k_pe, n_embd_head_qk_rope, n_head, n_tokens, 1), 0);
                cb(Kcur, "Kcur", il);

                if (inp_attn_scale) {
@@ -501,21 +501,10 @@ ggml_tensor * llama_model_deepseek4::graph::build_hca_compressed_kv_from_state(
    comp = build_norm(comp, norm, nullptr, LLM_NORM_RMS, il);
    cb(comp, name, il);

-    ggml_tensor * comp_nope = ggml_view_3d(ctx0, comp, n_embd_head_nope, 1, n_blocks,
-            ggml_row_size(comp->type, n_embd_head),
-            ggml_row_size(comp->type, n_embd_head),
-            0);
-    ggml_tensor * comp_pe = ggml_view_3d(ctx0, comp, n_embd_head_rope, 1, n_blocks,
-            ggml_row_size(comp->type, n_embd_head),
-            ggml_row_size(comp->type, n_embd_head),
-            ggml_row_size(comp->type, n_embd_head_nope));
-
-    comp_pe = ggml_rope_ext(ctx0, comp_pe, comp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig,
+    comp = ggml_rope_ext(ctx0, comp, comp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig,
            hparams.dsv4_compress_rope_base, freq_scale, ext_factor,
            dsv4_rope_attn_factor(freq_scale, ext_factor), beta_fast, beta_slow);
-    cb(comp_pe, name, il);
-
-    comp = ggml_concat(ctx0, comp_nope, comp_pe, 0);
+    comp = ggml_rope_set_offset(comp, n_embd_head_nope);
    cb(comp, name, il);

    return comp;
@@ -585,21 +574,10 @@ ggml_tensor * llama_model_deepseek4::graph::build_overlap_compressed_kv_from_sta
    comp = build_norm(comp, norm, nullptr, LLM_NORM_RMS, il);
    cb(comp, name, il);

-    ggml_tensor * comp_nope = ggml_view_3d(ctx0, comp, n_embd_head_nope, 1, n_blocks,
-            ggml_row_size(comp->type, n_embd_head),
-            ggml_row_size(comp->type, n_embd_head),
-            0);
-    ggml_tensor * comp_pe = ggml_view_3d(ctx0, comp, n_embd_head_rope, 1, n_blocks,
-            ggml_row_size(comp->type, n_embd_head),
-            ggml_row_size(comp->type, n_embd_head),
-            ggml_row_size(comp->type, n_embd_head_nope));
-
-    comp_pe = ggml_rope_ext(ctx0, comp_pe, comp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig,
+    comp = ggml_rope_ext(ctx0, comp, comp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig,
            hparams.dsv4_compress_rope_base, freq_scale, ext_factor,
            dsv4_rope_attn_factor(freq_scale, ext_factor), beta_fast, beta_slow);
-    cb(comp_pe, name, il);
-
-    comp = ggml_concat(ctx0, comp_nope, comp_pe, 0);
+    comp = ggml_rope_set_offset(comp, n_embd_head_nope);
    cb(comp, name, il);

    return comp;
@@ -628,21 +606,12 @@ ggml_tensor * llama_model_deepseek4::graph::build_lid_top_k(
    indexer_q = ggml_reshape_3d(ctx0, indexer_q, n_embd_indexer_head, n_indexer_head, nt);
    cb(indexer_q, "lid_q", il);

-    ggml_tensor * indexer_q_nope = ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_nope, n_indexer_head, nt,
-            ggml_row_size(indexer_q->type, n_embd_indexer_head),
-            ggml_row_size(indexer_q->type, n_embd_indexer_head)*n_indexer_head,
-            0);
-    ggml_tensor * indexer_q_pe = ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_rope, n_indexer_head, nt,
-            ggml_row_size(indexer_q->type, n_embd_indexer_head),
-            ggml_row_size(indexer_q->type, n_embd_indexer_head)*n_indexer_head,
-            ggml_row_size(indexer_q->type, n_embd_indexer_head_nope));
-
-    indexer_q_pe = ggml_rope_ext(ctx0, indexer_q_pe, inp_pos, nullptr, n_embd_indexer_head_rope,
+    indexer_q = ggml_rope_ext(ctx0, indexer_q, inp_pos, nullptr, n_embd_indexer_head_rope,
            rope_type, n_ctx_orig, hparams.dsv4_compress_rope_base, freq_scale,
            ext_factor, dsv4_rope_attn_factor(freq_scale, ext_factor), beta_fast, beta_slow);
-    cb(indexer_q_pe, "lid_q_pe", il);
+    indexer_q = ggml_rope_set_offset(indexer_q, n_embd_indexer_head_nope);
+    cb(indexer_q, "lid_q_rope", il);

-    indexer_q = ggml_concat(ctx0, indexer_q_nope, indexer_q_pe, 0);
    indexer_q = llama_mul_mat_hadamard(ctx0, indexer_q, inp_lid.k_rot);
    cb(indexer_q, "lid_q_rot", il);

@@ -945,18 +914,9 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention_impl(
    q = ggml_rms_norm(ctx0, q, norm_rms_eps);
    cb(q, "q_norm", il);

