cuda : fix CUB argsort corruption caused by in-place keys (#28389 )

argsort_f32_i32_cuda_cub called the one-shot DeviceRadixSort::SortPairs API with d_keys_in == d_keys_out (temp_keys, temp_keys). CUB's internal double-buffer ping-pong requires distinct key buffers: with aliased buffers the sort partially overwrites its own input mid-pass and emits a corrupted permutation, surfacing as intermittent garbage indices (e.g. backend top_k over a 248k-column vocab on Maxwell/CUDA 12.5/CCCL 2.x, which then triggered out-of-bounds gathers in downstream get_rows). Use a distinct keys-out buffer for all six call sites (plain and segmented, ascending and descending, size-query and execute). --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Oliver Simons <osimons@nvidia.com>
opencl: add support for bin kernel flash_attn_f32_f16_bin (#29046 )
2026-10-29 07:57:38 -05:00 · 2026-09-19 07:32:52 +02:00 · 2026-09-18 16:32:31 -07:00 · 2026-09-18 15:05:10 -07:00 · 2026-09-18 14:20:48 -07:00 · 2026-09-18 13:15:08 -07:00
132 changed files with 9558 additions and 6976 deletions
@@ -1,5 +1,5 @@
-ARG OPENVINO_VERSION_MAJOR=2026.3.1
-ARG OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d
+ARG OPENVINO_VERSION_MAJOR=2026.4
+ARG OPENVINO_VERSION_FULL=2026.4.0.22959.99c81491cc3
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
@@ -10,9 +10,9 @@ ARG COMPUTE_RUNTIME_VERSION_FULL=26.31.39395.13-0
 ARG IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.35.0
-ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
-ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1
+ARG NPU_DRIVER_VERSION=v1.38.0
+ARG NPU_DRIVER_FULL=v1.38.0.20260910-34487311128
+ARG LIBZE1_VERSION=1.32.0-1~24.04~ppa1

 # Optional proxy build arguments
 ARG http_proxy=
@@ -173,7 +173,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
    fi; \
    DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
    if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260830T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
    fi; \
    mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
    apt-get update; \
@@ -29,7 +29,7 @@ concurrency:

 jobs:
  android-ndk-snapdragon:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.7'
    defaults:
@@ -59,7 +59,7 @@ jobs:
          path: pkg-snapdragon/llama.cpp

  linux-iot-snapdragon:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.7'
    defaults:
@@ -33,7 +33,7 @@ env:

 jobs:
  default:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest

    steps:
      - name: Clone
@@ -49,7 +49,7 @@ jobs:
          distribution: zulu

      - name: Setup Android SDK
-        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        uses: android-actions/setup-android@be39fa834029ff78f1a44aa3bb0819b8fc2bd8fd # v4.0.4
        with:
          log-accepted-android-sdk-licenses: false

@@ -59,7 +59,7 @@ jobs:
          ./gradlew build --no-daemon

  ndk:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
    defaults:
@@ -93,7 +93,7 @@ jobs:
          path: pkg-adb/llama.cpp

  arm64:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest

    env:
      NDK_VERSION: "29.0.14206865"
@@ -123,7 +123,7 @@ jobs:
          distribution: temurin

      - name: Setup Android SDK
-        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        uses: android-actions/setup-android@be39fa834029ff78f1a44aa3bb0819b8fc2bd8fd # v4.0.4
        with:
          log-accepted-android-sdk-licenses: false

@@ -41,8 +41,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

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

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

    steps:
      - name: Clone
@@ -41,8 +41,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

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

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

    steps:
      - name: Clone
@@ -412,8 +412,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

    steps:
      - name: Clone
@@ -19,7 +19,7 @@ on:

 jobs:
  check-vendor:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim

    steps:
      - name: Checkout
@@ -11,10 +11,12 @@ on:
    paths:
      - .github/workflows/copilot-setup-steps.yml

+cache-mode: none
+
 jobs:
  # The job MUST be called `copilot-setup-steps` or it will not be picked up by Copilot.
  copilot-setup-steps:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04

    # Set the permissions to the lowest permissions possible needed for your steps.
    # Copilot will be given its own token for its operations.
@@ -31,8 +33,8 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.24
        with:
-          key: copilot-setup-steps
-          evict-old-files: 1d
+          restore: false
+          save: false

      - name: Dependencies
        id: depends
@@ -21,7 +21,7 @@ on:
 jobs:
  deploy:

-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest

    steps:
    - uses: actions/checkout@v6
@@ -13,6 +13,16 @@ on:
        required: true
        type: boolean
        default: true
+      skip_apiabi_check:
+        description: 'Skip API/ABI compatibility check'
+        required: false
+        type: boolean
+        default: false
+      apiabi_compare_tag:
+        description: 'Tag to compare against for API/ABI check (default: latest release)'
+        required: false
+        type: string
+        default: ''

 env:
  GH_TOKEN: ${{ github.token }}
@@ -23,7 +33,7 @@ permissions:

 jobs:
  make-release:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest

    steps:
      - name: Checkout
@@ -33,12 +43,18 @@ jobs:
          ref: ${{ inputs.commit != '' && inputs.commit || github.ref_name }}
          fetch-depth: 0

+      - name: Install API/ABI check tools
+        if: ${{ github.event.inputs.skip_apiabi_check != 'true' }}
+        run: sudo apt-get install -y abi-compliance-checker abigail-tools
+
      - name: Run release checks
        id: checks
        run: bash scripts/make-release-checks.sh ${{ github.event.inputs.dry_run == 'true' && '--dry-run' || '' }}
        env:
          GITHUB_REPOSITORY: ${{ github.repository }}
          RELEASE_BRANCH: ${{ github.ref_name }}
+          SKIP_APIABI_CHECK: ${{ github.event.inputs.skip_apiabi_check }}
+          APIABI_COMPARE_TAG: ${{ github.event.inputs.apiabi_compare_tag }}

      - name: Create release tag
        if: ${{ github.event.inputs.dry_run == 'false' }}
@@ -12,7 +12,7 @@ on:

 jobs:
    pre-tokenizer-hashes:
-        runs-on: [self-hosted, fast]
+        runs-on: ubuntu-slim

        steps:
        - name: Checkout repository
@@ -20,7 +20,7 @@ concurrency:

 jobs:
  python-check-requirements:
-    runs-on: [self-hosted, CPU, fast]
+    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
    name: check-requirements
    steps:
      - name: Check out source repository
@@ -21,7 +21,7 @@ concurrency:

 jobs:
  flake8-lint:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
    name: Lint
    steps:
      - name: Check out source repository
@@ -22,7 +22,7 @@ concurrency:

 jobs:
  python-type-check:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
    name: python type-check
    steps:
      - name: Check out source repository
@@ -106,6 +106,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-${{ matrix.os }}-${{ matrix.arch }}
+          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -190,6 +191,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-${{ matrix.os }}-cpu
+          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -275,6 +277,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-${{ matrix.os }}-vulkan
+          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -453,7 +456,7 @@ jobs:
    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest

    #permissions:
    #  actions: write
@@ -481,10 +484,9 @@ jobs:
          distribution: temurin

      - name: Setup Android SDK
-        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        uses: android-actions/setup-android@be39fa834029ff78f1a44aa3bb0819b8fc2bd8fd # v4.0.4
        with:
          log-accepted-android-sdk-licenses: false
-          packages: 'platform-tools'

      - name: Install NDK
        run: |
@@ -501,6 +503,7 @@ jobs:
      #  uses: ggml-org/ccache-action@v1.2.24
      #  with:
      #    key: release-android-arm64
+      #    evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -555,8 +558,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

    steps:
      - name: Set OpenVINO version output
@@ -579,6 +582,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-ubuntu-24.04-openvino-release-no-preset-v1
+          evict-old-files: 1d

      - name: Dependencies
        run: |
@@ -669,8 +673,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.4"
+      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"

    steps:
      - name: Set OpenVINO version output
@@ -822,6 +826,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+          evict-old-files: 1d

      - name: Build
        shell: cmd
@@ -1066,6 +1071,7 @@ jobs:
      #  uses: ggml-org/ccache-action@v1.2.24
      #  with:
      #    key: release-windows-2025-${{ matrix.arch }}-${{ matrix.backend }}
+      #    evict-old-files: 1d

      - name: Install OpenCL Headers and Libs
        id: install_opencl
@@ -1154,6 +1160,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -1250,6 +1257,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-windows-2022-x64-sycl
+          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -1368,6 +1376,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -16,7 +16,7 @@ on:

 jobs:
    update-ops-docs:
-        runs-on: [self-hosted, fast, ARM64]
+        runs-on: ubuntu-slim

        steps:
        - name: Checkout repository
@@ -8,7 +8,7 @@ on:
 jobs:
  update:
    name: Update Winget Package
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04  # previously ubuntu-latest
    if: github.repository_owner == 'ggml-org'

    steps:
@@ -97,7 +97,6 @@
 /src/models/                            @CISC
 /tests/                                 @ggerganov
 /tests/test-chat.*                      @pwilkin
-/tests/test-llama-archs.cpp             @JohannesGaessler
 /tools/batched-bench/                   @ggerganov
 /tools/cli/                             @ngxson
 /tools/completion/                      @ggerganov
@@ -104,6 +104,12 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
    const std::string GEN_PROMPT   = "<｜Assistant｜>";
    const std::string TC_SEPARATOR = "\n\n";

+    // lets the server find user turns in the prompt and place context checkpoints there
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, GEN_PROMPT },
+        { COMMON_CHAT_ROLE_USER, "<｜User｜>" },
+    };
+
    data.prompt = common_chat_template_direct_apply_impl(
        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
    data.generation_prompt = common_chat_template_generation_prompt_impl(
@@ -1861,6 +1861,9 @@ class TextModel(ModelBase):
        if chkhsh == "972da7b59cec44d1f0a490a86c96df53859e486e481563e5dddac155013d87ac":
            # ref: https://huggingface.co/poolside/Laguna-XS.2
            res = "laguna"
+        if chkhsh == "653660222fb704f61cbf2b618a8ae6502b7f8b20c980f9a5de07ed78e13319cd":
+            # ref: https://huggingface.co/ufakai/ufakzeka-1
+            res = "ufakzeka"

        if res is None:
            logger.warning("\n")
@@ -163,6 +163,7 @@ models = [
    {"name": "granite-embed-multi-311m", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-embedding-311m-multilingual-r2", },
    {"name": "mellum2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Base"},
    {"name": "laguna",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/poolside/Laguna-XS.2", },
+    {"name": "ufakzeka",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ufakai/ufakzeka-1", },
 ]

 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -12,6 +12,8 @@ The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a t
 - Compiles and caches the model for the target device.
 - Binds GGML tensor memory to OpenVINO inference tensors and runs inference.

+For guidance on contributing to the OpenVINO backend, see the [OpenVINO Backend Contributing Guide](https://github.com/ravi9/llamacpp-ov-dev-guide/blob/main/contributing-llamacpp-ov.md).
+
 ## Contents

 - [Supported Devices](#supported-devices)
@@ -96,7 +98,7 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
  - **SL** = Stateless (`GGML_OPENVINO_STATEFUL_EXECUTION=0`)
  - **SF** = Stateful (`GGML_OPENVINO_STATEFUL_EXECUTION=1`)
  - Note: The NPU operates in stateless mode only.
- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel OpenCL GPU Driver 26.31.39395.13-0 | Intel NPU Driver 1.35.0.
+- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel Graphics Compiler 2.41.5 | Intel OpenCL GPU Driver 26.31.39395.13-0 | Intel NPU Driver 1.38.0.
 - See [Known Limitations](#known-limitations) for context on observed failures.

 | Model | CPU (SL / SF) | GPU (SL / SF) | NPU (SL) |
@@ -117,9 +119,9 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [lmstudio-community/Qwen3.5-9B-Q4_K_M](https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
 |  |  |  |  |
 | [unsloth/gemma-3-4b-it-Q4_K_M](https://huggingface.co/unsloth/gemma-3-4b-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [bartowski/google_gemma-4-E4B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
-| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
+| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✗ |
+| [bartowski/google_gemma-4-E4B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✗ |
 |  |  |  |  |
 | [bartowski/Phi-3-mini-4k-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/Phi-3.5-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
@@ -132,9 +134,9 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [bartowski/DeepSeek-R1-Distill-Llama-8B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
-| [ibm-granite/granite-4.0-350m-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF) | ✓ / ✓ | ✗ / ✗ | ✓ |
+| [ibm-granite/granite-4.0-350m-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [ibm-granite/granite-4.0-micro-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-micro-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [ibm-granite/granite-4.0-1b-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-1b-GGUF) | ✓ / ✓ | ✗ / ✗ | ✗ |
+| [ibm-granite/granite-4.0-1b-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-1b-GGUF) | ✓ / ✓ | ✓ / ✓ | ✗ |
 | [ibm-research/granite-3.2-8b-instruct-Q4_K_M](https://huggingface.co/ibm-research/granite-3.2-8b-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
 | [HuggingFaceTB/smollm2-1.7b-instruct-q4_k_m](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
@@ -242,8 +244,8 @@ chmod +x build-llamacpp-ov.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.3.1"
-OPENVINO_VERSION_FULL="2026.3.1.22476.56d9685302d"
+OPENVINO_VERSION_MAJOR="2026.4"
+OPENVINO_VERSION_FULL="2026.4.0.22959.99c81491cc3"

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -340,7 +342,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.4` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.

 </details>

@@ -370,8 +372,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================

-set "OPENVINO_VERSION_MAJOR=2026.3.1"
-set "OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d"
+set "OPENVINO_VERSION_MAJOR=2026.4"
+set "OPENVINO_VERSION_FULL=2026.4.0.22959.99c81491cc3"

 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -550,7 +552,7 @@ endlocal
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.4` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.

 </details>

@@ -1,5 +1,7 @@
-set(TARGET llama-convert-llama2c-to-ggml)
-add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+if (GGML_CPU)
+    set(TARGET llama-convert-llama2c-to-ggml)
+    add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
+    install(TARGETS ${TARGET} RUNTIME)
+    target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+    target_compile_features(${TARGET} PRIVATE cxx_std_17)
+endif()
@@ -25,7 +25,7 @@ GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);

 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type([[maybe_unused]] int main_device, const float * tensor_split);

 // Tensor parallelism (--split-mode tensor): comm_init/free/allreduce_tensor
 // trio queried by the meta-backend via ggml_backend_reg_get_proc_address.
@@ -1630,7 +1630,10 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
            ggml_backend_synchronize(sched->backends[i]);
        }

-        ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
+        if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
+            GGML_LOG_ERROR("%s: failed to reserve graph buffers\n", __func__);
+            return false;
+        }
        if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
            GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__);
            return false;
@@ -1329,7 +1329,9 @@ UseGgmlGemm1:;
        const size_t nbw3 = nbw2*ne12;

        assert(params->wsize >= ne13*nbw3);
-        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+        GGML_ASSERT(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
+        // the F16 path below writes plain floats into wdata, so it needs an F32 vec_dot_type
+        GGML_ASSERT(src1->type == GGML_TYPE_F32 || vec_dot_type == GGML_TYPE_F32);

    #if 0
        for (int64_t i13 = 0; i13 < ne13; ++i13) {
@@ -1348,9 +1350,15 @@ UseGgmlGemm1:;
                    size_t bs = ggml_blck_size(vec_dot_type);
                    int64_t ne10_block_start = (ith * ne10/bs) / nth;
                    int64_t ne10_block_end   = ((ith + 1) * ne10/bs) / nth;
-                    from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10),
-                               (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0),
-                               (ne10_block_end - ne10_block_start) * bs);
+                    const char * src1_block = (const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10;
+                    char * dst_block = wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0;
+                    const int64_t n_block = (ne10_block_end - ne10_block_start) * bs;
+
+                    if (src1->type == GGML_TYPE_F32) {
+                        from_float((const float *) src1_block, dst_block, n_block);
+                    } else {
+                        ggml_cpu_fp16_to_fp32((const ggml_fp16_t *) src1_block, (float *) dst_block, n_block);
+                    }
                }
            }
        }
@@ -451,6 +451,10 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
                op->type != GGML_TYPE_IQ1_S   &&
                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
        case GGML_OP_MUL_MAT:
+            if (ggml_get_op_params_i32(op, 1) == GGML_HINT_SRC0_IS_HADAMARD &&
+                src0->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) {
+                return src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16;
+            }
            return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
        case GGML_OP_SOFT_MAX_BACK: {
            if (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type != GGML_TYPE_F32) {
@@ -12015,11 +12015,20 @@ void ggml_compute_forward_opt_step_sgd(const ggml_compute_params * params, ggml_
    }
 }

-static void ggml_compute_forward_fwht_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+static inline float ggml_fwht_load(const float value) {
+    return value;
+}
+
+static inline float ggml_fwht_load(const ggml_fp16_t value) {
+    return ggml_fp16_to_fp32(value);
+}
+
+template<typename src_t>
+static void ggml_compute_forward_fwht_impl(const ggml_compute_params * params, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];

-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == (std::is_same_v<src_t, float> ? GGML_TYPE_F32 : GGML_TYPE_F16));
    GGML_ASSERT(dst->type == GGML_TYPE_F32);

    GGML_TENSOR_BINARY_OP_LOCALS
@@ -12046,11 +12055,11 @@ static void ggml_compute_forward_fwht_f32(const ggml_compute_params * params, gg
        const int64_t i12 = (r - i13 * ne11 * ne12) / ne11;
        const int64_t i11 = r - i13 * ne11 * ne12 - i12 * ne11;