-    ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_nope, n_head, nt,
-            ggml_row_size(q->type, n_embd_head),
-            ggml_row_size(q->type, n_embd_head)*n_head,
-            0);
-    ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_rope, n_head, nt,
-            ggml_row_size(q->type, n_embd_head),
-            ggml_row_size(q->type, n_embd_head)*n_head,
-            ggml_row_size(q->type, n_embd_head_nope));
-    q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig_l,
+    q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig_l,
            freq_base_l, freq_scale_l, ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);
-    cb(q_pe, "q_pe", il);
-    q = ggml_concat(ctx0, q_nope, q_pe, 0);
+    q = ggml_rope_set_offset(q, n_embd_head_nope);
    cb(q, "q", il);

    ggml_tensor * kv = build_lora_mm(layer.wkv, cur);
@@ -964,18 +924,9 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention_impl(
    kv = ggml_reshape_3d(ctx0, kv, n_embd_head, 1, nt);
    cb(kv, "kv_norm", il);

-    ggml_tensor * kv_nope = ggml_view_3d(ctx0, kv, n_embd_head_nope, 1, nt,
-            ggml_row_size(kv->type, n_embd_head),
-            ggml_row_size(kv->type, n_embd_head),
-            0);
-    ggml_tensor * kv_pe = ggml_view_3d(ctx0, kv, n_embd_head_rope, 1, nt,
-            ggml_row_size(kv->type, n_embd_head),
-            ggml_row_size(kv->type, n_embd_head),
-            ggml_row_size(kv->type, n_embd_head_nope));
-    kv_pe = ggml_rope_ext(ctx0, kv_pe, inp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig_l,
+    kv = ggml_rope_ext(ctx0, kv, inp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig_l,
            freq_base_l, freq_scale_l, ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);
-    cb(kv_pe, "kv_pe", il);
-    kv = ggml_concat(ctx0, kv_nope, kv_pe, 0);
+    kv = ggml_rope_set_offset(kv, n_embd_head_nope);
    cb(kv, "kv", il);

    const int64_t ratio = hparams.dsv4_compress_ratios[il];
@@ -1225,7 +1176,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention_impl(
    if (inp_mtp) {
        out = build_attn(inp_mtp,
                nullptr, nullptr, nullptr,
-                q, kv, nullptr,
+                q, kv, kv,
                nullptr, layer.attn_sinks, nullptr,
                1.0f/sqrtf(float(n_embd_head)), il);
        cb(out, "attn_raw", il);
@@ -1245,17 +1196,9 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention_impl(
    }

    out = ggml_reshape_3d(ctx0, out, n_embd_head, n_head, nt);
-    ggml_tensor * out_nope = ggml_view_3d(ctx0, out, n_embd_head_nope, n_head, nt,
-            ggml_row_size(out->type, n_embd_head),
-            ggml_row_size(out->type, n_embd_head)*n_head,
-            0);
-    ggml_tensor * out_pe = ggml_view_3d(ctx0, out, n_embd_head_rope, n_head, nt,
-            ggml_row_size(out->type, n_embd_head),
-            ggml_row_size(out->type, n_embd_head)*n_head,
-            ggml_row_size(out->type, n_embd_head_nope));
-    out_pe = ggml_rope_ext_back(ctx0, out_pe, inp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig_l,
+    out = ggml_rope_ext_back(ctx0, out, inp_pos, nullptr, n_embd_head_rope, rope_type, n_ctx_orig_l,
            freq_base_l, freq_scale_l, ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);
-    out = ggml_concat(ctx0, out_nope, out_pe, 0);
+    out = ggml_rope_set_offset(out, n_embd_head_nope);
    cb(out, "attn_derope", il);

    out = ggml_reshape_3d(ctx0, out, o_group_dim, n_groups, nt);
@@ -591,17 +591,9 @@ llama_model_dflash::graph_dsv4::graph_dsv4(const llama_model & model, const llm_
            kv = build_norm(kv, layer.attn_kv_norm, nullptr, LLM_NORM_RMS, il);
            kv = ggml_reshape_3d(ctx0, kv, n_embd_head, 1, n_tokens);

-            ggml_tensor * kv_nope = ggml_view_3d(ctx0, kv, n_embd_head_nope, 1, n_tokens,
-                    ggml_row_size(kv->type, n_embd_head),
-                    ggml_row_size(kv->type, n_embd_head),
-                    0);
-            ggml_tensor * kv_pe = ggml_view_3d(ctx0, kv, n_embd_head_rope, 1, n_tokens,
-                    ggml_row_size(kv->type, n_embd_head),
-                    ggml_row_size(kv->type, n_embd_head),
-                    ggml_row_size(kv->type, n_embd_head_nope));
-            kv_pe = ggml_rope_ext(ctx0, kv_pe, inp_pos, nullptr, n_embd_head_rope, rope_type, 0,
+            kv = ggml_rope_ext(ctx0, kv, inp_pos, nullptr, n_embd_head_rope, rope_type, 0,
                    freq_base, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
-            kv = ggml_concat(ctx0, kv_nope, kv_pe, 0);
+            kv = ggml_rope_set_offset(kv, n_embd_head_nope);
            cb(kv, "kv_injected", il);

            if (inp_attn->self_k_rot_swa) {
@@ -0,0 +1,480 @@
+#include "models.h"
+
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-dsa.h"
+
+// note: code adapted from deepseek32.cpp (DSA indexer + absorbed MLA) and step35.cpp (head-wise output gate)
+
+void llama_model_dots3note::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    hparams.f_norm_eps = 1e-6;  // eps for the indexer k_norm layer norm
+
+    // TODO: use MTP layer
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");
+
+    // MoE parameters
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,  hparams.n_layer_dense_lead);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func);
+
+    // MLA parameters of the full-attention layers
+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,      hparams.n_lora_q);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl);
+
+    // MLA parameters of the sliding-window layers
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK_SWA,     hparams.n_lora_kv_swa);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,   hparams.n_embd_head_k_mla_swa);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, hparams.n_embd_head_v_mla_swa);
+
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa);
+    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
+
+    // DSA parameters
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,      hparams.indexer_top_k);
+    ml.get_arr(LLM_KV_ATTENTION_INDEXER_TYPES, hparams.is_indexer_full_impl);
+
+    switch (hparams.n_layer()) {
+        case 46: type = LLM_TYPE_288B_A19B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_dots3note::load_arch_tensors(llama_model_loader & ml) {
+    LLAMA_LOAD_LOCALS;
+    GGML_UNUSED(ml);
+
+    if (!hparams.is_mla()) {
+        throw std::runtime_error("DOTS3NOTE architecture requires MLA");
+    }
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+
+    const int64_t q_lora_rank     = hparams.n_lora_q;
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer_all; ++i) {
+        auto & layer = layers[i];
+
+        const bool is_mtp = i >= n_layer;
+        // the NextN/MTP block uses the sliding-attention geometry
+        const bool is_swa = is_mtp || hparams.is_swa(i);
+
+        // MTP tensors are preserved in the GGUF but there is no MTP graph yet
+        const int flags = is_mtp ? TENSOR_SKIP | TENSOR_NOT_REQUIRED : 0;
+
+        const int64_t n_head_l = hparams.n_head(i);
+
+        const int64_t kv_lora_rank       = is_swa ? hparams.n_lora_kv_swa         : hparams.n_lora_kv;
+        const int64_t n_embd_head_k_mla  = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();
+        const int64_t n_embd_head_v_mla  = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();
+        const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, flags);
+        layer.attn_q_a_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM,  "weight", i), {q_lora_rank}, flags);
+        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
+        // norm applied on the shared rope key before rope
+        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_qk_rope}, flags);
+
+        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
+        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head_l * n_embd_head_k_mla}, flags);
+
+        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
+
+        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head_l}, flags);
+        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head_l}, flags);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head_l * n_embd_head_v_mla, n_embd}, flags);
+
+        // head-wise sigmoid output gate
+        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, flags);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+        // DSA indexer
+        if (!is_mtp && hparams.is_indexer_full(i)) {
+            layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
+            layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
+            layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
+            layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
+            layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+        }
+
+        if (is_mtp || i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
+        } else {
+            if (n_expert == 0 || n_expert_used == 0) {
+                throw std::runtime_error("n_expert and n_expert_used must be > 0");
+            }
+
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, flags);
+
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+        }
+
+        if (is_mtp) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd }, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd }, flags);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), { n_embd, n_vocab }, flags);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_dots3note::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_dots3note::graph::graph(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    GGML_ASSERT(hparams.is_mla());
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+
+    const int64_t n_indexer_head      = hparams.indexer_n_head;
+    const int64_t n_embd_indexer_head = hparams.indexer_head_size;
+    const uint32_t n_indexer_top_k = hparams.indexer_top_k;
+
+    // the indexer head layout is [rope | nope]
+    GGML_ASSERT(hparams.n_rot() <= n_embd_indexer_head);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    llm_graph_input_attn_k_dsa_iswa * inp_attn = build_attn_inp_k_dsa_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        const bool is_swa = hparams.is_swa(il);
+
+        const int64_t n_head_l = hparams.n_head(il);
+
+        const int64_t kv_lora_rank        = is_swa ? hparams.n_lora_kv_swa         : hparams.n_lora_kv;
+        const int64_t n_embd_head_k_mla   = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();
+        const int64_t n_embd_head_v_mla   = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();
+        const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+        const float kq_scale    = 1.0f/sqrtf(float(n_embd_head_k_mla));
+        const float freq_base_l = model.get_rope_freq_base(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * attn_inp = cur;
+
+            ggml_tensor * qr = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+            cb(qr, "qr", il);
+
+            qr = build_norm(qr, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(qr, "qr", il);
+
+            ggml_tensor * top_k = nullptr;
+
+            // lightning indexer (full-attention layers only)
+            if (!is_swa) {
+                ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
+                cb(indexer_q, "indexer_q", il);
+
+                // {n_embd_indexer_head, n_indexer_head, n_tokens}
+                indexer_q = ggml_reshape_3d(ctx0, indexer_q, n_embd_indexer_head, n_indexer_head, n_tokens);
+                indexer_q = ggml_rope_ext(ctx0, indexer_q, inp_pos, nullptr, n_rot,
+                                     LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(indexer_q, "indexer_q", il);
+
+                ggml_tensor * indexer_k = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_k, cur);
+                cb(indexer_k, "indexer_k", il);
+
+                indexer_k = build_norm(indexer_k, model.layers[il].indexer_k_norm, model.layers[il].indexer_k_norm_b, LLM_NORM, il);
+                cb(indexer_k, "indexer_k", il);
+
+                // {n_embd_indexer_head, 1, n_tokens}
+                indexer_k = ggml_reshape_3d(ctx0, indexer_k, n_embd_indexer_head, 1, n_tokens);
+                indexer_k = ggml_rope_ext(ctx0, indexer_k, inp_pos, nullptr, n_rot,
+                                     LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(indexer_k, "indexer_k", il);
+
+                // perform Hadamard transform on indexer q and k
+                indexer_q = ggml_mul_mat(ctx0, inp_attn->get_dsa()->self_k_rot_lid, indexer_q);