-        const float * src_row = (const float *) ((const char *) src1->data + i11 * nb11 + i12 * nb12 + i13 * nb13);
+        const src_t * src_row = (const src_t *) ((const char *) src1->data + i11 * nb11 + i12 * nb12 + i13 * nb13);
        float * dst_row = (float *) ((char *) dst->data + i11 * nb1 + i12 * nb2 + i13 * nb3);

        for (int64_t j = 0; j < n; j++) {
-            dst_row[j] = src_row[j] * scale;
+            dst_row[j] = ggml_fwht_load(src_row[j]) * scale;
        }

        // Scalar passes
@@ -12097,12 +12106,17 @@ void ggml_compute_forward_fwht(const ggml_compute_params * params, ggml_tensor *
    switch (src1->type) {
        case GGML_TYPE_F32:
            {
-                ggml_compute_forward_fwht_f32(params, dst);
+                ggml_compute_forward_fwht_impl<float>(params, dst);
+            }
+            break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_fwht_impl<ggml_fp16_t>(params, dst);
            }
            break;
        default:
            {
-                GGML_ABORT("fatal error - fwht is F32 only");
+                GGML_ABORT("fatal error - fwht supports F32 and F16 input");
            }
    }
 }
@@ -51,9 +51,12 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                              cudaStream_t     stream) {
    ggml_cuda_pool_alloc<int>   temp_indices_alloc(pool, ncols * nrows);
    ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
+    // Device*Sort algorithms currently do not allow for in-place sorting/aliasing of input/outputs
+    ggml_cuda_pool_alloc<float> temp_keys_out_alloc(pool, ncols * nrows);

    int *   temp_indices = temp_indices_alloc.get();
    float * temp_keys    = temp_keys_alloc.get();
+    float * temp_keys_out = temp_keys_out_alloc.get();

    static const int block_size = 256;
    const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
@@ -85,18 +88,18 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,

    if (order == GGML_SORT_ORDER_ASC) {
        if (nrows == 1) {
-            CUDA_CHECK(DeviceRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+            CUDA_CHECK(DeviceRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys_out,  // keys in, keys out
                                                  temp_indices, dst,  // values (indices)
                                                  ncols, 0, sizeof(float) * 8, stream));
        } else if (is_capturing) {
            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(
-                nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+                nullptr, temp_storage_bytes, temp_keys, temp_keys_out,  // keys in, keys out
                temp_indices, dst,                                  // values (indices)
                ncols * nrows, nrows,                               // num items, num segments
                offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
        } else {
            CUDA_CHECK(DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys,
-                                                      temp_keys,             // keys (in-place)
+                                                      temp_keys_out, // keys out
                                                      temp_indices, dst,     // values (indices)
                                                      ncols * nrows, nrows,  // num items, num segments
                                                      offset_iterator, offset_iterator + 1, stream));
@@ -104,15 +107,15 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
    } else {
        if (nrows == 1) {
            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys,
-                                                            temp_keys,          // keys (in-place)
+                                                            temp_keys_out, // keys out
                                                            temp_indices, dst,  // values (indices)
                                                            ncols, 0, sizeof(float) * 8, stream));
        } else if (is_capturing) {
            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairsDescending(
-                nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst, ncols * nrows, nrows,
+                nullptr, temp_storage_bytes, temp_keys, temp_keys_out, temp_indices, dst, ncols * nrows, nrows,
                offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
        } else {
-            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys,
+            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys_out,
                                                                temp_indices, dst, ncols * nrows, nrows,
                                                                offset_iterator, offset_iterator + 1, stream));
        }
@@ -124,31 +127,31 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
    if (order == GGML_SORT_ORDER_ASC) {
        if (nrows == 1) {
            CUDA_CHECK(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys,
-                                                  temp_keys,          // keys (in-place)
+                                                  temp_keys_out, // keys out
                                                  temp_indices, dst,  // values (indices)
                                                  ncols, 0, sizeof(float) * 8, stream));
        } else if (is_capturing) {
-            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys_out,
                                                           temp_indices, dst, ncols * nrows, nrows, offset_iterator,
                                                           offset_iterator + 1, 0, sizeof(float) * 8, stream));
        } else {
-            CUDA_CHECK(DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+            CUDA_CHECK(DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys_out,
                                                      temp_indices, dst, ncols * nrows, nrows, offset_iterator,
                                                      offset_iterator + 1, stream));
        }
    } else {
        if (nrows == 1) {
            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys,
-                                                            temp_keys,          // keys (in-place)
+                                                            temp_keys_out, // keys out
                                                            temp_indices, dst,  // values (indices)
                                                            ncols, 0, sizeof(float) * 8, stream));
        } else if (is_capturing) {
            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairsDescending(
-                d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst, ncols * nrows, nrows,
+                d_temp_storage, temp_storage_bytes, temp_keys, temp_keys_out, temp_indices, dst, ncols * nrows, nrows,
                offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
        } else {
            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys,
-                                                                temp_keys, temp_indices, dst, ncols * nrows, nrows,
+                                                                temp_keys_out, temp_indices, dst, ncols * nrows, nrows,
                                                                offset_iterator, offset_iterator + 1, stream));
        }
    }
@@ -4000,7 +4000,9 @@ static bool ggml_hexagon_flash_attn_is_hmx_eligible(
    const uint32_t DK = q->ne[0];
    const uint32_t DV = v->ne[0];

-    if (DK % 64 != 0 || DV % 64 != 0) {
+    // Head dims that are not multiples of 64 are handled by internally padding to
+    // DK_pad/DV_pad = round_up(.,64) and zero-filling the tail lanes.
+    if (DK % 8 != 0 || DV % 8 != 0) {
        return false;
    }

@@ -4073,8 +4075,13 @@ static bool ggml_hexagon_precompute_flash_attn_params(
    // Check HMX eligibility
    const struct ggml_tensor * sinks = op->src[4];
    if (ggml_hexagon_flash_attn_is_hmx_eligible(sess, q, k, v, sinks)) {
+        // HMX tiles head_dim in units of 64; when DK/DV are not 64-aligned the kernel
+        // operates on padded dims with zero-filled tail lanes. VTCM budget and chunk-size
+        // are sized for the padded tiles.
+        const uint32_t DK_pad = hex_round_up(DK, 64);
+        const uint32_t DV_pad = hex_round_up(DV, 64);
        size_t Br = 0, Bc = 0;
-        int ret = hmx_fa_find_chunk_size(&Br, &Bc, G, DK, DV, neq1, nek1, sess->vtcm_size, sess->n_threads, kparams->is_q_fp32 != 0);
+        int ret = hmx_fa_find_chunk_size(&Br, &Bc, G, DK_pad, DV_pad, neq1, nek1, sess->vtcm_size, sess->n_threads, kparams->is_q_fp32 != 0);
        if (ret == 0) {
            kparams->kernel_type = HTP_FA_KERNEL_HMX;
            kparams->Br = Br;
@@ -4084,7 +4091,7 @@ static bool ggml_hexagon_precompute_flash_attn_params(

            kparams->u.hmx.g_br = hex_align_up(G * Br, 32);
            kparams->u.hmx.pipeline = (kparams->n_kv_blocks >= 3 && sess->n_threads >= 2) ? 1 : 0;
-            kparams->vtcm_size = hmx_fa_compute_vtcm_usage(G, DK, DV, Br, Bc, kparams->n_threads, kparams->u.hmx.pipeline != 0, kparams->is_q_fp32 != 0);
+            kparams->vtcm_size = hmx_fa_compute_vtcm_usage(G, DK_pad, DV_pad, Br, Bc, kparams->n_threads, kparams->u.hmx.pipeline != 0, kparams->is_q_fp32 != 0);

            const size_t row_vec_bytes = hex_align_up(Bc * sizeof(uint16_t), 256);
            kparams->u.hmx.row_buf_stride = row_vec_bytes / 128; // HVX vector is 128 bytes
@@ -5523,11 +5530,6 @@ static bool ggml_hexagon_supported_im2col(const struct ggml_hexagon_session * se
    const struct ggml_tensor * src1 = op->src[1];
    const struct ggml_tensor * dst  = op;

-    const bool is_2D = ((const int32_t *) op->op_params)[6] == 1;
-    if (!is_2D) {
-        return false;
-    }
-
    // For now support F32->F32 and F32->F16 only.
    if (src1->type != GGML_TYPE_F32 || (dst->type != GGML_TYPE_F16 && dst->type != GGML_TYPE_F32)) {
        return false;
@@ -5537,13 +5539,6 @@ static bool ggml_hexagon_supported_im2col(const struct ggml_hexagon_session * se
        return false;
    }

-    // For now keep padded OPs on CPU. Will revisit once we expand coverage past patch-embed OPs.
-    const int32_t p0 = ((const int32_t *) op->op_params)[2];
-    const int32_t p1 = ((const int32_t *) op->op_params)[3];
-    if (p0 != 0 || p1 != 0) {
-        return false;
-    }
-
    GGML_UNUSED(sess);
    return true;
 }
@@ -5704,6 +5699,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
        case GGML_OP_TRI:             return HTP_OP_TRI;
        case GGML_OP_PAD:             return HTP_OP_PAD;
        case GGML_OP_IM2COL:          return HTP_OP_IM2COL;
+        case GGML_OP_ROLL:            return HTP_OP_ROLL;

        case GGML_OP_UNARY:
            switch (ggml_get_unary_op(t)) {
@@ -6636,6 +6632,31 @@ static bool ggml_hexagon_supported_fill(const struct ggml_hexagon_session * sess
    GGML_UNUSED(sess);
 }

+static bool ggml_hexagon_supported_roll(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    GGML_UNUSED(sess);
+
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    if (src0->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (!ggml_are_same_shape(src0, dst)) {
+        return false;
+    }
+
+    if (src0->nb[0] != ggml_type_size(src0->type) || dst->nb[0] != ggml_type_size(dst->type)) {
+        return false;
+    }
+
+    if (!ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
    auto dev_ctx = static_cast<ggml_backend_hexagon_device_context *>(dev->context);
    auto sess    = dev_ctx->session();
@@ -6803,6 +6824,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
            supp = ggml_hexagon_supported_pad(sess, op);
            break;

+        case GGML_OP_ROLL:
+            supp = ggml_hexagon_supported_roll(sess, op);
+            break;
+
        default:
            break;
    }
@@ -43,6 +43,7 @@ add_library(${HTP_LIB} SHARED
    pad-ops.c
    argsort-ops.c
    im2col-ops.c
+    roll-ops.c
    allreduce-ops.c
 )

@@ -108,6 +108,7 @@ struct hmx_fa_context {

    // Dimensions
    uint32_t     DK, DV;
+    uint32_t     DK_pad, DV_pad;  // head_dim rounded up to 64 for HMX tiling
    uint32_t     n_kv;        // kv_len
    uint32_t     n_kv_heads;  // number of KV heads
    uint32_t     n_heads;     // number of Q heads
@@ -652,7 +653,7 @@ static void fa_k_interleave_thread(unsigned int n, unsigned int i, void * data)
            hvx_dequantize_row_q8_0_f16(row_k, row_k, factx->DK);
        }
    }
-    hmx_interleave_rows_to_tiles(factx->vtcm_k_tiles[args->buf_idx], (const __fp16 *) args->curr_k, total_rows, factx->DK,
+    hmx_interleave_rows_to_tiles(factx->vtcm_k_tiles[args->buf_idx], (const __fp16 *) args->curr_k, total_rows, factx->DK_pad,
                             args->src_stride, start, end);
    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_FA_K_PREP, (uint16_t) (args->kv_start + start));
 }
@@ -706,7 +707,7 @@ static void fa_v_interleave_thread(unsigned int n, unsigned int i, void * data)
            hvx_dequantize_row_q8_0_f16(row_v, row_v, factx->DV);
        }
    }
-    hmx_interleave_cols_to_tiles(v_tiles_dst, (const __fp16 *) args->v_src, total_rows, factx->DV,
+    hmx_interleave_cols_to_tiles(v_tiles_dst, (const __fp16 *) args->v_src, total_rows, factx->DV_pad,
                             args->src_stride, (uint32_t) args->n_col_tiles, start, end);
    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_FA_V_PREP, (uint16_t) (args->kv_start + start));
 }
@@ -832,17 +833,22 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
        const uint32_t            kv_head = args->kv_head;
        const uint32_t            ib3     = args->ib3;

-        assert(factx->DK == factx->DV);
-
        const bool use_q_dma = (factx->vtcm_q_dma != NULL);

        __fp16 * q_tiles = factx->vtcm_q_tiles;
+        const size_t DK_pad = factx->DK_pad;
        if (use_q_dma) {
            const size_t g_rows_end = hex_smin(end, n_rows_g);
            const uint32_t d_limit = factx->is_q_fp32 ? DK / 32 : DK / 64;

            uint8_t * q_flat  = (uint8_t *) factx->vtcm_q_dma;
-            if (factx->is_q_fp32) {
+            if (DK_pad != DK) {
+                if (factx->is_q_fp32) {
+                    hmx_fa_q_prep_fp32_pad(q_tiles, q_flat, start, end, g_rows_end, DK, DK_pad, G, args->n_rows_q, &factx->div_G, args->q_transposed);
+                } else {
+                    hmx_fa_q_prep_fp16_pad(q_tiles, q_flat, start, end, g_rows_end, DK, DK_pad, G, args->n_rows_q, &factx->div_G, args->q_transposed);
+                }
+            } else if (factx->is_q_fp32) {
                switch (d_limit) {
                case 2:  hmx_fa_q_prep_fp32_d2(q_tiles, q_flat, start, end, g_rows_end, DK, G, args->n_rows_q, &factx->div_G, args->q_transposed); break;
                case 4:  hmx_fa_q_prep_fp32_d4(q_tiles, q_flat, start, end, g_rows_end, DK, G, args->n_rows_q, &factx->div_G, args->q_transposed); break;
@@ -858,7 +864,7 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
        } else {
            // Fallback: direct-from-DDR/L2 path
            hmx_fa_q_prep_fallback(q_tiles, q->data, q->nb[1], q->nb[2], q->nb[3],
-                                   q_start, kv_head, ib3, start, end, n_rows_g, G, DK, factx->is_q_fp32, &factx->div_G);
+                                   q_start, kv_head, ib3, start, end, n_rows_g, G, DK, DK_pad, factx->is_q_fp32, &factx->div_G);
        }
    }

@@ -952,6 +958,8 @@ static void fa_o_store_thread_f32(unsigned int n, unsigned int i, void * data) {
    const uint32_t            kv_head    = args->kv_head;
    const uint32_t            ib3        = args->ib3;

+    const size_t DV_pad = factx->DV_pad;
+
    size_t q_idx = fastdiv(start, &factx->div_G);
    size_t h_idx = fastmodulo(start, G, &factx->div_G);

@@ -961,7 +969,7 @@ static void fa_o_store_thread_f32(unsigned int n, unsigned int i, void * data) {

        size_t         r0            = r / HMX_FP16_TILE_N_ROWS;
        size_t         r1            = r % HMX_FP16_TILE_N_ROWS;
-        const __fp16 * tile_row_base = o_tile_src + r0 * HMX_FP16_TILE_N_ROWS * DV;
+        const __fp16 * tile_row_base = o_tile_src + r0 * HMX_FP16_TILE_N_ROWS * DV_pad;

        for (uint32_t d = 0; d < DV / 32; ++d) {
            const HVX_Vector * in_tile = (const HVX_Vector *) (tile_row_base + d * HMX_FP16_TILE_N_ELMS);
@@ -972,6 +980,16 @@ static void fa_o_store_thread_f32(unsigned int n, unsigned int i, void * data) {
                *(HVX_UVector *) (out + d * 32) = Q6_V_hi_W(vp);
            }
        }
+        // Ragged tail: DV not a multiple of 32 (e.g. 72 -> last 8 lanes). Partial vector-write
+        // for the remaining (DV % 32) floats.
+        const uint32_t d_tail = DV / 32;
+        const uint32_t rem    = DV - d_tail * 32;
+        if (rem) {
+            const HVX_Vector * in_tile = (const HVX_Vector *) (tile_row_base + d_tail * HMX_FP16_TILE_N_ELMS);
+            HVX_VectorPair     vp      = hvx_vec_f16_to_f32_shuff(in_tile[r1 / 2]);
+            HVX_Vector         vd      = (r1 % 2 == 0) ? Q6_V_lo_W(vp) : Q6_V_hi_W(vp);
+            hvx_vec_store_u((void *) (out + d_tail * 32), rem * sizeof(float), vd);
+        }

        h_idx++;
        if (h_idx == G) {
@@ -1006,6 +1024,9 @@ static void fa_o_store_thread_f16(unsigned int n, unsigned int i, void * data) {
    const uint32_t            kv_head    = args->kv_head;
    const uint32_t            ib3        = args->ib3;

+    // O-tiles use the padded head dim (DV_pad); dst holds the real DV lanes.
+    const size_t DV_pad = factx->DV_pad;
+
    size_t q_idx = fastdiv(start, &factx->div_G);
    size_t h_idx = fastmodulo(start, G, &factx->div_G);