+                cb(indexer_q, "indexer_q", il);
+                indexer_k = ggml_mul_mat(ctx0, inp_attn->get_dsa()->self_k_rot_lid, indexer_k);
+                cb(indexer_k, "indexer_k", il);
+
+                // store indexer keys to KV cache
+                const auto * mctx_lid = inp_attn->get_dsa()->mctx->get_lid();
+                const auto & k_idxs_lid = inp_attn->get_dsa()->get_k_idxs_lid();
+                ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, indexer_k, k_idxs_lid, il));
+
+                ggml_tensor * indexer_weights = ggml_mul_mat(ctx0, model.layers[il].indexer_proj, cur);
+                cb(indexer_weights, "indexer_weights", il);
+
+                indexer_k = mctx_lid->get_k(ctx0, il);
+
+                // split the batch into streams if needed
+                const auto n_stream = indexer_k->ne[3];
+                indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream, indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
+                indexer_weights = ggml_view_4d(ctx0, indexer_weights, indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream, indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
+
+                // pre-scale weights to avoid scaling operations on huge indexer_score tensor
+                indexer_weights = ggml_scale(ctx0, indexer_weights, 1.0f / sqrtf(float(n_embd_indexer_head * n_indexer_head)));
+                cb(indexer_weights, "indexer_weights", il);
+
+                ggml_tensor * indexer_score = nullptr;
+                if (cparams.fused_lid) {
+                    indexer_score = ggml_lightning_indexer(ctx0, indexer_q, indexer_k, indexer_weights, inp_attn->get_dsa()->get_kq_mask_lid());
+                    cb(indexer_score, "indexer_score", il);
+                    res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, indexer_score, il});
+                } else {
+                    indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
+                    cb(indexer_q, "indexer_q", il);
+                    indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
+                    cb(indexer_k, "indexer_k", il);
+
+                    ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
+                    cb(indexer_kq, "indexer_kq", il);
+
+                    // ReLU requires contiguous tensors
+                    indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
+                    cb(indexer_kq, "indexer_kq", il);
+
+                    indexer_score = ggml_relu(ctx0, indexer_kq);
+                    cb(indexer_score, "indexer_score", il);
+
+                    indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // sum by q n_indexer_head dimension
+                    indexer_score = ggml_sum_rows(ctx0, indexer_score);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // permute result to match KQ mask
+                    indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
+                    cb(indexer_score, "indexer_score", il);
+
+                    ggml_tensor * indexer_kq_mask = inp_attn->get_dsa()->get_kq_mask_lid();
+                    indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
+                    cb(indexer_score, "indexer_score", il);
+                }
+
+                // get indices of top k indexer scores
+                uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k;
+                top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k));
+                cb(top_k, "top_k", il);
+            }
+
+            ggml_tensor * q = ggml_mul_mat(ctx0, model.layers[il].wq_b, qr);
+            cb(q, "q", il);
+
+            // split into {n_embd_head_qk_nope, n_head_l, n_tokens}
+            ggml_tensor * q_nope =
+                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),
+                             ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, 0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_embd_head_qk_rope, n_head_l, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(
+                ctx0, q, n_embd_head_qk_rope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),
+                ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_cmpr =
+                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // and {n_embd_head_qk_rope, 1, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            // norm on the shared rope key, applied before rope
+            k_pe = build_norm(k_pe, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
+            cb(k_pe, "k_pe", il);
+
+            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(q_pe, "q_pe", il);
+
+            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(k_pe, "k_pe", il);
+
+            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // MLA attention with the absorption optimization
+            {
+                // {n_embd_head_qk_nope, n_tokens, n_head_l}
+                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                cb(q_nope, "q_nope_perm", il);
+
+                // {n_embd_head_qk_nope, kv_lora_rank, n_head_l} x {n_embd_head_qk_nope, n_tokens, n_head_l}
+                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                // {kv_lora_rank, n_head_l, n_tokens}
+                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, n_head_l, n_tokens}
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+                cb(Qcur, "Qcur", il);
+
+                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+                cb(Kcur, "Kcur", il);
+
+                // {kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Vcur = kv_cmpr;
+                cb(Vcur, "Vcur", il);
+
+                // apply the head-wise output gate before o_proj, so wo stays out of build_attn
+                if (is_swa) {
+                    cur = build_attn(inp_attn->get_swa(),
+                            nullptr, nullptr, nullptr,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
+                } else {
+                    cur = build_attn(inp_attn->get_dsa(),
+                            nullptr, nullptr, nullptr,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
+                }
+                cb(cur, "attn_out", il);
+
+                ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+                cb(gate, "attn_gate", il);
+
+                gate = ggml_sigmoid(ctx0, gate);
+                cb(gate, "attn_gate_sigmoid", il);
+
+                // broadcast the per-head gate over the head dimension
+                ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, cur, n_embd_head_v_mla, n_head_l, n_tokens);
+                ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate,                1, n_head_l, n_tokens);
+                attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
+                cb(attn_3d, "attn_gated", il);
+
+                cur = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v_mla * n_head_l, n_tokens);
+
+                cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
+                cb(cur, "attn_output", il);
+            }
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+                model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il,
+                nullptr,
+                model.layers[il].ffn_gate_up_exps,
+                model.layers[il].ffn_up_exps_s,
+                model.layers[il].ffn_gate_exps_s,
+                model.layers[il].ffn_down_exps_s);
+            cb(moe_out, "ffn_moe_out", il);
+
+            ggml_tensor * ffn_shexp =
+                build_ffn(cur,
+                    model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,
+                    model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
+                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
@@ -2,6 +2,8 @@
 #include "../llama-memory-hybrid-iswa.h"
 #include "../llama-memory-hybrid.h"

+#include <algorithm>
+
 void llama_model_lfm2::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -202,15 +204,20 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
        }
        GGML_ASSERT(bx->ne[0] > conv->ne[0]);

-        // last d_conv columns is a new conv state
-        auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
-                                       (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
-        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
+        // write conv states: slot 0 = the final state, slot s = the state s tokens back (partial rollback)
+        const int64_t K         = hparams.causal_attn && cparams.n_rs_seq > 0 ? (int64_t) cparams.n_rs_seq + 1 : 1;
+        const int64_t n_written = std::min<int64_t>(n_seq_tokens, K);
+        const auto    mem_size  = mctx_cur->get_size();
+        const size_t  row_size  = ggml_row_size(conv_state->type, (int64_t) d_conv * n_embd);

-        // write new conv conv state
-        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
-                                               ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
-                                                            kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
+        for (int64_t slot = 0; slot < n_written; ++slot) {
+            auto * conv_snap = ggml_view_3d(ctx0, bx, d_conv, bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
+                                            (bx->ne[0] - d_conv - slot) * ggml_element_size(bx));
+            ggml_build_forward_expand(gf, ggml_cpy(ctx0, conv_snap,
+                                                   ggml_view_2d(ctx0, conv_state, (int64_t) d_conv * n_embd, n_seqs,
+                                                                conv_state->nb[1],
+                                                                ((size_t) slot * mem_size + kv_head) * row_size)));
+        }

        auto * conv_kernel = model.layers[il].shortconv.conv;
        auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);
@@ -242,6 +249,8 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = cur;
+
        const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);

        auto * prev_cur = cur;
@@ -115,19 +115,9 @@ llama_model_minicpm3::graph::graph(const llama_model & model, const llm_graph_pa
            q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
            cb(q, "q", il);

-            // split into {n_head * n_embd_head_qk_nope, n_tokens}
-            ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k()),
-                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
-                    0);
-            cb(q_nope, "q_nope", il);
-
-            // and {n_head * n_embd_head_qk_rope, n_tokens}
-            ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k()),
-                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
-                    ggml_row_size(q->type, n_embd_head_qk_nope));
-            cb(q_pe, "q_pe", il);
+            // {n_embd_head_k, n_head, n_tokens}, RoPE is applied to the trailing dims only
+            q = ggml_reshape_3d(ctx0, q, hparams.n_embd_head_k(), n_head, n_tokens);
+            cb(q, "q", il);

            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
@@ -172,12 +162,13 @@ llama_model_minicpm3::graph::graph(const llama_model & model, const llm_graph_pa
            v_states = ggml_cont(ctx0, v_states);
            cb(v_states, "v_states", il);

-            q_pe = ggml_rope_ext(
-                    ctx0, q_pe, inp_pos, rope_factors,
+            q = ggml_rope_ext(
+                    ctx0, q, inp_pos, rope_factors,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                    );
-            cb(q_pe, "q_pe", il);
+            q = ggml_rope_set_offset(q, n_embd_head_qk_nope);
+            cb(q, "q_rope", il);

            // shared RoPE key
            k_pe = ggml_rope_ext(
@@ -187,10 +178,11 @@ llama_model_minicpm3::graph::graph(const llama_model & model, const llm_graph_pa
                    );
            cb(k_pe, "k_pe", il);

-            ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+            ggml_tensor * q_states = q;
            cb(q_states, "q_states", il);

-            ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+            ggml_tensor * k_states = ggml_concat(ctx0, k_nope,
+                    ggml_repeat_4d(ctx0, k_pe, n_embd_head_qk_rope, n_head, n_tokens, 1), 0);
            cb(k_states, "k_states", il);

            cur = build_attn(inp_attn,
@@ -1156,6 +1156,18 @@ struct llama_model_deepseek32 : public llama_model_base {
 };


+struct llama_model_dots3note : public llama_model_base {
+    llama_model_dots3note(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
 struct llama_model_deepseek4 : public llama_model_base {
    llama_model_deepseek4(const struct llama_model_params & params) : llama_model_base(params) {}
    void load_arch_hparams(llama_model_loader & ml) override;
@@ -81,19 +81,9 @@ llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params
            q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
            cb(q, "q", il);

-            // split into {n_head * n_embd_head_qk_nope, n_tokens}
-            ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k()),
-                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
-                    0);
-            cb(q_nope, "q_nope", il);
-
-            // and {n_head * n_embd_head_qk_rope, n_tokens}
-            ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k()),
-                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
-                    ggml_row_size(q->type, n_embd_head_qk_nope));
-            cb(q_pe, "q_pe", il);
+            // {n_embd_head_k, n_head, n_tokens}, RoPE is applied to the trailing dims only
+            q = ggml_reshape_3d(ctx0, q, hparams.n_embd_head_k(), n_head, n_tokens);
+            cb(q, "q", il);