@@ -1015,7 +1036,7 @@ static void fa_o_store_thread_f16(unsigned int n, unsigned int i, void * data) {

        size_t         r0            = r / HMX_FP16_TILE_N_ROWS;
        size_t         r1            = r % HMX_FP16_TILE_N_ROWS;
-        const __fp16 * tile_row_base = o_tile_src + r0 * HMX_FP16_TILE_N_ROWS * DV;
+        const __fp16 * tile_row_base = o_tile_src + r0 * HMX_FP16_TILE_N_ROWS * DV_pad;

        for (uint32_t d = 0; d < DV / 64; ++d) {
            const __fp16 *     in_dtile = tile_row_base + d * HMX_FP16_TILE_N_ELMS * 2;
@@ -1028,6 +1049,17 @@ static void fa_o_store_thread_f16(unsigned int n, unsigned int i, void * data) {
                *(HVX_UVector *) (out + d * 64) = Q6_V_hi_W(vp);
            }
        }
+        // Ragged tail when DV is not a multiple of 64.
+        const uint32_t d_tail = DV / 64;
+        const uint32_t rem    = DV - d_tail * 64;
+        if (rem) {
+            const __fp16 *     in_dtile = tile_row_base + d_tail * HMX_FP16_TILE_N_ELMS * 2;
+            const HVX_Vector * pv_in0   = ((const HVX_Vector *) in_dtile) + r1 / 2;
+            const HVX_Vector * pv_in1   = pv_in0 + 16;
+            HVX_VectorPair     vp       = Q6_W_vdeal_VVR(*pv_in1, *pv_in0, -2);
+            HVX_Vector         vd       = (r1 % 2 == 0) ? Q6_V_lo_W(vp) : Q6_V_hi_W(vp);
+            hvx_vec_store_u((void *) (out + d_tail * 64), rem * sizeof(__fp16), vd);
+        }

        h_idx++;
        if (h_idx == G) {
@@ -1829,8 +1861,11 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    const uint32_t DK = neq0;
    const uint32_t DV = nev0;

-    // HMX requires head_dim to be multiple of 32
-    if (DK % 32 != 0 || DV % 32 != 0) {
+    // HMX tiles head_dim in units of 64. head_dim need not be 64- (or 32-) aligned:
+    // we can operate on DK/DV rounded up to 64 with tail lanes [D, D_pad) zero-filled.
+    const uint32_t DK_pad = hex_round_up(DK, 64);
+    const uint32_t DV_pad = hex_round_up(DV, 64);
+    if (DK == 0 || DV == 0) {
        return HTP_STATUS_NO_SUPPORT;
    }

@@ -1847,6 +1882,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    factx.n_threads      = kparams->n_threads;
    factx.DK             = DK;
    factx.DV             = DV;
+    factx.DK_pad         = DK_pad;
+    factx.DV_pad         = DV_pad;
    factx.n_kv           = nek1;
    factx.n_kv_heads     = n_kv_heads;
    factx.n_heads        = neq2;
@@ -1905,16 +1942,18 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {

    // ======== VTCM allocation (GQA-aware) ========
    // K/V row sizes drive the DMA descriptors (not the VTCM layout) and are used
-    // throughout the KV loop below.
+    // throughout the KV loop below. The DMA copies only the real DK/DV columns; the
+    // staging rows are padded to hold DK_pad/DV_pad columns (tail zero-filled below)
+    // so the HMX interleave/tile logic can operate on 64-aligned head dims.
    const size_t size_k_row        = htp_tensor_get_row_size(k->type, DK);
    const size_t size_v_row        = htp_tensor_get_row_size(v->type, DV);
-    const size_t size_k_row_padded = hex_round_up(DK * sizeof(__fp16), 128);
-    const size_t size_v_row_padded = hex_round_up(DV * sizeof(__fp16), 128);
+    const size_t size_k_row_padded = hex_round_up(DK_pad * sizeof(__fp16), 128);
+    const size_t size_v_row_padded = hex_round_up(DV_pad * sizeof(__fp16), 128);

    // Build the VTCM layout once (shared with the host estimator) and place every
-    // scratch buffer at its computed offset.
+    // scratch buffer at its computed offset. Padded head dims size the HMX tiles.
    struct hmx_fa_vtcm_layout L;
-    hmx_fa_vtcm_layout_build(&L, G, DK, DV, Br, Bc, n_threads, pipeline, factx.is_q_fp32);
+    hmx_fa_vtcm_layout_build(&L, G, DK_pad, DV_pad, Br, Bc, n_threads, pipeline, factx.is_q_fp32);

    if (L.total_bytes > ctx->vtcm_size) {
        return HTP_STATUS_VTCM_TOO_SMALL;
@@ -1961,6 +2000,24 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {

    dma_cache_init(&factx.m_cache, (uint8_t *) factx.vtcm_mask_buf, L.m_buf_slot_bytes, HMX_FA_DMA_CACHE_SIZE);

+    // Head-dim padding: the K/V DMA staging buffers and the flat-Q buffer are laid out
+    // with padded row strides (size_{k,v,q}_row_padded, covering D_pad columns) but the
+    // DMA only writes the real D columns per row. Zero the whole staging buffers once up
+    // front so tail lanes [D, D_pad) stay zero for all KV blocks. No-op when already aligned.
+    if (DK_pad != DK || DV_pad != DV) {
+        const size_t k_buf_bytes = (size_t) factx.Bc * size_k_row_padded;
+        const size_t v_buf_bytes = (size_t) factx.Bc * size_v_row_padded;
+        hvx_splat_u8_a((char *) factx.vtcm_k_fp16[0], 0, k_buf_bytes);
+        hvx_splat_u8_a((char *) factx.vtcm_k_fp16[1], 0, k_buf_bytes);
+        hvx_splat_u8_a((char *) factx.vtcm_v_fp16[0], 0, v_buf_bytes);
+        hvx_splat_u8_a((char *) factx.vtcm_v_fp16[1], 0, v_buf_bytes);
+        // Flat-Q DMA scratch
+        if (factx.vtcm_q_dma) {
+            const size_t q_dma_bytes = hex_align_up(factx.g_br * DK * (factx.is_q_fp32 ? sizeof(float) : sizeof(__fp16)), 128);
+            hvx_splat_u8_a((char *) factx.vtcm_q_dma, 0, q_dma_bytes);
+        }
+    }
+
    // ======== Initialize HMX output scales ========
    hmx_init_column_scales(factx.vtcm_hmx_scales_id, Q6_V_vsplat_R(0x3c00)); // 1.0
    hmx_init_column_scales(factx.vtcm_hmx_scales_qk, hvx_vec_splat_f16(factx.scale));
@@ -2072,7 +2129,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                    qk_job[0].s_tiles        = factx.vtcm_s_tiles[0];
                    qk_job[0].n_row_tiles    = n_row_tiles;
                    qk_job[0].n_col_tiles    = hmx_ceil_div(kv_rows0, HMX_FP16_TILE_N_COLS);
-                    qk_job[0].n_dot_tiles    = DK / 32;
+                    qk_job[0].n_dot_tiles    = DK_pad / 32;
                    qk_job[0].n_tiles_per_bc = n_tiles_per_bc;
                    qk_job[0].hmx_scales     = factx.vtcm_hmx_scales_qk;
                    hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_qk_dot_worker, &qk_job[0]));
@@ -2116,7 +2173,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                                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;
+                            ou_job[prev_buf].DV               = DV_pad;
                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
                        }

@@ -2134,7 +2191,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            qk_job[next_buf].s_tiles        = factx.vtcm_s_tiles[next_buf];
                            qk_job[next_buf].n_row_tiles    = n_row_tiles;
                            qk_job[next_buf].n_col_tiles    = hmx_ceil_div(next_rows, HMX_FP16_TILE_N_COLS);
-                            qk_job[next_buf].n_dot_tiles    = DK / 32;
+                            qk_job[next_buf].n_dot_tiles    = DK_pad / 32;
                            qk_job[next_buf].n_tiles_per_bc = n_tiles_per_bc;
                            qk_job[next_buf].hmx_scales     = factx.vtcm_hmx_scales_qk;
                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_qk_dot_worker, &qk_job[next_buf]));
@@ -2198,7 +2255,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                        ou_job[0].n_col_tiles      = last_cols;
                        ou_job[0].n_row_tiles_g_br = n_row_tiles_g_br;
                        ou_job[0].n_tiles_per_bc   = n_tiles_per_bc;
-                        ou_job[0].DV               = DV;
+                        ou_job[0].DV               = DV_pad;
                        hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[0]));

                        // Overlapped: run HVX build diag inv L while HMX is busy executing the update
@@ -2246,7 +2303,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            qk_job.s_tiles        = factx.vtcm_s_tiles[0];
                            qk_job.n_row_tiles    = n_row_tiles;
                            qk_job.n_col_tiles    = n_col_tiles;
-                            qk_job.n_dot_tiles    = (size_t) (DK / 32);
+                            qk_job.n_dot_tiles    = (size_t) (DK_pad / 32);
                            qk_job.n_tiles_per_bc = n_tiles_per_bc;
                            qk_job.hmx_scales     = factx.vtcm_hmx_scales_qk;

@@ -2302,7 +2359,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            ou_job.n_col_tiles      = n_col_tiles;
                            ou_job.n_row_tiles_g_br = n_row_tiles_g_br;
                            ou_job.n_tiles_per_bc   = n_tiles_per_bc;
-                            ou_job.DV               = DV;
+                            ou_job.DV               = DV_pad;

                            hmx_queue_push(ctx->hmx_queue, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job));
                            if (kv_blk + 1 == factx.n_kv_blocks) {
@@ -2380,7 +2437,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                    on_job.hmx_scales       = factx.vtcm_hmx_scales_id;
                    on_job.n_row_tiles      = n_row_tiles;
                    on_job.n_row_tiles_g_br = n_row_tiles_g_br;
-                    on_job.DV               = DV;
+                    on_job.DV               = DV_pad;
                    hmx_queue_push(ctx->hmx_queue, hmx_queue_make_desc(hmx_fa_o_norm_worker, &on_job));
                    hmx_queue_pop(ctx->hmx_queue);
                }
@@ -495,12 +495,140 @@ static inline void hmx_fa_q_prep_fp16(
 }


+// Head-dim-padded Q-prep (f32). Used when DK is not a multiple of 64.
+static inline void hmx_fa_q_prep_fp32_pad(__fp16 *                      vtcm_q_tiles,
+                                          const uint8_t *               temp_q_vtcm,
+                                          size_t                        start,
+                                          size_t                        end,
+                                          size_t                        g_rows_end,
+                                          size_t                        dk_in,
+                                          size_t                        dk_out,
+                                          size_t                        G,
+                                          size_t                        n_rows_q,
+                                          const struct fastdiv_values * div_G,
+                                          bool                          q_transposed) {
+    const uint32_t n_out_tiles = (uint32_t) (dk_out / 32);
+    for (size_t r = start; r < end; r += 2) {
+        size_t   r0       = r / HMX_FP16_TILE_N_ROWS;
+        size_t   r1       = r % HMX_FP16_TILE_N_ROWS;
+        __fp16 * out_base = vtcm_q_tiles + r0 * HMX_FP16_TILE_N_ROWS * dk_out;
+
+        if (r >= g_rows_end) {
+            for (uint32_t d = 0; d < n_out_tiles; ++d) {
+                ((HVX_Vector *) (out_base + d * HMX_FP16_TILE_N_ELMS))[r1 / 2] = Q6_V_vzero();
+            }
+            continue;
+        }
+
+        const size_t q_idx0 = fastdiv(r + 0, div_G);
+        const size_t h_idx0 = fastmodulo(r + 0, G, div_G);
+        const size_t q_idx1 = fastdiv(r + 1, div_G);
+        const size_t h_idx1 = fastmodulo(r + 1, G, div_G);
+
+        const size_t offset0 = q_transposed ? (h_idx0 * n_rows_q + q_idx0) : (q_idx0 * G + h_idx0);
+        const size_t offset1 = q_transposed ? (h_idx1 * n_rows_q + q_idx1) : (q_idx1 * G + h_idx1);
+
+        const HVX_UVector * pv_in0 = (const HVX_UVector *) (temp_q_vtcm + offset0 * dk_in * sizeof(float));
+        const HVX_UVector * pv_in1 = (r + 1 < g_rows_end) ? (const HVX_UVector *) (temp_q_vtcm + offset1 * dk_in * sizeof(float)) : NULL;
+
+        for (uint32_t d = 0; d < n_out_tiles; ++d) {
+            HVX_Vector * out_tile   = (HVX_Vector *) (out_base + d * HMX_FP16_TILE_N_ELMS);
+            const size_t base_lane  = (size_t) d * 32;
+            const size_t real_lanes = (base_lane < dk_in) ? hex_smin(32, dk_in - base_lane) : 0;
+
+            if (real_lanes == 0) {
+                out_tile[r1 / 2] = Q6_V_vzero();
+                continue;
+            }
+
+            HVX_Vector v0 = pv_in0[d];
+            HVX_Vector v1 = pv_in1 ? pv_in1[d] : Q6_V_vzero();
+            if (real_lanes < 32) {
+                // Straddle tile: keep the first real_lanes floats, zero the padded tail so
+                // the packed f16 lanes beyond DK are zero.
+                const HVX_VectorPred keep = Q6_Q_vsetq_R((uint32_t) (real_lanes * sizeof(float)));
+                v0                        = Q6_V_vmux_QVV(keep, v0, Q6_V_vzero());
+                v1                        = Q6_V_vmux_QVV(keep, v1, Q6_V_vzero());
+            }
+            out_tile[r1 / 2] = hvx_vec_f32_to_f16_shuff(v0, v1);
+        }
+    }
+}
+
+// Head-dim-padded Q-prep (f16). Used when DK is not a multiple of 64.
+static inline void hmx_fa_q_prep_fp16_pad(__fp16 *                      vtcm_q_tiles,
+                                          const uint8_t *               temp_q_vtcm,
+                                          size_t                        start,
+                                          size_t                        end,
+                                          size_t                        g_rows_end,
+                                          size_t                        dk_in,
+                                          size_t                        dk_out,
+                                          size_t                        G,
+                                          size_t                        n_rows_q,
+                                          const struct fastdiv_values * div_G,
+                                          bool                          q_transposed) {
+    const uint32_t n_out_pairs = (uint32_t) (dk_out / 64);
+    for (size_t r = start; r < end; r += 2) {
+        size_t   r0       = r / HMX_FP16_TILE_N_ROWS;
+        size_t   r1       = r % HMX_FP16_TILE_N_ROWS;
+        __fp16 * out_base = vtcm_q_tiles + r0 * HMX_FP16_TILE_N_ROWS * dk_out;
+
+        if (r >= g_rows_end) {
+            for (uint32_t d = 0; d < n_out_pairs; ++d) {
+                __fp16 *     out_dtile = out_base + d * HMX_FP16_TILE_N_ELMS * 2;
+                HVX_Vector * pv_out0   = ((HVX_Vector *) out_dtile) + r1 / 2;
+                HVX_Vector * pv_out1   = pv_out0 + 16;
+                *pv_out0               = Q6_V_vzero();
+                *pv_out1               = Q6_V_vzero();
+            }
+            continue;
+        }
+
+        const size_t q_idx0 = fastdiv(r + 0, div_G);
+        const size_t h_idx0 = fastmodulo(r + 0, G, div_G);
+        const size_t q_idx1 = fastdiv(r + 1, div_G);
+        const size_t h_idx1 = fastmodulo(r + 1, G, div_G);
+
+        const size_t offset0 = q_transposed ? (h_idx0 * n_rows_q + q_idx0) : (q_idx0 * G + h_idx0);
+        const size_t offset1 = q_transposed ? (h_idx1 * n_rows_q + q_idx1) : (q_idx1 * G + h_idx1);
+
+        const HVX_UVector * pv_in0 = (const HVX_UVector *) (temp_q_vtcm + offset0 * dk_in * sizeof(__fp16));
+        const HVX_UVector * pv_in1 = (r + 1 < g_rows_end) ? (const HVX_UVector *) (temp_q_vtcm + offset1 * dk_in * sizeof(__fp16)) : NULL;
+
+        for (uint32_t d = 0; d < n_out_pairs; ++d) {
+            __fp16 *     out_dtile = out_base + d * HMX_FP16_TILE_N_ELMS * 2;
+            HVX_Vector * pv_out0   = ((HVX_Vector *) out_dtile) + r1 / 2;
+            HVX_Vector * pv_out1   = pv_out0 + 16;
+
+            const size_t base_lane  = (size_t) d * 64;
+            const size_t real_lanes = (base_lane < dk_in) ? hex_smin(64, dk_in - base_lane) : 0;
+
+            if (real_lanes == 0) {
+                *pv_out0 = Q6_V_vzero();
+                *pv_out1 = Q6_V_vzero();
+                continue;
+            }
+
+            HVX_Vector v0 = pv_in0[d];
+            HVX_Vector v1 = pv_in1 ? pv_in1[d] : Q6_V_vzero();
+            if (real_lanes < 64) {
+                const HVX_VectorPred keep = Q6_Q_vsetq_R((uint32_t) (real_lanes * sizeof(__fp16)));
+                v0                        = Q6_V_vmux_QVV(keep, v0, Q6_V_vzero());
+                v1                        = Q6_V_vmux_QVV(keep, v1, Q6_V_vzero());
+            }
+            HVX_VectorPair vp = Q6_W_vshuff_VVR(v1, v0, -2);
+            *pv_out0          = Q6_V_lo_W(vp);
+            *pv_out1          = Q6_V_hi_W(vp);
+        }
+    }
+}
+
 static inline void hmx_fa_q_prep_fallback(
    __fp16 * vtcm_q_tiles, uintptr_t q_data,
    size_t q_nb1, size_t q_nb2, size_t q_nb3,
    uint32_t q_start, uint32_t kv_head, uint32_t ib3,
    size_t start, size_t end, size_t n_rows_g,
-    size_t G, size_t DK, bool is_q_fp32,
+    size_t G, size_t dk_in, size_t dk_out, bool is_q_fp32,
    const struct fastdiv_values * div_G
 ) {
    for (size_t r = start; r < end; r += 2) {
@@ -518,33 +646,55 @@ static inline void hmx_fa_q_prep_fallback(

        size_t   r0       = r / HMX_FP16_TILE_N_ROWS;
        size_t   r1       = r % HMX_FP16_TILE_N_ROWS;
-        __fp16 * out_base = vtcm_q_tiles + r0 * HMX_FP16_TILE_N_ROWS * DK;
+        __fp16 * out_base = vtcm_q_tiles + r0 * HMX_FP16_TILE_N_ROWS * dk_out;

        if (is_q_fp32) {
            const HVX_UVector * pv_in0 = q_ptr0 ? (const HVX_UVector *) q_ptr0 : NULL;
            const HVX_UVector * pv_in1 = q_ptr1 ? (const HVX_UVector *) q_ptr1 : NULL;