            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
@@ -143,12 +133,13 @@ llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params
                    0);
            cb(v_states, "v_states", il);

-            q_pe = ggml_rope_ext(
-                    ctx0, q_pe, inp_pos, nullptr,
+            q = ggml_rope_ext(
+                    ctx0, q, inp_pos, nullptr,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                    );
-            cb(q_pe, "q_pe", il);
+            q = ggml_rope_set_offset(q, n_embd_head_qk_nope);
+            cb(q, "q_rope", il);

            // shared RoPE key
            k_pe = ggml_rope_ext(
@@ -158,10 +149,11 @@ llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params
                    );
            cb(k_pe, "k_pe", il);

-            ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+            ggml_tensor * q_states = q;
            cb(q_states, "q_states", il);

-            ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+            ggml_tensor * k_states = ggml_concat(ctx0, k_nope,
+                    ggml_repeat_4d(ctx0, k_pe, n_embd_head_qk_rope, n_head, n_tokens, 1), 0);
            cb(k_states, "k_states", il);

            cur = build_attn(inp_attn,
@@ -7085,9 +7085,10 @@ struct test_flash_attn_ext : public test_case {
    const ggml_type type_V;
    std::array<int32_t, 4> permute;
    const bool kv_view; // create K/V as views of a larger buffer (like a KV cache)
+    const bool v_is_view_of_k;

    std::string vars() override {
-        return VARS_TO_STR15(hsk, hsv, nh, nr23, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_K, type_V, permute, kv_view);
+        return VARS_TO_STR16(hsk, hsv, nh, nr23, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_K, type_V, permute, kv_view, v_is_view_of_k);
    }

    double max_nmse_err() override {
@@ -7104,9 +7105,9 @@ struct test_flash_attn_ext : public test_case {
    test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, std::array<int64_t, 2> nr23 = {1, 1}, int64_t kv = 96, int64_t nb = 8,
                        bool mask = true, bool sinks = false, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
                        ggml_type type_K = GGML_TYPE_F16, ggml_type type_V = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3},
-                        bool kv_view = true)
+                        bool kv_view = true, bool v_is_view_of_k = false)
        : hsk(hsk), hsv(hsv), nh(nh), nr23(nr23), kv(kv), nb(nb), mask(mask), sinks(sinks), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec),
-          type_K(type_K), type_V(type_V), permute(permute), kv_view(kv_view) {}
+          type_K(type_K), type_V(type_V), permute(permute), kv_view(kv_view), v_is_view_of_k(v_is_view_of_k) {}

    ggml_tensor * build_graph(ggml_context * ctx) override {
        const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_K));
@@ -7138,14 +7139,14 @@ struct test_flash_attn_ext : public test_case {
        ggml_set_name(k, "k");

        ggml_tensor * v = nullptr;
-        if (type_K == type_V && hsk_padded == 576 && hsv_padded == 512) {
-            // TODO: this branch should become a separate test case parameter instead of hardcoding this for these head shapes
-
-            // in this branch, the V cache is sub-view of the K cache. this is used by some MLA-based models
+        if (v_is_view_of_k) {
+            // the V cache is a sub-view of the K cache. this is used by some MLA-based models
            // for more info:
            //   - https://github.com/ggml-org/llama.cpp/pull/13435
            //   - https://github.com/ggml-org/llama.cpp/pull/18953#issuecomment-3774948392
            //   - https://github.com/ggml-org/llama.cpp/pull/18986
+            GGML_ASSERT(type_K == type_V && hsv_padded <= hsk_padded);
+
            v = ggml_view_4d(ctx, k, hsv_padded, kv, nh, nr23[1], k->nb[1], k->nb[2], k->nb[3], 0);
        } else {
            v = create_permuted(type_V,        hsv_padded, kv, nh,         nr23[1], kv_view); // the V tensor is usually a view of the V cache
@@ -9298,6 +9299,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {

    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 6, 4096, 5120, {1, 1}, {1, 1}));

+    // K not a multiple of 32
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32,  65, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32,  80, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 64, 32,  80, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 32,  80, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32, 588, {1, 1}, {1, 1})); // 14*14*3, e.g. conv_2d im2col
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32,  80, {4, 1}, {1, 1}));
+
 #if 0
    // test the mat-mat path for Metal
    for (int k = 1; k < 512; ++k) {
@@ -9898,12 +9907,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                                    if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
                                                    for (ggml_type type_KV : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0, GGML_TYPE_IQ4_NL}) {
                                                        if (type_KV != GGML_TYPE_F16 && hsk != 64 && hsk != 72) continue;
+                                                        // DeepSeek MLA: the V cache is a sub-view of the K cache
+                                                        const bool v_is_view_of_k = hsk == 576;
                                                        test_cases.emplace_back(new test_flash_attn_ext(
-                                                                    hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV));
+                                                                    hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV, {0, 1, 2, 3}, true, v_is_view_of_k));
                                                        // run fewer test cases permuted
                                                        if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
                                                            test_cases.emplace_back(new test_flash_attn_ext(
-                                                                        hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV, {0, 2, 1, 3}));
+                                                                        hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV, {0, 2, 1, 3}, true, v_is_view_of_k));
                                                        }
                                                    }
                                                }
@@ -9934,6 +9945,25 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
    test_cases.emplace_back(new test_flash_attn_ext(64, 128, 4, {1, 1}, 128, 2, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q4_0, GGML_TYPE_Q2_0));
    test_cases.emplace_back(new test_flash_attn_ext(128, 64, 4, {1, 1}, 64, 2, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q2_0, GGML_TYPE_F16));

+    // q8_0 KV cases: decode and prompt batches, KV pad, permuted KV, feature flags, and long context
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},   113,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 2},  1025,   1, true, true,  8, 30, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  1025,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 16384,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+
+    // MLA shape: the V cache is a sub-view of the K cache, with quantized KV
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1},  113,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+
+    // more V-is-sub-view-of-K cases: other head shapes, and full views with equal head sizes
+    test_cases.emplace_back(new test_flash_attn_ext(320, 256, 1, {32, 1}, 512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(192, 128, 4, {8, 1},  512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(128, 128, 8, {4, 1},  512, 8, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(64,  64,  4, {1, 1},  512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+
    // large-KV F16 cases (Qwen3.6-27B geometry and a llama-class control): the upstream matrix
    // stops at kv=1024, blind to long-context FA bugs (e.g. the oneDNN SDPA ordering race on BMG).
    for (int64_t kv : { 4096, 16384 }) {
@@ -10325,6 +10355,21 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
    test_cases.emplace_back(new test_flash_attn_ext(64, 64, 8, {8, 1}, 7680,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
    test_cases.emplace_back(new test_flash_attn_ext(64, 64, 8, {8, 1}, 7680, 512, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));

+    // q8_0 KV cases with long context (decode and prompt)
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},   128, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},   512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  1024, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  2048, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  4096, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 10000, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 20000, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 10000, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 20000, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 10000, 512, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 20000, 512, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 10000, 512, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 20000, 512, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+
    for (int kv : { 4096, 8192, 16384, }) {
        for (int hs : { 64, 128, }) {
            for (int nr : { 1, 4, }) {
@@ -1564,6 +1564,70 @@ int main() {
                space ::= | " " | "\n"{1,2} [ \t]{0,20}
            )""",
        });
+
+        run({
+            SUCCESS,
+            "unanchored regexp",
+            R"""({
+                "type": "string",
+                "pattern": "[0-9]+"
+            })""",
+            R"""(
+                char ::= [^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})
+                root ::= string
+                space ::= | " " | "\n"{1,2} [ \t]{0,20}
+                string ::= "\"" char* "\""
+            )""",
+        });
+
+        // the rules of the partial conversion (here "root-0") must not leak into the grammar
+        run({
+            SUCCESS,
+            "regexp with unsupported shorthand",
+            R"""({
+                "type": "string",
+                "pattern": "^[0-9]{3}\\w$"
+            })""",
+            R"""(
+                char ::= [^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})
+                root ::= string
+                space ::= | " " | "\n"{1,2} [ \t]{0,20}
+                string ::= "\"" char* "\""
+            )""",
+        });
+
+        // a regexp that is invalid under any flavor is still an error
+        run({
+            FAILURE,
+            "regexp with unbalanced parentheses",
+            R"""({
+                "type": "string",
+                "pattern": "^(a$"
+            })""",
+            ""
+        });
+
+        // only the property with the bad pattern degrades
+        run({
+            SUCCESS,
+            "unsupported regexp in a property",
+            R"""({
+                "type": "object",
+                "properties": {
+                    "a": { "type": "string", "pattern": "^[a-z\\-]+$" }
+                },
+                "required": ["a"],
+                "additionalProperties": false
+            })""",
+            R"""(
+                a ::= string
+                a-kv ::= "\"a\"" space ":" space a
+                char ::= [^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})
+                root ::= "{" space a-kv space "}"
+                space ::= | " " | "\n"{1,2} [ \t]{0,20}
+                string ::= "\"" char* "\""
+            )""",
+        });
    }