-            for (uint32_t d = 0; d < DK / 32; ++d) {
-                HVX_Vector v0   = pv_in0 ? pv_in0[d] : Q6_V_vzero();
-                HVX_Vector v1   = pv_in1 ? pv_in1[d] : Q6_V_vzero();
-                HVX_Vector v_hf = hvx_vec_f32_to_f16_shuff(v0, v1);
+            for (uint32_t d = 0; d < dk_out / 32; ++d) {
+                HVX_Vector * out_tile   = (HVX_Vector *) (out_base + d * HMX_FP16_TILE_N_ELMS);
+                const size_t base_lane  = (size_t) d * 32;
+                const size_t real_lanes = (base_lane < dk_in) ? hex_smin(32, dk_in - base_lane) : 0;

-                HVX_Vector * out_tile = (HVX_Vector *) (out_base + d * HMX_FP16_TILE_N_ELMS);
-                out_tile[r1 / 2]      = v_hf;
+                if (real_lanes == 0) {
+                    out_tile[r1 / 2] = Q6_V_vzero();
+                    continue;
+                }
+                HVX_Vector v0 = pv_in0 ? pv_in0[d] : Q6_V_vzero();
+                HVX_Vector v1 = pv_in1 ? pv_in1[d] : Q6_V_vzero();
+                if (real_lanes < 32) {
+                    const HVX_VectorPred keep = Q6_Q_vsetq_R((uint32_t) (real_lanes * sizeof(float)));
+                    v0 = Q6_V_vmux_QVV(keep, v0, Q6_V_vzero());
+                    v1 = Q6_V_vmux_QVV(keep, v1, Q6_V_vzero());
+                }
+                out_tile[r1 / 2] = hvx_vec_f32_to_f16_shuff(v0, v1);
            }
        } else {
            const HVX_UVector * pv_in0 = q_ptr0 ? (const HVX_UVector *) q_ptr0 : NULL;
            const HVX_UVector * pv_in1 = q_ptr1 ? (const HVX_UVector *) q_ptr1 : NULL;

-            for (uint32_t d = 0; d < DK / 64; ++d) {
-                HVX_Vector     v0 = pv_in0 ? pv_in0[d] : Q6_V_vzero();
-                HVX_Vector     v1 = pv_in1 ? pv_in1[d] : Q6_V_vzero();
-                HVX_VectorPair vp = Q6_W_vshuff_VVR(v1, v0, -2);
-
+            for (uint32_t d = 0; d < dk_out / 64; ++d) {
                __fp16 *     out_dtile = out_base + d * HMX_FP16_TILE_N_ELMS * 2;
                HVX_Vector * pv_out0   = ((HVX_Vector *) out_dtile) + r1 / 2;
                HVX_Vector * pv_out1   = pv_out0 + 16;

+                const size_t base_lane  = (size_t) d * 64;
+                const size_t real_lanes = (base_lane < dk_in) ? hex_smin(64, dk_in - base_lane) : 0;
+
+                if (real_lanes == 0) {
+                    *pv_out0 = Q6_V_vzero();
+                    *pv_out1 = Q6_V_vzero();
+                    continue;
+                }
+                HVX_Vector v0 = pv_in0 ? pv_in0[d] : Q6_V_vzero();
+                HVX_Vector v1 = pv_in1 ? pv_in1[d] : Q6_V_vzero();
+                if (real_lanes < 64) {
+                    const HVX_VectorPred keep = Q6_Q_vsetq_R((uint32_t) (real_lanes * sizeof(__fp16)));
+                    v0 = Q6_V_vmux_QVV(keep, v0, Q6_V_vzero());
+                    v1 = Q6_V_vmux_QVV(keep, v1, Q6_V_vzero());
+                }
+                HVX_VectorPair vp = Q6_W_vshuff_VVR(v1, v0, -2);
                *pv_out0 = Q6_V_lo_W(vp);
                *pv_out1 = Q6_V_hi_W(vp);
            }
@@ -175,5 +175,6 @@ int op_gated_delta_net(struct htp_ops_context * octx);
 int op_pad(struct htp_ops_context * octx);
 int op_im2col(struct htp_ops_context * octx);
 int op_allreduce(struct htp_ops_context * octx);
+int op_roll(struct htp_ops_context * octx);

 #endif /* HTP_CTX_H */
@@ -104,6 +104,7 @@ enum htp_op_code {
    HTP_OP_ALLREDUCE_ADD,
    HTP_OP_GLU_SWIGLU_CLAMP,
    HTP_OP_MDEV_GROUP,
+    HTP_OP_ROLL,

    HTP_OP_INVALID
 };
@@ -25,17 +25,20 @@ struct htp_im2col_context {
    uint32_t                 npatches;             // number of patches assigned to this dev
    uint32_t                 npatches_per_thread;  // patches = N*OH*OW (pure-DDR kernel)

-    uint32_t pe_row_base;              // first N*OH row index assigned to this dev (DMA path)
-    uint32_t pe_nrows;                 // number of N*OH rows assigned to this dev (DMA path)
-    uint32_t pe_rows_per_thread;       // N*OH rows per worker
-    uint32_t pe_src_row_bytes;         // one output row's source: IC*KH*IW*4, rounded 256
-    uint32_t pe_dst_row_bytes;         // one output row's dst: OW*patch_stride*2, rounded 256
+    uint32_t pe_row_base;                          // first N*OH row index assigned to this dev (DMA path)
+    uint32_t pe_nrows;                             // number of N*OH rows assigned to this dev (DMA path)
+    uint32_t pe_rows_per_thread;                   // N*OH rows per worker
+    uint32_t pe_src_row_bytes;                     // one output row's source: IC*KH*IW*4, rounded 256
+    uint32_t pe_dst_row_bytes;                     // one output row's dst: OW*patch_stride*2, rounded 256

    // Patch-embed DMA path VTCM ping-pong.
    uint8_t * pe_vtcm_src;                         // base of the 2x src buffers region
    uint8_t * pe_vtcm_dst;                         // base of the 2x dst buffers region
    uint32_t  pe_src_size_per_thread;              // 2 * pe_src_row_bytes
    uint32_t  pe_dst_size_per_thread;              // 2 * pe_dst_row_bytes
+
+    uint32_t pe_owb;                               // output-col block size
+    uint32_t pe_wb;                                // staged source window width
 };

 // Per-op VTCM layout for the patch-embed DMA path
@@ -59,83 +62,253 @@ static inline void htp_im2col_vtcm_layout_build(struct htp_im2col_vtcm_layout *
    L->total_bytes = L->off_dst + L->dst_bytes_per_thread * n_threads;
 }

-#define IM2COL_PATCHEMBED_BODY(FNAME, DST_CTYPE, COPY_FN, SPLAT_FN, DST_ELEM, TAG)                        \
-    static void FNAME(unsigned int nth, unsigned int ith, void * data) {                                  \
-        struct htp_im2col_context * ictx        = (struct htp_im2col_context *) data;                     \
-        struct htp_ops_context *    octx        = ictx->octx;                                             \
-        struct htp_thread_trace * restrict tr   = &octx->ctx->trace[ith];                                 \
-        const struct htp_tensor * restrict src0 = octx->src[0];                                           \
-        const struct htp_tensor * restrict src1 = octx->src[1];                                           \
-        const struct htp_tensor * restrict dst  = octx->dst;                                              \
-        const int32_t s0 = octx->op_params[0], s1 = octx->op_params[1];                                   \
-        const int32_t p0 = octx->op_params[2], p1 = octx->op_params[3];                                   \
-        const int32_t d0 = octx->op_params[4], d1 = octx->op_params[5];                                   \
-        const uint32_t N = src1->ne[3], IC = src1->ne[2], IH = src1->ne[1], IW = src1->ne[0];             \
-        const uint32_t KH                       = src0->ne[1], KW = src0->ne[0];                          \
-        const uint32_t OH                       = dst->ne[2];                                             \
-        const uint32_t OW                       = dst->ne[1];                                             \
-        const uint32_t patch_stride             = IC * KH * KW;                                           \
-        const float * restrict src_data         = (const float *) src1->data;                             \
-        DST_CTYPE * restrict dst_data           = (DST_CTYPE *) dst->data;                                \
-        const uint32_t patch_end                = ictx->patch_base + ictx->npatches;                      \
-        const uint32_t patch_start              = ictx->patch_base + ictx->npatches_per_thread * ith;     \
-        const uint32_t patch_stop               = MIN(patch_start + ictx->npatches_per_thread, patch_end);\
-        if (patch_start >= patch_stop) {                                                                  \
-            return;                                                                                       \
-        }                                                                                                 \
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, patch_start);                                   \
-        for (uint32_t p = patch_start; p < patch_stop; p++) {                                             \
-            const uint32_t iow             = p % OW;                                                      \
-            const uint32_t ioh             = (p / OW) % OH;                                               \
-            const uint32_t in              = p / (OW * OH);                                               \
-            DST_CTYPE * restrict dst_patch = dst_data + (uint64_t) p * patch_stride;                      \
-            for (uint32_t iic = 0; iic < IC; iic++) {                                                     \
-                const float * restrict src_plane = src_data + ((uint64_t) in * IC + iic) * IH * IW;       \
-                for (uint32_t ikh = 0; ikh < KH; ikh++) {                                                 \
-                    const int32_t iih            = (int32_t) ioh * s1 + (int32_t) ikh * d1 - p1;          \
-                    DST_CTYPE * restrict out_run = dst_patch + iic * (KH * KW) + ikh * KW;                \
-                    if (iih < 0 || iih >= (int32_t) IH) {                                                 \
-                        SPLAT_FN(out_run, 0.0f, KW);                                                      \
-                        continue;                                                                         \
-                    }                                                                                     \
-                    const int32_t iiw0             = (int32_t) iow * s0 - p0;                             \
-                    const float * restrict src_run = src_plane + (uint64_t) iih * IW + iiw0;              \
-                    if (d0 == 1) {                                                                        \
-                        /* contiguous source run: [lo,hi) is in-bounds, tails are zero pad */             \
-                        const int32_t lo = iiw0 < 0 ? -iiw0 : 0;                                          \
-                        int32_t       hi = (int32_t) IW - iiw0;                                           \
-                        if (hi > (int32_t) KW) {                                                          \
-                            hi = (int32_t) KW;                                                            \
-                        }                                                                                 \
-                        if (hi <= lo) {                                                                   \
-                            SPLAT_FN(out_run, 0.0f, KW);                                                  \
-                        } else {                                                                          \
-                            if (lo > 0) {                                                                 \
-                                SPLAT_FN(out_run, 0.0f, (uint32_t) lo);                                   \
-                            }                                                                             \
-                            COPY_FN((uint8_t *) (out_run + lo), (const uint8_t *) (src_run + lo),         \
-                                    (uint32_t) (hi - lo));                                                \
-                            if (hi < (int32_t) KW) {                                                      \
-                                SPLAT_FN(out_run + hi, 0.0f, (KW - (uint32_t) hi));                       \
-                            }                                                                             \
-                        }                                                                                 \
-                        continue;                                                                         \
-                    }                                                                                     \
-                    for (uint32_t ikw = 0; ikw < KW; ikw++) {                                             \
-                        const int32_t iiw = (int32_t) iow * s0 + (int32_t) ikw * d0 - p0;                 \
-                        out_run[ikw]      = (iiw < 0 || iiw >= (int32_t) IW) ?                            \
-                                                (DST_CTYPE) 0.0f :                                        \
-                                                (DST_CTYPE) src_plane[(uint64_t) iih * IW + iiw];         \
-                    }                                                                                     \
-                }                                                                                         \
-            }                                                                                             \
-        }                                                                                                 \
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, patch_start);                                    \
+#define IM2COL_PATCHEMBED_BODY(FNAME, DST_CTYPE, COPY_FN, SPLAT_FN, DST_ELEM, TAG)                         \
+    static void FNAME(unsigned int nth, unsigned int ith, void * data) {                                   \
+        struct htp_im2col_context * ictx        = (struct htp_im2col_context *) data;                      \
+        struct htp_ops_context *    octx        = ictx->octx;                                              \
+        struct htp_thread_trace * restrict tr   = &octx->ctx->trace[ith];                                  \
+        const struct htp_tensor * restrict src0 = octx->src[0];                                            \
+        const struct htp_tensor * restrict src1 = octx->src[1];                                            \
+        const struct htp_tensor * restrict dst  = octx->dst;                                               \
+        const int32_t  s0                       = octx->op_params[0];                                      \
+        const int32_t  s1                       = octx->op_params[1];                                      \
+        const int32_t  p0                       = octx->op_params[2];                                      \
+        const int32_t  p1                       = octx->op_params[3];                                      \
+        const int32_t  d0                       = octx->op_params[4];                                      \
+        const int32_t  d1                       = octx->op_params[5];                                      \
+        const int32_t  is_2D                    = octx->op_params[6] == 1;                                 \
+        const uint32_t N                        = is_2D ? src1->ne[3] : src1->ne[2];                       \
+        const uint32_t IC                       = is_2D ? src1->ne[2] : src1->ne[1];                       \
+        const uint32_t IH                       = is_2D ? src1->ne[1] : 1;                                 \
+        const uint32_t IW                       = src1->ne[0];                                             \
+        const uint32_t KH                       = is_2D ? src0->ne[1] : 1;                                 \
+        const uint32_t KW                       = src0->ne[0];                                             \
+        const uint32_t OH                       = is_2D ? dst->ne[2] : 1;                                  \
+        const uint32_t OW                       = dst->ne[1];                                              \
+        const uint32_t patch_stride             = IC * KH * KW;                                            \
+        const float * restrict src_data         = (const float *) src1->data;                              \
+        DST_CTYPE * restrict dst_data           = (DST_CTYPE *) dst->data;                                 \
+        const uint32_t patch_end                = ictx->patch_base + ictx->npatches;                       \
+        const uint32_t patch_start              = ictx->patch_base + ictx->npatches_per_thread * ith;      \
+        const uint32_t patch_stop               = MIN(patch_start + ictx->npatches_per_thread, patch_end); \
+        if (patch_start >= patch_stop) {                                                                   \
+            return;                                                                                        \
+        }                                                                                                  \
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, patch_start);                                    \
+        for (uint32_t p = patch_start; p < patch_stop; p++) {                                              \
+            const uint32_t iow             = p % OW;                                                       \
+            const uint32_t ioh             = (p / OW) % OH;                                                \
+            const uint32_t in              = p / (OW * OH);                                                \
+            DST_CTYPE * restrict dst_patch = dst_data + (uint64_t) p * patch_stride;                       \
+            for (uint32_t iic = 0; iic < IC; iic++) {                                                      \
+                const float * restrict src_plane = src_data + ((uint64_t) in * IC + iic) * IH * IW;        \
+                for (uint32_t ikh = 0; ikh < KH; ikh++) {                                                  \
+                    const int32_t iih            = (int32_t) ioh * s1 + (int32_t) ikh * d1 - p1;           \
+                    DST_CTYPE * restrict out_run = dst_patch + iic * (KH * KW) + ikh * KW;                 \
+                    if (iih < 0 || iih >= (int32_t) IH) {                                                  \
+                        SPLAT_FN(out_run, 0.0f, KW);                                                       \
+                        continue;                                                                          \
+                    }                                                                                      \
+                    const int32_t iiw0             = (int32_t) iow * s0 - p0;                              \
+                    const float * restrict src_run = src_plane + (uint64_t) iih * IW + iiw0;               \
+                    if (d0 == 1) {                                                                         \
+                        /* contiguous source run: [lo,hi) is in-bounds, tails are zero pad */              \
+                        const int32_t lo = iiw0 < 0 ? -iiw0 : 0;                                           \
+                        int32_t       hi = (int32_t) IW - iiw0;                                            \
+                        if (hi > (int32_t) KW) {                                                           \
+                            hi = (int32_t) KW;                                                             \
+                        }                                                                                  \
+                        if (hi <= lo) {                                                                    \
+                            SPLAT_FN(out_run, 0.0f, KW);                                                   \
+                        } else {                                                                           \
+                            if (lo > 0) {                                                                  \
+                                SPLAT_FN(out_run, 0.0f, (uint32_t) lo);                                    \
+                            }                                                                              \
+                            COPY_FN((uint8_t *) (out_run + lo), (const uint8_t *) (src_run + lo),          \
+                                    (uint32_t) (hi - lo));                                                 \
+                            if (hi < (int32_t) KW) {                                                       \
+                                SPLAT_FN(out_run + hi, 0.0f, (KW - (uint32_t) hi));                        \
+                            }                                                                              \
+                        }                                                                                  \
+                        continue;                                                                          \
+                    }                                                                                      \
+                    for (uint32_t ikw = 0; ikw < KW; ikw++) {                                              \
+                        const int32_t iiw = (int32_t) iow * s0 + (int32_t) ikw * d0 - p0;                  \
+                        out_run[ikw]      = (iiw < 0 || iiw >= (int32_t) IW) ?                             \
+                                                (DST_CTYPE) 0.0f :                                         \
+                                                (DST_CTYPE) src_plane[(uint64_t) iih * IW + iiw];          \
+                    }                                                                                      \
+                }                                                                                          \
+            }                                                                                              \
+        }                                                                                                  \
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, patch_start);                                     \
    }

 IM2COL_PATCHEMBED_BODY(im2col_patchembed_thread, __fp16, hvx_copy_f16_f32_uu, hvx_splat_f16_u, sizeof(__fp16), "f32-f16")
 IM2COL_PATCHEMBED_BODY(im2col_patchembed_f32_thread, float, hvx_copy_f32_uu, hvx_splat_f32_u, sizeof(float), "f32-f32")

+// Software-pipelined 2-deep: while HVX computes block bi from buffer slot
+// (bi&1), the DMA engine stages block bi+1 into the other slot concurrently.
+// A single dma_queue_flush per iteration (after issuing the next stage-in and
+// this block's store-out) waits for both - safe because the ring is strict
+// FIFO and each buffer slot is only reused after its prior consumer (compute
+// or store-out) already finished in program order.
+#define IM2COL_BLOCKED_DMA_BODY(FNAME, DST_CTYPE, COPY_FN, SPLAT_FN, DST_ELEM, TAG)                                  \
+    static void FNAME(unsigned int nth, unsigned int ith, void * data) {                                             \
+        struct htp_im2col_context * ictx        = (struct htp_im2col_context *) data;                                \
+        struct htp_ops_context *    octx        = ictx->octx;                                                        \
+        struct htp_thread_trace * restrict tr   = &octx->ctx->trace[ith];                                            \
+        const struct htp_tensor * restrict src1 = octx->src[1];                                                      \
+        const struct htp_tensor * restrict dst  = octx->dst;                                                         \
+        const int32_t  s0 = octx->op_params[0], s1 = octx->op_params[1];                                             \
+        const int32_t  p0 = octx->op_params[2], p1 = octx->op_params[3];                                             \
+        const int32_t  d0 = octx->op_params[4], d1 = octx->op_params[5];                                             \
+        const int32_t  is_2D = octx->op_params[6] == 1;                                                              \
+        const uint32_t N     = is_2D ? src1->ne[3] : src1->ne[2];                                                    \
+        const uint32_t IC    = is_2D ? src1->ne[2] : src1->ne[1];                                                    \
+        const uint32_t IH    = is_2D ? src1->ne[1] : 1;                                                              \
+        const uint32_t IW    = src1->ne[0];                                                                          \
+        const uint32_t KH    = is_2D ? octx->src[0]->ne[1] : 1;                                                      \
+        const uint32_t KW    = octx->src[0]->ne[0];                                                                  \
+        const uint32_t OH    = is_2D ? dst->ne[2] : 1;                                                               \
+        const uint32_t OW    = dst->ne[1];                                                                           \
+        const uint32_t owb = ictx->pe_owb, Wb = ictx->pe_wb;                                                         \
+        const uint32_t patch_stride     = IC * KH * KW;                                                              \
+        const float * restrict src_data = (const float *) src1->data;                                                \
+        DST_CTYPE * restrict dst_data   = (DST_CTYPE *) dst->data;                                                   \
+        dma_queue *    dmaq             = octx->ctx->dma[ith];                                                       \
+        uint8_t *      srcb_base        = ictx->pe_vtcm_src + ith * ictx->pe_src_size_per_thread;                    \
+        uint8_t *      dstb_base        = ictx->pe_vtcm_dst + ith * ictx->pe_dst_size_per_thread;                    \
+        float *        srcb2[2]         = { (float *) srcb_base, (float *) (srcb_base + ictx->pe_src_row_bytes) };   \
+        DST_CTYPE *    dstb2[2]   = { (DST_CTYPE *) dstb_base, (DST_CTYPE *) (dstb_base + ictx->pe_dst_row_bytes) }; \
+        const uint32_t nrows      = N * OH;                                                                          \
+        const uint32_t per_thread = ictx->pe_rows_per_thread;                                                        \
+        const uint32_t row_start  = per_thread * ith;                                                                \
+        const uint32_t row_end    = MIN(row_start + per_thread, nrows);                                              \
+        if (row_start >= row_end)                                                                                    \
+            return;                                                                                                  \
+        const uint32_t nbpr         = (OW + owb - 1) / owb;                                                          \
+        const uint32_t nrows_local  = row_end - row_start;                                                           \
+        const uint32_t total_blocks = nrows_local * nbpr;                                                            \
+        for (uint32_t bi = 0; bi < total_blocks; bi++) {                                                             \
+            const uint32_t buf  = bi & 1u;                                                                           \
+            float *        srcb = srcb2[buf];                                                                        \
+            DST_CTYPE *    dstb = dstb2[buf];                                                                        \
+            const uint32_t r    = row_start + bi / nbpr;                                                             \
+            const uint32_t in   = r / OH;                                                                            \
+            const uint32_t ioh  = r % OH;                                                                            \
+            const uint32_t c0   = (bi % nbpr) * owb;                                                                 \
+            const uint32_t nb   = MIN(owb, OW - c0);                                                                 \
+            const int32_t  win0 = (int32_t) c0 * s0 - p0;                                                            \
+            if (bi == 0) {                                                                                           \
+                /* prologue: stage block 0 and wait - nothing to overlap with yet */                                 \
+                for (uint32_t ikh = 0; ikh < KH; ikh++) {                                                            \
+                    const int32_t iih = (int32_t) ioh * s1 + (int32_t) ikh * d1 - p1;                                \
+                    if (iih < 0 || iih >= (int32_t) IH)                                                              \
+                        continue;                                                                                    \
+                    const int32_t lo = win0 < 0 ? -win0 : 0;                                                         \
+                    int32_t       hi = (int32_t) IW - win0;                                                          \
+                    if (hi > (int32_t) Wb)                                                                           \
+                        hi = (int32_t) Wb;                                                                           \
+                    if (hi <= lo)                                                                                    \
+                        continue;                                                                                    \
+                    const uint32_t cpw  = (uint32_t) (hi - lo);                                                      \
+                    float *        vdst = srcb + (uint64_t) ikh * Wb + (uint32_t) lo;                                \
+                    const float *  vsrc = src_data + ((uint64_t) (in * IC) * IH + iih) * IW + (win0 + lo);           \
+                    while (!dma_queue_push(dmaq, dma_make_ptr((uint8_t *) vdst, (const uint8_t *) vsrc),             \
+                                           (size_t) KH * Wb * sizeof(float), (size_t) IH * IW * sizeof(float),       \
+                                           cpw * sizeof(float), IC)) {                                               \
+                        dma_queue_pop(dmaq);                                                                         \
+                    }                                                                                                \
+                }                                                                                                    \
+                dma_queue_flush(dmaq);                                                                               \
+            }                                                                                                        \
+            if (bi + 1 < total_blocks) {                                                                             \
+                /* prefetch: stage block bi+1 into the other slot; overlaps with this block's compute below */       \
+                const uint32_t nbuf  = 1u - buf;                                                                     \
+                float *        nsrcb = srcb2[nbuf];                                                                  \
+                const uint32_t nr    = row_start + (bi + 1) / nbpr;                                                  \
+                const uint32_t nin   = nr / OH;                                                                      \
+                const uint32_t nioh  = nr % OH;                                                                      \
+                const uint32_t nc0   = ((bi + 1) % nbpr) * owb;                                                      \
+                const int32_t  nwin0 = (int32_t) nc0 * s0 - p0;                                                      \
+                for (uint32_t ikh = 0; ikh < KH; ikh++) {                                                            \
+                    const int32_t iih = (int32_t) nioh * s1 + (int32_t) ikh * d1 - p1;                               \
+                    if (iih < 0 || iih >= (int32_t) IH)                                                              \
+                        continue;                                                                                    \
+                    const int32_t lo = nwin0 < 0 ? -nwin0 : 0;                                                       \
+                    int32_t       hi = (int32_t) IW - nwin0;                                                         \
+                    if (hi > (int32_t) Wb)                                                                           \
+                        hi = (int32_t) Wb;                                                                           \
+                    if (hi <= lo)                                                                                    \
+                        continue;                                                                                    \
+                    const uint32_t cpw  = (uint32_t) (hi - lo);                                                      \
+                    float *        vdst = nsrcb + (uint64_t) ikh * Wb + (uint32_t) lo;                               \
+                    const float *  vsrc = src_data + ((uint64_t) (nin * IC) * IH + iih) * IW + (nwin0 + lo);         \
+                    while (!dma_queue_push(dmaq, dma_make_ptr((uint8_t *) vdst, (const uint8_t *) vsrc),             \
+                                           (size_t) KH * Wb * sizeof(float), (size_t) IH * IW * sizeof(float),       \
+                                           cpw * sizeof(float), IC)) {                                               \
+                        dma_queue_pop(dmaq);                                                                         \
+                    }                                                                                                \
+                }                                                                                                    \
+            }                                                                                                        \
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, r);                                                    \
+            for (uint32_t j = 0; j < nb; j++) {                                                                      \
+                const uint32_t iow       = c0 + j;                                                                   \
+                DST_CTYPE *    dst_patch = dstb + (uint64_t) j * patch_stride;                                       \
+                const int32_t  iiw0      = (int32_t) iow * s0 - p0;                                                  \
+                for (uint32_t ikh = 0; ikh < KH; ikh++) {                                                            \
+                    const int32_t iih = (int32_t) ioh * s1 + (int32_t) ikh * d1 - p1;                                \
+                    const int     okh = (iih >= 0 && iih < (int32_t) IH);                                            \
+                    for (uint32_t iic = 0; iic < IC; iic++) {                                                        \
+                        DST_CTYPE * out_run = dst_patch + iic * (KH * KW) + ikh * KW;                                \
+                        if (!okh) {                                                                                  \
+                            SPLAT_FN(out_run, 0.0f, KW);                                                             \
+                            continue;                                                                                \
+                        }                                                                                            \
+                        const float * vrow = srcb + ((uint64_t) (iic * KH + ikh)) * Wb; /* col win0 at idx 0*/       \
+                        if (d0 == 1) {                                                                               \
+                            /* contiguous run within the staged window: [lo,hi) in-bounds, tails zero pad */         \
+                            const int32_t lo = iiw0 < 0 ? -iiw0 : 0;                                                 \
+                            int32_t       hi = (int32_t) IW - iiw0;                                                  \
+                            if (hi > (int32_t) KW) {                                                                 \
+                                hi = (int32_t) KW;                                                                   \
+                            }                                                                                        \
+                            if (hi <= lo) {                                                                          \
+                                SPLAT_FN(out_run, 0.0f, KW);                                                         \
+                            } else {                                                                                 \
+                                if (lo > 0) {                                                                        \
+                                    SPLAT_FN(out_run, 0.0f, (uint32_t) lo);                                          \
+                                }                                                                                    \
+                                COPY_FN((uint8_t *) (out_run + lo), (const uint8_t *) (vrow + (iiw0 + lo - win0)),   \
+                                        (uint32_t) (hi - lo));                                                       \
+                                if (hi < (int32_t) KW) {                                                             \
+                                    SPLAT_FN(out_run + hi, 0.0f, (KW - (uint32_t) hi));                              \
+                                }                                                                                    \
+                            }                                                                                        \
+                            continue;                                                                                \
+                        }                                                                                            \
+                        for (uint32_t ikw = 0; ikw < KW; ikw++) {                                                    \
+                            const int32_t iiw = iiw0 + (int32_t) ikw * d0;                                           \
+                            out_run[ikw] =                                                                           \
+                                (iiw < 0 || iiw >= (int32_t) IW) ? (DST_CTYPE) 0.0f : (DST_CTYPE) vrow[iiw - win0];  \
+                        }                                                                                            \
+                    }                                                                                                \
+                }                                                                                                    \
+            }                                                                                                        \
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, r);                                                     \
+            DST_CTYPE * ddr = dst_data + ((uint64_t) (in * OH + ioh) * OW + c0) * patch_stride;                      \
+            dma_queue_push_vtcm_to_ddr(dmaq, dma_make_ptr((uint8_t *) ddr, (uint8_t *) dstb),                        \
+                                       nb * patch_stride * (DST_ELEM), nb * patch_stride * (DST_ELEM), 1);           \
+            dma_queue_flush(dmaq);                                                                                   \
+        }                                                                                                            \
+    }
+IM2COL_BLOCKED_DMA_BODY(im2col_blocked_dma_thread,     __fp16, hvx_copy_f16_f32_uu, hvx_splat_f16_u, sizeof(__fp16), "blk-dma-f16")
+IM2COL_BLOCKED_DMA_BODY(im2col_blocked_dma_f32_thread, float,  hvx_copy_f32_uu,     hvx_splat_f32_u, sizeof(float),  "blk-dma-f32")
+
+// Exact-tiling patch-embed DMA fast path (s0==KW, p0=0, d0=1; and 2D s1==KH,
+// p1=0, d1=1). Intentionally reads no stride/pad/dilation params so the inner
+// copy stays tight and fully hoisted - do NOT graft the general gather in here.
 #define IM2COL_PATCHEMBED_DMA_BODY(FNAME, DST_CTYPE, COPY_FN, SPLAT_FN, DST_ELEM, TAG)                               \
    static void FNAME(unsigned int nth, unsigned int ith, void * data) {                                             \
        struct htp_im2col_context * ictx        = (struct htp_im2col_context *) data;                                \
@@ -143,21 +316,27 @@ IM2COL_PATCHEMBED_BODY(im2col_patchembed_f32_thread, float, hvx_copy_f32_uu, hvx
        struct htp_thread_trace * restrict tr   = &octx->ctx->trace[ith];                                            \
        const struct htp_tensor * restrict src1 = octx->src[1];                                                      \
        const struct htp_tensor * restrict dst  = octx->dst;                                                         \
-        const uint32_t N = src1->ne[3], IC = src1->ne[2], IH = src1->ne[1], IW = src1->ne[0];                        \
-        const uint32_t KH = octx->src[0]->ne[1], KW = octx->src[0]->ne[0];                                           \
-        const uint32_t OH = dst->ne[2], OW = dst->ne[1];                                                             \
-        const uint32_t patch_stride     = IC * KH * KW;                                                              \
-        const float * restrict src_data = (const float *) src1->data;                                                \
-        DST_CTYPE * restrict dst_data   = (DST_CTYPE *) dst->data;                                                   \
-        dma_queue *    dmaq             = octx->ctx->dma[ith];                                                       \
-        uint8_t *      src_base         = ictx->pe_vtcm_src + ith * ictx->pe_src_size_per_thread;                    \
-        uint8_t *      dst_base         = ictx->pe_vtcm_dst + ith * ictx->pe_dst_size_per_thread;                    \
-        float *        srcb             = (float *) src_base;                                                        \
-        DST_CTYPE *    dstb             = (DST_CTYPE *) dst_base;                                                    \
-        const uint32_t row_end_max      = ictx->pe_row_base + ictx->pe_nrows;                                        \
-        const uint32_t per_thread       = ictx->pe_rows_per_thread;                                                  \
-        const uint32_t row_start        = ictx->pe_row_base + per_thread * ith;                                      \
-        const uint32_t row_end          = MIN(row_start + per_thread, row_end_max);                                  \
+        const int32_t  is_2D                    = octx->op_params[6] == 1;                                           \
+        const uint32_t N                        = is_2D ? src1->ne[3] : src1->ne[2];                                 \
+        const uint32_t IC                       = is_2D ? src1->ne[2] : src1->ne[1];                                 \
+        const uint32_t IH                       = is_2D ? src1->ne[1] : 1;                                           \
+        const uint32_t IW                       = src1->ne[0];                                                       \
+        const uint32_t KH                       = is_2D ? octx->src[0]->ne[1] : 1;                                   \
+        const uint32_t KW                       = octx->src[0]->ne[0];                                               \
+        const uint32_t OH                       = is_2D ? dst->ne[2] : 1;                                            \
+        const uint32_t OW                       = dst->ne[1];                                                        \
+        const uint32_t patch_stride             = IC * KH * KW;                                                      \
+        const float * restrict src_data         = (const float *) src1->data;                                        \
+        DST_CTYPE * restrict dst_data           = (DST_CTYPE *) dst->data;                                           \
+        dma_queue *    dmaq                     = octx->ctx->dma[ith];                                               \
+        uint8_t *      src_base                 = ictx->pe_vtcm_src + ith * ictx->pe_src_size_per_thread;            \
+        uint8_t *      dst_base                 = ictx->pe_vtcm_dst + ith * ictx->pe_dst_size_per_thread;            \
+        float *        srcb                     = (float *) src_base;                                                \
+        DST_CTYPE *    dstb                     = (DST_CTYPE *) dst_base;                                            \
+        const uint32_t row_end_max              = ictx->pe_row_base + ictx->pe_nrows;                                \
+        const uint32_t per_thread               = ictx->pe_rows_per_thread;                                          \
+        const uint32_t row_start                = ictx->pe_row_base + per_thread * ith;                              \
+        const uint32_t row_end                  = MIN(row_start + per_thread, row_end_max);                          \
        if (row_start >= row_end)                                                                                    \
            return;                                                                                                  \
        for (uint32_t r = row_start; r < row_end; r++) {                                                             \
@@ -209,21 +388,30 @@ static bool im2col_use_patchembed_dma(const struct htp_ops_context * octx) {
    const int32_t p0 = octx->op_params[2], p1 = octx->op_params[3];
    const int32_t d0 = octx->op_params[4], d1 = octx->op_params[5];
    const int     is_2D = octx->op_params[6] == 1;
-    if (!is_2D) {
-        return false;
-    }
    if (octx->dst->type != HTP_TYPE_F16 && octx->dst->type != HTP_TYPE_F32) {
        return false;
    }
-    const uint32_t KH = octx->src[0]->ne[1], KW = octx->src[0]->ne[0];
-    if (s0 != (int32_t) KW || s1 != (int32_t) KH) {
-        return false;  // non-overlapping
+    const uint32_t KH = is_2D ? octx->src[0]->ne[1] : 1;
+    const uint32_t KW = octx->src[0]->ne[0];
+    if (s0 != (int32_t) KW) {
+        return false;  // non-overlapping (width)
    }
-    if (p0 != 0 || p1 != 0) {
-        return false;  // no padding
+    if (p0 != 0) {
+        return false;  // no padding (width)
    }
-    if (d0 != 1 || d1 != 1) {
-        return false;  // no dilation
+    if (d0 != 1) {
+        return false;  // no dilation (width)
+    }
+    if (is_2D) {
+        if (s1 != (int32_t) KH) {
+            return false;  // non-overlapping (height)
+        }
+        if (p1 != 0) {
+            return false;  // no padding (height)
+        }
+        if (d1 != 1) {
+            return false;  // no dilation (height)
+        }
    }
    return true;
 }
@@ -233,8 +421,11 @@ static bool im2col_use_patchembed_dma(const struct htp_ops_context * octx) {
 static bool im2col_patchembed_dma_fits(struct htp_ops_context *    octx,
                                       struct htp_im2col_context * ictx,
                                       uint32_t                    n_threads) {
-    const uint32_t IC = octx->src[1]->ne[2], IW = octx->src[1]->ne[0];
-    const uint32_t KH = octx->src[0]->ne[1], KW = octx->src[0]->ne[0];
+    const int32_t  is_2D = octx->op_params[6] == 1;
+    const uint32_t IC = is_2D ? octx->src[1]->ne[2] : octx->src[1]->ne[1];
+    const uint32_t IW = octx->src[1]->ne[0];
+    const uint32_t KH = is_2D ? octx->src[0]->ne[1] : 1;
+    const uint32_t KW = octx->src[0]->ne[0];
    const uint32_t OW           = octx->dst->ne[1];
    const uint32_t patch_stride = IC * KH * KW;