    if (getenv("LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR")) {
@@ -104,6 +104,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
    } else if (arch == LLM_ARCH_DEEPSEEK2
            || arch == LLM_ARCH_DEEPSEEK32
            || arch == LLM_ARCH_GLM_DSA
+            || arch == LLM_ARCH_DOTS3NOTE
            || arch == LLM_ARCH_KIMI_LINEAR
            || arch == LLM_ARCH_BAILINGMOE3
            || arch == LLM_ARCH_KIMI_K3
@@ -166,6 +167,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
    if (arch == LLM_ARCH_DEEPSEEK2
            || arch == LLM_ARCH_DEEPSEEK32
            || arch == LLM_ARCH_GLM_DSA
+            || arch == LLM_ARCH_DOTS3NOTE
            || arch == LLM_ARCH_KIMI_LINEAR
            || arch == LLM_ARCH_BAILINGMOE3
            || arch == LLM_ARCH_KIMI_K3
@@ -175,6 +177,22 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
        ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,       uint32_t(64));
        ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   uint32_t(192));
        ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, uint32_t(128));
+        if (arch == LLM_ARCH_DOTS3NOTE) {
+            // SWA layers reuse the same MLA geometry as the full layers in this fixture
+            ms.add_kv(LLM_KV_ATTENTION_KV_LORA_RANK_SWA,     uint32_t(512));
+            ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA,       uint32_t(576));
+            ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,     uint32_t(512));
+            ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,   uint32_t(192));
+            ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, uint32_t(128));
+            ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,             10000.0f);
+            // indexer on the full-attention layers (inverse of the swa pattern)
+            std::vector<uint32_t> indexer_types;
+            indexer_types.reserve(n_layer);
+            for (uint32_t il = 0; il < n_layer; il++) {
+                indexer_types.push_back(il % 2 ? 0 : 1);
+            }
+            ms.add_kv(LLM_KV_ATTENTION_INDEXER_TYPES, indexer_types);
+        }
    } else if (arch == LLM_ARCH_MINIMAX_M3) {
        // partial rotary: n_rot must not exceed the indexer key length (64)
        ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,       uint32_t(64));
@@ -197,7 +215,8 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
        ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,              10000.0f);
        // SWA pattern: every 5th layer is full attention (matches E2B layer_types)
        ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, uint32_t(5));
-    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_GRANITE_SWA) {
+    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 ||
+            arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_GRANITE_SWA || arch == LLM_ARCH_DOTS3NOTE) {
        std::vector<uint32_t> pattern;
        pattern.reserve(n_layer);
        for (uint32_t il = 0; il < n_layer; il++) {
@@ -365,6 +384,7 @@ static bool moe_mandatory(const llm_arch arch) {
        case LLM_ARCH_DEEPSEEK:
        case LLM_ARCH_DEEPSEEK2:
        case LLM_ARCH_DEEPSEEK32:
+        case LLM_ARCH_DOTS3NOTE:
        case LLM_ARCH_GLM4_MOE:
        case LLM_ARCH_GLM_DSA:
        case LLM_ARCH_EXAONE_MOE:
@@ -162,6 +162,7 @@
 | `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
 | `--mmproj-auto, --no-mmproj, --no-mmproj-auto` | whether to use multimodal projector file (if available), useful when using -hf (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_AUTO) |
 | `--mmproj-offload, --no-mmproj-offload` | whether to enable GPU offloading for multimodal projector (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_OFFLOAD) |
+| `-mmdev, --mmproj-device DEVICE` | device to use for multimodal projector (none = don't offload, default: auto)<br/>use --list-devices to see a list of available devices<br/>(env: MTMD_BACKEND_DEVICE) |
 | `--image, --audio, --video FILE` | path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files |
 | `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) |
 | `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) |
@@ -186,14 +186,13 @@ struct clip_ctx {
            throw std::runtime_error("failed to initialize CPU backend");
        }
        if (ctx_params.use_gpu) {
-            auto * backend_name = std::getenv("MTMD_BACKEND_DEVICE");
-            if (backend_name != nullptr) {
-                backend = ggml_backend_init_by_name(backend_name, nullptr);
+            if (ctx_params.device != nullptr) {
+                backend = ggml_backend_dev_init(ctx_params.device, nullptr);
                if (!backend) {
-                    LOG_WRN("%s: Warning: Failed to initialize \"%s\" backend, falling back to default GPU backend\n", __func__, backend_name);
+                    throw std::runtime_error(string_format("%s: failed to initialize \"%s\" backend\n",
+                                                           __func__, ggml_backend_dev_name(ctx_params.device)));
                }
-            }
-            if (!backend) {
+            } else {
                backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
                backend = backend ? backend : ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU, nullptr);
            }
@@ -48,6 +48,7 @@ enum clip_flash_attn_type {

 struct clip_context_params {
    bool use_gpu;
+    ggml_backend_dev_t device;
    enum clip_flash_attn_type flash_attn_type;
    int image_min_tokens;
    int image_max_tokens;
@@ -84,6 +84,7 @@ int main(int argc, char ** argv) {
        const char * clip_path = params.mmproj.path.c_str();
        mtmd_context_params mparams = mtmd_context_params_default();
        mparams.use_gpu          = params.mmproj_use_gpu;
+        mparams.device           = params.mmproj_device;
        mparams.print_timings    = true;
        mparams.n_threads        = params.cpuparams.n_threads;
        mparams.flash_attn_type  = params.flash_attn_type;
@@ -154,6 +154,7 @@ struct mtmd_cli_context {
        const char * clip_path = params.mmproj.path.c_str();
        mtmd_context_params mparams = mtmd_context_params_default();
        mparams.use_gpu          = params.mmproj_use_gpu;
+        mparams.device           = params.mmproj_device;
        mparams.print_timings    = true;
        mparams.n_threads        = params.cpuparams.n_threads;
        mparams.flash_attn_type  = params.flash_attn_type;
@@ -456,6 +456,7 @@ static clip_flash_attn_type mtmd_get_clip_flash_attn_type(enum llama_flash_attn_
 mtmd_context_params mtmd_context_params_default() {
    mtmd_context_params params {
        /* use_gpu           */ true,
+        /* device            */ nullptr,
        /* print_timings     */ true,
        /* n_threads         */ 4,
        /* image_marker      */ nullptr,
@@ -564,6 +565,7 @@ struct mtmd_context {

        clip_context_params ctx_clip_params {
            /* use_gpu           */ ctx_params.use_gpu,
+            /* device            */ ctx_params.device,
            /* flash_attn_type   */ mtmd_get_clip_flash_attn_type(ctx_params.flash_attn_type),
            /* image_min_tokens  */ ctx_params.image_min_tokens,
            /* image_max_tokens  */ ctx_params.image_max_tokens,
--- a/Show More
+++ b/Show More