@@ -257,6 +448,45 @@ static bool im2col_patchembed_dma_fits(struct htp_ops_context *    octx,
    return true;
 }

+// Sizes a per-thread 2x(src,dst) VTCM ping-pong for the blocked general kernel.
+// Stages Wb=(owb-1)*s0+(KW-1)*d0+1 source cols per (iic,ikh) row and owb patches
+// of dst. Picks the largest owb that fits; returns false if even owb=1 does not.
+static bool im2col_blocked_dma_fits(struct htp_ops_context *    octx,
+                                    struct htp_im2col_context * ictx,
+                                    uint32_t                    n_threads) {
+    const int32_t  is_2D = octx->op_params[6] == 1;
+    const int32_t  s0 = octx->op_params[0];
+    const int32_t  d0 = octx->op_params[4];
+    const uint32_t IC = is_2D ? octx->src[1]->ne[2] : octx->src[1]->ne[1];
+    const uint32_t KH = is_2D ? octx->src[0]->ne[1] : 1;
+    const uint32_t KW = octx->src[0]->ne[0];
+    const uint32_t OW = octx->dst->ne[1];
+    const uint32_t patch_stride = IC * KH * KW;
+    const uint32_t dst_elem = (octx->dst->type == HTP_TYPE_F16) ? sizeof(__fp16) : sizeof(float);
+
+    for (uint32_t owb = (OW < 256 ? OW : 256); owb >= 1; owb--) {
+        const uint32_t Wb = (owb - 1) * (uint32_t) s0 + (KW - 1) * (uint32_t) d0 + 1;
+        const uint32_t src_row_bytes = hex_round_up(IC * KH * Wb * sizeof(float), 256);
+        const uint32_t dst_row_bytes = hex_round_up(owb * patch_stride * dst_elem, 256);
+        struct htp_im2col_vtcm_layout L;
+        htp_im2col_vtcm_layout_build(&L, src_row_bytes, dst_row_bytes, n_threads);
+        if (L.total_bytes <= octx->ctx->vtcm_size) {
+            uint8_t * const base = octx->ctx->vtcm_base;
+            ictx->pe_owb                 = owb;
+            ictx->pe_wb                  = Wb;
+            ictx->pe_src_row_bytes       = src_row_bytes;
+            ictx->pe_dst_row_bytes       = dst_row_bytes;
+            ictx->pe_vtcm_src            = VTCM_LAYOUT_PTR(uint8_t, base, L.off_src);
+            ictx->pe_vtcm_dst            = VTCM_LAYOUT_PTR(uint8_t, base, L.off_dst);
+            ictx->pe_src_size_per_thread = (uint32_t) L.src_bytes_per_thread;
+            ictx->pe_dst_size_per_thread = (uint32_t) L.dst_bytes_per_thread;
+            return true;
+        }
+        if (owb == 1) break;  // avoid unsigned underflow
+    }
+    return false;
+}
+
 int op_im2col(struct htp_ops_context * octx) {
    const struct htp_tensor * src1 = octx->src[1];
    const struct htp_tensor * dst  = octx->dst;
@@ -270,8 +500,9 @@ int op_im2col(struct htp_ops_context * octx) {
        return HTP_STATUS_OK;
    }

-    const uint32_t N             = src1->ne[3];
-    const uint32_t OH            = dst->ne[2];
+    const int32_t  is_2D         = octx->op_params[6] == 1;
+    const uint32_t N             = is_2D ? src1->ne[3] : src1->ne[2];
+    const uint32_t OH            = is_2D ? dst->ne[2] : 1;
    const uint32_t OW            = dst->ne[1];
    const uint32_t total_patches = N * OH * OW;
    const uint32_t total_rows    = N * OH;
@@ -280,8 +511,11 @@ int op_im2col(struct htp_ops_context * octx) {
    uint32_t npatches   = total_patches;
    if (octx->ctx->mdev.count > 1) {
        const uint32_t patch_size = dst->nb[1];
-        const uint32_t patches_per_chunk = (patch_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(patch_size, HEX_L2_LINE_SIZE)) : 1;
-        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_patches, htp_tensor_mdev_data_aligned(dst) ? patches_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        const uint32_t patches_per_chunk =
+            (patch_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(patch_size, HEX_L2_LINE_SIZE)) : 1;
+        const struct htp_tensor_mdev_range range =
+            htp_tensor_mdev_partition(total_patches, htp_tensor_mdev_data_aligned(dst) ? patches_per_chunk : 0,
+                                      octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
        patch_base = range.start;
        npatches   = range.count;
    }
@@ -290,8 +524,11 @@ int op_im2col(struct htp_ops_context * octx) {
    uint32_t nrows    = total_rows;
    if (octx->ctx->mdev.count > 1) {
        const uint32_t row_size = dst->nb[2];
-        const uint32_t rows_per_chunk = (row_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(row_size, HEX_L2_LINE_SIZE)) : 1;
-        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, htp_tensor_mdev_data_aligned(dst) ? rows_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        const uint32_t rows_per_chunk =
+            (row_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(row_size, HEX_L2_LINE_SIZE)) : 1;
+        const struct htp_tensor_mdev_range range =
+            htp_tensor_mdev_partition(total_rows, htp_tensor_mdev_data_aligned(dst) ? rows_per_chunk : 0,
+                                      octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
        row_base = range.start;
        nrows    = range.count;
    }
@@ -309,22 +546,32 @@ int op_im2col(struct htp_ops_context * octx) {
    ictx.npatches_per_thread = (npatches + n_threads - 1) / n_threads;

    // Clean non-overlapping patch-embed -> DMA kernel (if it fits VTCM);
-    // everything else (padding/dilation/stride edges) -> pure-DDR kernel.
-    if (im2col_use_patchembed_dma(octx) && nrows > 0) {
+    // everything else (padding/dilation/stride edges) -> blocked-staging DMA
+    // kernel; if neither fits VTCM -> pure-DDR kernel.
+    if (nrows > 0) {
        const uint32_t pth = MIN(octx->n_threads, nrows);
-        if (pth > 0 && im2col_patchembed_dma_fits(octx, &ictx, pth)) {
-            ictx.pe_row_base        = row_base;
-            ictx.pe_nrows           = nrows;
-            ictx.pe_rows_per_thread = (nrows + pth - 1) / pth;
-            if (dst->type == HTP_TYPE_F16) {
-                work_queue_run(octx->ctx->work_queue, im2col_patchembed_dma_thread, &ictx, pth);
-            } else {
-                work_queue_run(octx->ctx->work_queue, im2col_patchembed_dma_f32_thread, &ictx, pth);
+        if (pth > 0) {
+            ictx.pe_row_base = row_base;
+            ictx.pe_nrows    = nrows;
+            const bool exact = im2col_use_patchembed_dma(octx);
+            if (exact && im2col_patchembed_dma_fits(octx, &ictx, pth)) {
+                ictx.pe_rows_per_thread = (nrows + pth - 1) / pth;
+                work_queue_run(octx->ctx->work_queue,
+                    dst->type == HTP_TYPE_F16 ? im2col_patchembed_dma_thread
+                                              : im2col_patchembed_dma_f32_thread, &ictx, pth);
+                return HTP_STATUS_OK;
+            }
+            if (!exact && im2col_blocked_dma_fits(octx, &ictx, pth)) {
+                ictx.pe_rows_per_thread = (nrows + pth - 1) / pth;
+                work_queue_run(octx->ctx->work_queue,
+                    dst->type == HTP_TYPE_F16 ? im2col_blocked_dma_thread
+                                              : im2col_blocked_dma_f32_thread, &ictx, pth);
+                return HTP_STATUS_OK;
            }
-            return HTP_STATUS_OK;
        }
-        // else: doesn't fit -> fall through to the pure-DDR kernel below.
    }
+    // Fall through to pure-DDR.
+

    if (npatches == 0) {
        return HTP_STATUS_OK;
@@ -879,6 +879,9 @@ static int execute_op(struct htp_ops_context * octx) {
        case HTP_OP_IM2COL:
            return op_im2col(octx);

+        case HTP_OP_ROLL:
+            return op_roll(octx);
+
        case HTP_OP_CONCAT:
            return op_concat(octx);

@@ -0,0 +1,316 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <string.h>
+
+#include "dma-queue.h"
+#include "hvx-utils.h"
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "hex-common.h"
+#include "hex-profile.h"
+#include "htp-ops.h"
+#include "htp-tensor.h"
+
+struct htp_roll_context {
+    struct htp_ops_context * octx;
+
+    uint32_t row_start;
+    uint32_t nrows;
+    uint32_t nrows_per_thread;
+
+    struct fastdiv_values div_ne1;
+    struct fastdiv_values div_ne2_ne1;
+};
+
+static inline uint32_t htp_roll_wrap(int32_t i, uint32_t ne) {
+    if (i < 0) {
+        return (uint32_t) (i + (int32_t) ne);
+    }
+    if ((uint32_t) i >= ne) {
+        return (uint32_t) i - ne;
+    }
+    return (uint32_t) i;
+}
+
+#define htp_roll_preamble                             \
+    const struct htp_tensor * src0 = octx->src[0];    \
+    const struct htp_tensor * dst  = octx->dst;       \
+                                                      \
+    const uint32_t ne0 = dst->ne[0];                  \
+    const uint32_t ne1 = dst->ne[1];                  \
+    const uint32_t ne2 = dst->ne[2];                  \
+    const uint32_t ne3 = dst->ne[3];                  \
+                                                      \
+    const uint32_t nb01 = src0->nb[1];                \
+    const uint32_t nb02 = src0->nb[2];                \
+    const uint32_t nb03 = src0->nb[3];                \
+                                                      \
+    const uint32_t nb1 = dst->nb[1];                  \
+    const uint32_t nb2 = dst->nb[2];                  \
+    const uint32_t nb3 = dst->nb[3];                  \
+                                                      \
+    const int32_t s0 = octx->op_params[0];            \
+    const int32_t s1 = octx->op_params[1];            \
+    const int32_t s2 = octx->op_params[2];            \
+    const int32_t s3 = octx->op_params[3];            \
+                                                      \
+    const uint32_t i0_src0 = htp_roll_wrap(-s0, ne0); \
+    const uint32_t n0      = ne0 - i0_src0;
+
+#define htp_roll_dma_preamble dma_queue * q = octx->ctx->dma[0];
+
+static inline void roll_dma_push(dma_queue * q,
+                                 uintptr_t   dst,
+                                 uintptr_t   src,
+                                 uint32_t    dst_stride,
+                                 uint32_t    src_stride,
+                                 uint32_t    bytes,
+                                 uint32_t    nrows) {
+    if (bytes == 0 || nrows == 0) {
+        return;
+    }
+
+    if (!dma_queue_push(q, dma_make_ptr((void *) dst, (const void *) src), dst_stride, src_stride, bytes, nrows)) {
+        dma_queue_flush(q);
+        dma_queue_push(q, dma_make_ptr((void *) dst, (const void *) src),
+                                       dst_stride, src_stride, bytes, nrows);
+    }
+}
+
+static inline void roll_dma_push_rows(dma_queue *               q,
+                                      const struct htp_tensor * dst,
+                                      const struct htp_tensor * src0,
+                                      uint32_t                  dst_row,
+                                      uint32_t                  src_row,
+                                      uint32_t                  nrows,
+                                      uint32_t                  row_size,
+                                      uint32_t                  i0_src0) {
+    const uintptr_t dst_base = dst->data + (uintptr_t) dst_row * row_size;
+    const uintptr_t src_base = src0->data + (uintptr_t) src_row * row_size;
+    const uint32_t  n0       = src0->ne[0] - i0_src0;
+
+    roll_dma_push(q, dst_base, src_base + (uintptr_t) i0_src0 * sizeof(float),
+                  row_size, row_size, n0 * sizeof(float), nrows);
+    roll_dma_push(q, dst_base + (uintptr_t) n0 * sizeof(float), src_base,
+                  row_size, row_size, i0_src0 * sizeof(float), nrows);
+}
+
+// Same row-wrap split as roll_dma_push_rows, but addressed with explicit byte strides so it
+// also works for a src0 that is row-contiguous only (e.g. a permuted view) rather than fully packed.
+static inline void roll_dma_push_range(dma_queue * q,
+                                       uintptr_t   dst_row,
+                                       uintptr_t   src_row,
+                                       uint32_t    dst_stride,
+                                       uint32_t    src_stride,
+                                       uint32_t    nrows,
+                                       uint32_t    i0_src0,
+                                       uint32_t    n0) {
+    roll_dma_push(q, dst_row, src_row + (uintptr_t) i0_src0 * sizeof(float),
+                  dst_stride, src_stride, n0 * sizeof(float), nrows);
+    roll_dma_push(q, dst_row + (uintptr_t) n0 * sizeof(float), src_row,
+                  dst_stride, src_stride, i0_src0 * sizeof(float), nrows);
+}
+
+static int roll_dma_f32_contiguous(struct htp_ops_context * octx) {
+    htp_roll_preamble;
+    htp_roll_dma_preamble;
+
+    const uint32_t row_size = ne0 * sizeof(float);
+
+    if (s1 == 0 && s2 == 0 && s3 == 0) {
+        roll_dma_push_rows(q, dst, src0, 0, 0, ne1 * ne2 * ne3, row_size, i0_src0);
+        dma_queue_flush(q);
+        return HTP_STATUS_OK;
+    }
+
+    if (s1 == 0) {
+        const uint32_t i2_src0 = htp_roll_wrap(-s2, ne2);
+        for (uint32_t i3 = 0; i3 < ne3; i3++) {
+            const uint32_t i03 = htp_roll_wrap((int32_t) i3 - s3, ne3);
+            const uint32_t dst_row0 = i3 * ne2 * ne1;
+            const uint32_t src_row0 = (i03 * ne2 + i2_src0) * ne1;
+            const uint32_t n2_first = ne2 - i2_src0;
+
+            roll_dma_push_rows(q, dst, src0, dst_row0, src_row0, n2_first * ne1,
+                               row_size, i0_src0);
+            roll_dma_push_rows(q, dst, src0, dst_row0 + n2_first * ne1, i03 * ne2 * ne1,
+                               i2_src0 * ne1, row_size, i0_src0);
+        }
+
+        dma_queue_flush(q);
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t i1_src0 = htp_roll_wrap(-s1, ne1);
+    const uint32_t n1_first = ne1 - i1_src0;
+
+    for (uint32_t i3 = 0; i3 < ne3; i3++) {
+        const uint32_t i03 = htp_roll_wrap((int32_t) i3 - s3, ne3);
+        for (uint32_t i2 = 0; i2 < ne2; i2++) {
+            const uint32_t i02 = htp_roll_wrap((int32_t) i2 - s2, ne2);
+            const uint32_t dst_row0 = (i3 * ne2 + i2) * ne1;
+            const uint32_t src_row0 = (i03 * ne2 + i02) * ne1;
+
+            roll_dma_push_rows(q, dst, src0, dst_row0, src_row0 + i1_src0,
+                               n1_first, row_size, i0_src0);
+            roll_dma_push_rows(q, dst, src0, dst_row0 + n1_first, src_row0,
+                               i1_src0, row_size, i0_src0);
+        }
+    }
+
+    dma_queue_flush(q);
+    return HTP_STATUS_OK;
+}
+
+// DMA path for a row-contiguous but otherwise arbitrarily strided src0 (e.g. a permuted view).
+// Same row-wrap split as above, one DMA push per (i2,i3), addressed via the real nb01/nb02/nb03
+// instead of assuming a packed layout.
+static int roll_dma_f32_strided(struct htp_ops_context * octx) {
+    htp_roll_preamble;
+    htp_roll_dma_preamble;
+
+    const uint32_t i1_src0  = htp_roll_wrap(-s1, ne1);
+    const uint32_t n1_first = ne1 - i1_src0;
+
+    for (uint32_t i3 = 0; i3 < ne3; i3++) {
+        const uint32_t i03 = htp_roll_wrap((int32_t) i3 - s3, ne3);
+        for (uint32_t i2 = 0; i2 < ne2; i2++) {
+            const uint32_t i02 = htp_roll_wrap((int32_t) i2 - s2, ne2);
+
+            const uintptr_t dst_row0 = dst->data  + (uintptr_t) i2  * nb2  + (uintptr_t) i3  * nb3;
+            const uintptr_t src_row0 = src0->data + (uintptr_t) i02 * nb02 + (uintptr_t) i03 * nb03;
+
+            roll_dma_push_range(q, dst_row0, src_row0 + (uintptr_t) i1_src0 * nb01,
+                                nb1, nb01, n1_first, i0_src0, n0);
+            roll_dma_push_range(q, dst_row0 + (uintptr_t) n1_first * nb1, src_row0,
+                                nb1, nb01, i1_src0, i0_src0, n0);
+        }
+    }
+
+    dma_queue_flush(q);
+    return HTP_STATUS_OK;
+}
+
+static void roll_thread_f32(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_roll_context * rctx = (struct htp_roll_context *) data;
+    struct htp_ops_context * octx = rctx->octx;
+
+    htp_roll_preamble;
+
+    const uint32_t row_start = rctx->row_start + rctx->nrows_per_thread * ith;
+    const uint32_t row_end   = MIN(row_start + rctx->nrows_per_thread, rctx->row_start + rctx->nrows);
+    if (row_start >= row_end) {
+        return;
+    }
+
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
+
+    for (uint32_t row = row_start; row < row_end; row++) {
+        const uint32_t i3  = fastdiv(row, &rctx->div_ne2_ne1);
+        const uint32_t rem = row - i3 * ne2 * ne1;
+        const uint32_t i2  = fastdiv(rem, &rctx->div_ne1);
+        const uint32_t i1  = rem - i2 * ne1;
+
+        const uint32_t i01 = htp_roll_wrap((int32_t) i1 - s1, ne1);
+        const uint32_t i02 = htp_roll_wrap((int32_t) i2 - s2, ne2);
+        const uint32_t i03 = htp_roll_wrap((int32_t) i3 - s3, ne3);
+
+        const uint8_t * src_row = (const uint8_t *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+        uint8_t * dst_row = (uint8_t *) dst->data + i1*nb1 + i2*nb2 + i3*nb3;
+
+        hex_l2fetch(src_row + i0_src0 * sizeof(float), n0 * sizeof(float), ne0 * sizeof(float), 1);
+        hvx_copy_uu(dst_row, src_row + i0_src0 * sizeof(float), n0, sizeof(float));
+
+        if (i0_src0 != 0) {
+            hex_l2fetch(src_row, i0_src0 * sizeof(float), ne0 * sizeof(float), 1);
+            hvx_copy_uu(dst_row + n0 * sizeof(float), src_row, i0_src0, sizeof(float));
+        }
+    }
+
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
+
+    FARF(HIGH, "roll %d/%d: (%ux%ux%ux%u) rows %u:%u shift=(%d,%d,%d,%d)\n",
+         ith, nth, ne0, ne1, ne2, ne3,
+         row_start, row_end, s0, s1, s2, s3);
+}
+
+int execute_op_roll_f32(struct htp_ops_context * octx) {
+    htp_roll_preamble;
+
+    if (src0->type != HTP_TYPE_F32 || dst->type != HTP_TYPE_F32) {
+        FARF(ERROR, "roll: unsupported type %u -> %u\n", src0->type, dst->type);
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (src0->nb[0] != sizeof(float) || dst->nb[0] != sizeof(float)) {
+        FARF(ERROR, "roll: unsupported nb0 %u -> %u\n", src0->nb[0], dst->nb[0]);
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (src0->ne[0] != ne0 || src0->ne[1] != ne1 ||
+        src0->ne[2] != ne2 || src0->ne[3] != ne3) {
+        FARF(ERROR, "roll: shape mismatch\n");
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t total_rows = ne1 * ne2 * ne3;
+    const size_t dst_row_size = ne0 * sizeof(float);
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = total_rows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, sizeof(float), (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    if (octx->ctx->mdev.count <= 1) {
+        if (htp_tensor_is_contiguous(src0, sizeof(float)) && htp_tensor_is_contiguous(dst, sizeof(float))) {
+            return roll_dma_f32_contiguous(octx);
+        }
+        return roll_dma_f32_strided(octx);
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+    struct htp_roll_context rctx = {
+        .octx             = octx,
+        .row_start        = row_start,
+        .nrows            = nrows,
+        .nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div),
+        .div_ne1          = init_fastdiv_values(dst->ne[1]),
+        .div_ne2_ne1      = init_fastdiv_values(dst->ne[2] * dst->ne[1]),
+    };
+
+    work_queue_run(octx->ctx->work_queue, roll_thread_f32, &rctx, n_threads);
+
+    return HTP_STATUS_OK;
+}
+
+int op_roll(struct htp_ops_context * octx) {
+    switch (octx->src[0]->type) {
+        case HTP_TYPE_F32:
+            return execute_op_roll_f32(octx);
+
+        default:
+            return HTP_STATUS_NO_SUPPORT;
+    }
+}
@@ -191,6 +191,7 @@ set(GGML_OPENCL_KERNELS
    gemv_noshuffle_q6_k_f32_tiled
    gemm_noshuffle_q6_k_f32
    gemm_noshuffle_q6_k_f32_tiled
+    gemv_noshuffle_q6_k_f32_32b_trans
    gemv_noshuffle_q5_k_f32
    gemm_noshuffle_q5_k_f32
    mul
@@ -232,6 +233,7 @@ set(GGML_OPENCL_KERNELS
    mul_mm_f16_f32_kq_kqv
    conv2d
    conv2d_f16_f32
+    flash_attn_repack
    flash_attn_pre_f16
    flash_attn_f32_f16
    flash_attn_f32_q8_0
@@ -567,6 +567,16 @@ struct ggml_opencl_fa_kernels {
    // attempted (variant, (dk, dv))
    // all attempted FA kernels appear here, but those not registered failed compilation
    std::set<std::pair<int, std::pair<int, int>>> variant_attempted;
+
+    // FA bin kernels
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    cl_kernel kernel_flash_attn_f32_f16_bin;
+
+    cl_kernel kernel_repack_q_for_wmm;
+    cl_kernel kernel_repack_k_for_wmm;
+    cl_kernel kernel_repack_v_for_wmm;
+    cl_kernel kernel_repack_mask_for_wmm;
+#endif
 };

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
@@ -1246,6 +1256,8 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_gemv_noshuffle_q6_K_f32_mc3;        // multi-column (N=3) verify GEMV
    cl_kernel kernel_gemm_noshuffle_q6_K_f32;
    cl_kernel kernel_gemm_noshuffle_q6_K_f32_cok;
+    cl_kernel kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin;
+    cl_kernel kernel_gemv_noshuffle_q6_k_f32_32b_trans;
    cl_kernel kernel_gemv_noshuffle_q5_k_f32;
    cl_kernel kernel_gemv_noshuffle_q5_k_f32_mc3;  // multi-column (N=3) verify GEMV (spec/MTP)
    cl_kernel kernel_gemm_noshuffle_q5_k_f32;
@@ -1440,6 +1452,7 @@ static void load_cl_kernels_argsort(ggml_backend_opencl_context *backend_ctx) {

 static bool use_adreno_bin_kernels(ggml_backend_opencl_context * backend_ctx) {
 #ifndef GGML_OPENCL_USE_ADRENO_BIN_KERNELS
+    GGML_UNUSED(backend_ctx);
    return false;
 #else
    if (backend_ctx->gpu_family != GPU_FAMILY::ADRENO) {
@@ -4366,6 +4379,43 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        }
    }

+    backend_ctx->kernel_gemv_noshuffle_q6_k_f32_32b_trans = nullptr;
+    backend_ctx->kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin = nullptr;
+    if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E) {
+        {
+            std::string opts = std::string("-cl-std=") + opencl_c_std +
+                                           " -cl-mad-enable "
+                                           " -DSIMDGROUP_WIDTH=" +
+                                           std::to_string(backend_ctx->adreno_wave_size);
+#ifdef GGML_OPENCL_EMBED_KERNELS
+            const std::string kernel_src {
+                #include "gemv_noshuffle_q6_k_f32_32b_trans.cl.h"
+            };
+#else
+            const std::string kernel_src = read_file("gemv_noshuffle_q6_k_f32_32b_trans.cl");
+#endif
+            cl_program prog = build_program_from_source(backend_ctx, kernel_src.c_str(), opts);
+            CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q6_k_f32_32b_trans =
+                clCreateKernel(prog, "kernel_gemv_noshuffle_q6_k_f32_32b_trans", &err), err));
+            CL_CHECK(clReleaseProgram(prog));
+            GGML_LOG_CONT(".");
+        }
+
+        if (use_adreno_bin_kernels(backend_ctx)) {
+            size_t bin_size = 0;
+            const char * kernel_bin = (const char *)backend_ctx->get_adreno_bin_kernel("gemm_noshuffle_q6_k_f32_32b_trans_ila_a8", &bin_size);
+            if (kernel_bin && bin_size > 0) {
+                cl_program bin_prog =
+                    build_program_from_binary(backend_ctx->context, backend_ctx->device, kernel_bin, "", bin_size);
+
+                CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin =
+                    clCreateKernel(bin_prog, "kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8", &err), err));
+                CL_CHECK(clReleaseProgram(bin_prog));
+                GGML_LOG_CONT(".");
+            }
+        }
+    }
+
    std::string CL_moe_compile_opts = std::string("-cl-std=") + opencl_c_std +
            " -cl-mad-enable "
            " -cl-fast-relaxed-math";
@@ -5132,6 +5182,43 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        CL_CHECK(clReleaseProgram(prog));
        GGML_LOG_CONT(".");
    }
+
+    // repack
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "flash_attn_repack.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("flash_attn_repack.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->fa.kernel_repack_q_for_wmm = clCreateKernel(prog, "kernel_repack_q_for_wmm", &err), err));
+        CL_CHECK((backend_ctx->fa.kernel_repack_k_for_wmm = clCreateKernel(prog, "kernel_repack_k_for_wmm", &err), err));
+        CL_CHECK((backend_ctx->fa.kernel_repack_v_for_wmm = clCreateKernel(prog, "kernel_repack_v_for_wmm", &err), err));
+        CL_CHECK((backend_ctx->fa.kernel_repack_mask_for_wmm = clCreateKernel(prog, "kernel_repack_mask_for_wmm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // kernel_flash_attn_f32_f16_bin
+    {
+        size_t bin_size = 0;
+        backend_ctx->fa.kernel_flash_attn_f32_f16_bin = nullptr;
+
+        if (use_adreno_bin_kernels(backend_ctx)) {
+            const char * kernel_bin = (const char *)backend_ctx->get_adreno_bin_kernel("flash_attn_f32_f16_wmm", &bin_size);
+            if (kernel_bin && bin_size > 0) {
+                cl_program prog =
+                    build_program_from_binary(backend_ctx->context, backend_ctx->device, kernel_bin, CL_moe_compile_opts, bin_size);
+
+                CL_CHECK((backend_ctx->fa.kernel_flash_attn_f32_f16_bin = clCreateKernel(prog, "flash_attn_f32_f16", &err), err));
+                CL_CHECK(clReleaseProgram(prog));
+                GGML_LOG_CONT(".");
+            }
+        }
+    }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
    GGML_LOG_CONT("\n");
    backend_ctx->kernels_loaded = true;
@@ -6309,6 +6396,8 @@ static void ggml_opencl_print_backend_info(ggml_backend_opencl_device_context *

    auto * backend_ctx = dev_ctx->backend_ctx;

+    GGML_LOG_INFO("ggml_opencl: OpenCL device: %s\n",
+        backend_ctx->device_name.c_str());
    GGML_LOG_INFO("ggml_opencl: OpenCL driver: %s\n",
        backend_ctx->driver_version.c_str());
    GGML_LOG_INFO("ggml_opencl: vector subgroup broadcast support: %s\n",
@@ -6325,11 +6414,11 @@ static void ggml_opencl_print_backend_info(ggml_backend_opencl_device_context *
        backend_ctx->global_mem_size/1024/1024);
    GGML_LOG_INFO("ggml_opencl: max mem alloc size: %zu MB\n",
        backend_ctx->max_alloc_size/1024/1024);
-    GGML_LOG_INFO("ggml_opencl: device max image buffer size (pixels): %lu\n",
+    GGML_LOG_INFO("ggml_opencl: device max image buffer size (pixels): %zu\n",
        backend_ctx->image_max_buffer_size);
-    GGML_LOG_INFO("ggml_opencl: device max image2d size: %lu x %lu\n",
+    GGML_LOG_INFO("ggml_opencl: device max image2d size: %zu x %zu\n",
        backend_ctx->image2d_max_width, backend_ctx->image2d_max_height);
-    GGML_LOG_INFO("ggml_opencl: device max workgroup size: %lu\n",
+    GGML_LOG_INFO("ggml_opencl: device max workgroup size: %zu\n",
        backend_ctx->max_workgroup_size);
    GGML_LOG_INFO("ggml_opencl: SVM coarse grain buffer support: %s\n",
        backend_ctx->svm_caps & CL_DEVICE_SVM_COARSE_GRAIN_BUFFER ? "true" : "false");
@@ -7291,6 +7380,8 @@ struct ggml_tensor_extra_cl_q6_K {
    cl_mem ql_img = nullptr;
    // Upper 2 bits of quantized weights.
    cl_mem qh = nullptr;
+    // Upper 2 bits as image1d_buffer_t
+    cl_mem qh_img = nullptr;
    // Scales for each block.
    cl_mem s  = nullptr;
    // Scales for each super block.
@@ -7326,6 +7417,10 @@ struct ggml_tensor_extra_cl_q6_K {
            CL_CHECK(clReleaseMemObject(ql_img));
            ql_img = nullptr;
        }
+        if (qh_img != nullptr) {
+            CL_CHECK(clReleaseMemObject(qh_img));
+            qh_img = nullptr;
+        }

        size_ql = 0;
        size_qh = 0;
@@ -7627,7 +7722,7 @@ static void ggml_cl_moe_bias_glu_fused(ggml_backend_t backend, ggml_tensor * gat
    size_t global_work_size[] = { (size_t)glu->ne[1]*nth, (size_t)glu->ne[2], 1 };
    size_t local_work_size[]  = { (size_t)nth, 1, 1 };

-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, (ggml_tensor *)glu);
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, glu);
 }

 // Fusion B: the MoE down-projection bias add feeding the combine.
@@ -7771,7 +7866,7 @@ static void ggml_cl_moe_bias_combine_fused(ggml_backend_t backend, const ggml_te

    size_t lws[2] = { 64, 1 };
    size_t gws[2] = { (size_t)(((n_embd4 + 63) / 64) * 64), (size_t)nt };
-    backend_ctx->enqueue_ndrange_kernel(kernel, 2, gws, lws, (ggml_tensor *)dst);
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, gws, lws, dst);
 }


@@ -8114,7 +8209,6 @@ static void ggml_cl_mul_mat_q4_k_glu_fused(ggml_backend_t backend, ggml_tensor *
    GGML_UNUSED(gate_tensor);
    GGML_UNUSED(up_tensor);
    GGML_UNUSED(glu_tensor);
-    GGML_ABORT("q4_K GLU fusion requires GGML_OPENCL_USE_ADRENO_KERNELS");
 #endif
 }

@@ -8485,6 +8579,28 @@ inline bool use_q4_0_bin_kernels(const ggml_backend_opencl_context *backend_ctx,
 #endif
 }

+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+static bool use_fa_bin_kernels_prefill(const ggml_backend_opencl_context * backend_ctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v) {
+    if (backend_ctx->fa.kernel_flash_attn_f32_f16_bin == nullptr) {
+        return false;
+    }
+
+    const bool is_mixed = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 && v->type == GGML_TYPE_F16;
+    const bool is_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 && v->type == GGML_TYPE_Q8_0;
+
+    const int n_q = q->ne[1];
+    const int dk = q->ne[0];
+    const int dv = v->ne[0];
+
+    constexpr bool prefill_only = true;
+
+    return (backend_ctx->gpu_family == GPU_FAMILY::ADRENO &&
+            (is_mixed || is_q8_0) && (dk == dv)
+            && (dk == 64 || dk == 128 || dk == 256 || dk == 512)
+            && (!prefill_only || n_q != 1));
+}
+#endif
+
 // The flat-GEMV large-m escape is OPT-IN (GGML_OPENCL_FLAT_LARGE_M=1) because it
 // is SLOWER than the route it replaces, not because it is unsafe. It was first
 // parked on the theory that it out-of-bounds-writes at vocab-scale shapes; that
@@ -8564,6 +8680,21 @@ static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_backend_opencl_cont
        && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }

+inline bool use_q6_k_bin_kernels(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    if (!backend_ctx->kernel_gemv_noshuffle_q6_k_f32_32b_trans ||
+        !backend_ctx->kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin) {
+        return false;
+    }
+    return (tensor->ne[0] % 256 == 0) && (tensor->ne[1] % 64 == 0) &&
+           !use_q6k_tiled(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor);
+#else
+    GGML_UNUSED(backend_ctx);
+    GGML_UNUSED(tensor);
+    return false;
+#endif
+}
+
 inline bool use_q4_k_bin_kernels(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    if (!backend_ctx->kernel_gemv_noshuffle_q4_k_f32_32b_trans ||
@@ -8928,6 +9059,11 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
        case GGML_OP_MEAN:
            return op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_FLASH_ATTN_EXT: {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+            if (use_fa_bin_kernels_prefill(backend_ctx, op->src[0], op->src[1], op->src[2])) {
+                return true;
+            }
+#endif
            // The E17 compilers segfault while building FA kernels, skip E17 for now
            if (adreno_e17_compiler_quirks(backend_ctx)) {
                return false;
@@ -11179,18 +11315,39 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
            cl_int M = tensor->ne[1];   // ne01
            cl_int K = tensor->ne[0];   // ne00

-            // Transpose ql as ushort
-            transpose_2d_as_16b(backend_ctx,
-                extra->ql, extra->ql, size_ql, K/4, M);
+            if (use_q6_k_bin_kernels(backend_ctx, tensor)) {
+                GGML_ASSERT(K % 256 == 0);
+                GGML_ASSERT(M % 64 == 0);

-            // Transpose qh as uchar
-            transpose_2d_as_8b(backend_ctx,
-                extra->qh, extra->qh, size_qh, K/4, M);
+                transpose_2d_as_32b(backend_ctx, extra->ql, extra->ql, size_ql, K/8,  M);
+                transpose_2d_as_32b(backend_ctx, extra->qh, extra->qh, size_qh, K/16, M);

-            // Transpose s as ushort
-            transpose_2d_as_16b(backend_ctx,
-                extra->s, extra->s, size_s, K/16/2, M);
+                cl_image_format wimg_fmt = { CL_R, CL_UNSIGNED_INT32 };
+                cl_image_desc   wimg_desc;
+                memset(&wimg_desc, 0, sizeof(wimg_desc));
+                wimg_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+                wimg_desc.image_width = static_cast<size_t>(ggml_nelements(tensor) / 8);
+                wimg_desc.buffer      = extra->ql;
+                CL_CHECK((extra->ql_img = clCreateImage(context, CL_MEM_READ_ONLY, &wimg_fmt, &wimg_desc, NULL, &err), err));

+                memset(&wimg_desc, 0, sizeof(wimg_desc));
+                wimg_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+                wimg_desc.image_width = static_cast<size_t>(ggml_nelements(tensor) / 16);
+                wimg_desc.buffer      = extra->qh;
+                CL_CHECK((extra->qh_img = clCreateImage(context, CL_MEM_READ_ONLY, &wimg_fmt, &wimg_desc, NULL, &err), err));
+            } else {
+                // Transpose ql as ushort
+                transpose_2d_as_16b(backend_ctx,
+                    extra->ql, extra->ql, size_ql, K/4, M);
+
+                // Transpose qh as uchar
+                transpose_2d_as_8b(backend_ctx,
+                    extra->qh, extra->qh, size_qh, K/4, M);
+
+                // Transpose s as ushort
+                transpose_2d_as_16b(backend_ctx,
+                    extra->s, extra->s, size_s, K/16/2, M);
+            }
            // Transpose d as ushort
            transpose_2d_as_16b(backend_ctx,
                extra->d, extra->d, size_d, K/256, M);
@@ -11336,7 +11493,6 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
        if (use_adreno_moe_kernels(backend_ctx, tensor)) {
-            cl_int err;
            cl_kernel kernel = backend_ctx->kernel_restore_block_q4_0_trans4_ns;

            cl_mem data_device = ggml_cl_create_temp_download_buffer(context, queue, ggml_nbytes(tensor), tensor->name);
@@ -11535,7 +11691,6 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
        if (use_adreno_moe_kernels(backend_ctx, tensor)) {
-            cl_int err;
            // TODO: use ggml_cl_buffer to manage this temporary buffer
            cl_mem data_device = ggml_cl_create_temp_download_buffer(context, queue, ggml_nbytes(tensor), tensor->name);
            GGML_ASSERT(data_device != NULL && "get_tensor: temp download buffer alloc failed");
@@ -11638,7 +11793,6 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
        if (use_adreno_moe_kernels(backend_ctx, tensor)) {
-            cl_int err;
            // TODO: use ggml_cl_buffer to manage this temporary buffer
            cl_mem data_device = ggml_cl_create_temp_download_buffer(context, queue, ggml_nbytes(tensor), tensor->name);
            GGML_ASSERT(data_device != NULL && "get_tensor: temp download buffer alloc failed");
@@ -12318,15 +12472,24 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,

            buf_trans_ql.allocate(backend_ctx->context, size_ql);
            buf_trans_qh.allocate(backend_ctx->context, size_qh);
-            buf_trans_s.allocate(backend_ctx->context, size_s);
            buf_trans_d.allocate(backend_ctx->context, size_d);
            buf_unpacked.allocate(backend_ctx->context, ggml_nbytes(tensor));

-            // transpose ql, qh, s and d back
-            transpose_2d_as_16b(backend_ctx, extra->ql, buf_trans_ql.buffer, size_ql, M, K/4);
-            transpose_2d_as_8b(backend_ctx,  extra->qh, buf_trans_qh.buffer, size_qh, M, K/4);
-            transpose_2d_as_16b(backend_ctx, extra->s,  buf_trans_s.buffer,  size_s,  M, K/16/2);
-            transpose_2d_as_16b(backend_ctx, extra->d,  buf_trans_d.buffer,  size_d,  M, K/256);
+            cl_mem s_buffer;
+            if (use_q6_k_bin_kernels(backend_ctx, tensor)) {
+                transpose_2d_as_32b(backend_ctx, extra->ql, buf_trans_ql.buffer, size_ql, M, K/8);
+                transpose_2d_as_32b(backend_ctx, extra->qh, buf_trans_qh.buffer, size_qh, M, K/16);
+                // s is left row-major, untransposed, for the binary layout.
+                s_buffer = extra->s;
+            } else {
+                // transpose ql, qh, s and d back
+                buf_trans_s.allocate(backend_ctx->context, size_s);
+                transpose_2d_as_16b(backend_ctx, extra->ql, buf_trans_ql.buffer, size_ql, M, K/4);
+                transpose_2d_as_8b(backend_ctx,  extra->qh, buf_trans_qh.buffer, size_qh, M, K/4);
+                transpose_2d_as_16b(backend_ctx, extra->s,  buf_trans_s.buffer,  size_s,  M, K/16/2);
+                s_buffer = buf_trans_s.buffer;
+            }
+            transpose_2d_as_16b(backend_ctx, extra->d, buf_trans_d.buffer, size_d, M, K/256);

            // unpack
            cl_uchar mask = 0xFF;
@@ -12334,7 +12497,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
            cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K_noshuffle;
            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_ql.buffer));
            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_qh.buffer));
-            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &buf_trans_s.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &s_buffer));
            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &buf_trans_d.buffer));
            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &buf_unpacked.buffer));
            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask));
@@ -17109,6 +17272,407 @@ static void ggml_cl_adreno_xmem_attn_run(

 #endif // GGML_OPENCL_USE_ADRENO_KERNELS

+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+static void ggml_cl_flash_attn_prefill_bin(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
+    const ggml_tensor * v = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+    GGML_ASSERT(q->extra);
+    GGML_ASSERT(k->extra);
+    GGML_ASSERT(v->extra);
+    GGML_ASSERT(dst->extra);
+    if (mask) {
+        GGML_ASSERT(mask->extra);
+    }
+    if (sinks) {
+        GGML_ASSERT(sinks->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_context context = backend_ctx->context;
+
+    const int n_q = q->ne[1];
+    const int n_kv = k->ne[1];
+    const int d_head_q = q->ne[0];
+    const int d_head_v = v->ne[0];
+    const int n_head = q->ne[2];
+    const int n_head_kv = k->ne[2];
+    const int n_batch = q->ne[3];
+
+    const std::pair<int, int> dk_dv = {d_head_q, d_head_v};
+    cl_kernel kernel = backend_ctx->fa.kernel_flash_attn_f32_f16_bin;
+    GGML_ASSERT(kernel != NULL);
+
+    ggml_tensor_extra_cl * extra_q = (ggml_tensor_extra_cl *)q->extra;
+    ggml_tensor_extra_cl * extra_k = (ggml_tensor_extra_cl *)k->extra;
+    ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
+    ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+    ggml_tensor_extra_cl * extra_sinks = sinks ? (ggml_tensor_extra_cl *)sinks->extra : NULL;
+
+    cl_ulong offset_q = extra_q->offset + q->view_offs;
+    cl_ulong offset_o = extra_o->offset + dst->view_offs;
+
+    cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
+    cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+    cl_mem   sinks_buffer = extra_sinks ? extra_sinks->data_device : NULL;
+    cl_ulong offset_sinks = extra_sinks ? extra_sinks->offset + sinks->view_offs : 0;
+
+    const cl_ulong q_nb1 = q->nb[1];
+    const cl_ulong q_nb2 = q->nb[2];
+    const cl_ulong q_nb3 = q->nb[3];
+
+    cl_mem   k_data_device = extra_k->data_device;
+    cl_ulong offset_k = extra_k->offset + k->view_offs;
+    cl_ulong k_nb1 = k->nb[1];
+    cl_ulong k_nb2 = k->nb[2];
+    cl_ulong k_nb3 = k->nb[3];
+
+    cl_mem   v_data_device = extra_v->data_device;
+    cl_ulong offset_v = extra_v->offset + v->view_offs;
+    cl_ulong v_nb1 = v->nb[1];
+    cl_ulong v_nb2 = v->nb[2];
+    cl_ulong v_nb3 = v->nb[3];
+
+    const cl_ulong o_nb1 = dst->nb[1];
+    const cl_ulong o_nb2 = dst->nb[2];
+    const cl_ulong o_nb3 = dst->nb[3];
+
+    const cl_ulong mask_nb1 = mask ? mask->nb[1] : 0;
+    const cl_ulong mask_nb2 = mask ? mask->nb[2] : 0;
+    const cl_ulong mask_nb3 = mask ? mask->nb[3] : 0;
+    const int mask_ne2 = mask ? mask->ne[2] : 0;
+    const int mask_ne3 = mask ? mask->ne[3] : 0;
+
+    float * params      = (float *)dst->op_params;
+    float scale         = params[0];
+    float max_bias      = params[1];
+    float logit_softcap = params[2];
+
+    const int is_causal = (mask == NULL && n_q > 1 && n_q == n_kv);     // redundant n_q > 1 check ?
+
+    const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
+    const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;
+    const float m0 = powf(2.0f, -(max_bias) / n_head_log2_f);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2_f);
+
+    const bool is_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 && v->type == GGML_TYPE_Q8_0;
+
+    ggml_cl_flash_attn_temp_buffer temp_k;
+    ggml_cl_flash_attn_temp_buffer temp_v;
+    ggml_cl_flash_attn_temp_buffer temp_k_aos;
+    ggml_cl_flash_attn_temp_buffer temp_v_aos;
+
+    if (is_q8_0) {
+        ggml_cl_flash_attn_reconstruct_aos(
+            backend_ctx, k, temp_k_aos, k_data_device, offset_k, k_nb1, k_nb2, k_nb3);
+
+        ggml_cl_flash_attn_reconstruct_aos(
+            backend_ctx, v, temp_v_aos, v_data_device, offset_v, v_nb1, v_nb2, v_nb3);
+
+        bool k_done = ggml_cl_flash_attn_dequant_kv_gpu(
+            backend_ctx, k, GGML_TYPE_F16, k_data_device, offset_k, k_nb1, k_nb2, k_nb3,
+            temp_k, k_data_device, offset_k, k_nb1, k_nb2, k_nb3);
+
+        bool v_done = ggml_cl_flash_attn_dequant_kv_gpu(
+            backend_ctx, v, GGML_TYPE_F16, v_data_device, offset_v, v_nb1, v_nb2, v_nb3,
+            temp_v, v_data_device, offset_v, v_nb1, v_nb2, v_nb3);
+
+        GGML_ASSERT(k_done && v_done);
+    }
+
+    // Allocate input/output memory buffers
+    cl_mem mem_matrixQ;
+    cl_mem mem_matrixK;
+    cl_mem mem_matrixV;
+    cl_mem mem_matrixO;
+    cl_buffer_region region;
+    cl_int err;
+
+    region.origin = offset_q;
+    region.size = ggml_nbytes(q);
+    mem_matrixQ = clCreateSubBuffer(extra_q->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+    CL_CHECK(err);
+
+    region.origin = offset_k;
+    region.size = is_q8_0 ? (size_t) k_nb3 * (size_t) k->ne[3] : ggml_nbytes(k);
+    mem_matrixK = clCreateSubBuffer(k_data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+    CL_CHECK(err);
+
+    region.origin = offset_v;
+    region.size = is_q8_0 ? (size_t) v_nb3 * (size_t) v->ne[3] : ggml_nbytes(v);
+    mem_matrixV = clCreateSubBuffer(v_data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+    CL_CHECK(err);
+
+    region.origin = offset_o;
+    region.size = ggml_nbytes(dst);
+    mem_matrixO = clCreateSubBuffer(extra_o->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+    CL_CHECK(err);
+
+    cl_image_format img_fmt_1d = { CL_RGBA, CL_FLOAT};
+    cl_image_desc img_desc_1d;
+
+    // use image 1d buffer used as fallback when on mask is applied
+    cl_mem  mem_tex_mask_fallback_1dbuf;
+    img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = 1;
+    img_desc_1d.buffer = mem_matrixK;
+    mem_tex_mask_fallback_1dbuf = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &err);
+    CL_CHECK(err);
+
+    cl_mem  mem_tex_matrixO_1dbuf;
+    img_fmt_1d = { CL_RGBA, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = ggml_nbytes(dst) / 4 / 4;
+    img_desc_1d.buffer = mem_matrixO;
+    mem_tex_matrixO_1dbuf = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &err);
+    CL_CHECK(err);
+
+    // The bin kernel requires 2d (or 3d) buffers packed for data loading/multiplication.
+    // These repack kernels launch across all buffers to ensure compatibility
+    cl_mem  mem_tex_matrixMask_1dbuf = NULL;
+    cl_mem  mem_matrixMask = NULL;
+    cl_mem  mem_matrixMask_padded = NULL;
+    cl_ulong mask_nb1_padded = mask_nb1, mask_nb2_padded = mask_nb2, mask_nb3_padded = mask_nb3;
+    if (extra_mask) {
+        // allocate mem_matrixMask w/ new padded size
+        size_t n_kv_padded = GGML_PAD(n_kv, 4);
+        size_t mask_nb_padded = n_kv_padded * sizeof(cl_half) * mask->ne[1] * mask->ne[2] * mask->ne[3];
+
+        // apply offset and create subBuffer for mask
+        region.origin = offset_mask;
+        region.size = ggml_nbytes(mask);
+        mem_matrixMask = clCreateSubBuffer(extra_mask->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        {
+            // create padded mask to contain all data
+            mem_matrixMask_padded = clCreateBuffer(context, CL_MEM_ALLOC_HOST_PTR, mask_nb_padded, NULL, &err);
+            CL_CHECK(err);
+
+            // pass extra_mask->data_device, mem_matrixMask to kernel for copying/padding
+            mask_nb1_padded = (cl_ulong)n_kv_padded * sizeof(cl_half);
+            mask_nb2_padded = mask_nb1_padded * (cl_ulong)mask->ne[1];
+            mask_nb3_padded = mask_nb2_padded * (cl_ulong)mask->ne[2];
+
+            cl_kernel repack_mask = backend_ctx->fa.kernel_repack_mask_for_wmm;
+            CL_CHECK(clSetKernelArg(repack_mask, 0, sizeof(cl_mem),   &mem_matrixMask));
+            CL_CHECK(clSetKernelArg(repack_mask, 1, sizeof(cl_ulong), &mask_nb1));
+            CL_CHECK(clSetKernelArg(repack_mask, 2, sizeof(cl_ulong), &mask_nb2));
+            CL_CHECK(clSetKernelArg(repack_mask, 3, sizeof(cl_ulong), &mask_nb3));
+            CL_CHECK(clSetKernelArg(repack_mask, 4, sizeof(int),      &mask_ne2));
+            CL_CHECK(clSetKernelArg(repack_mask, 5, sizeof(cl_mem),   &mem_matrixMask_padded));
+            CL_CHECK(clSetKernelArg(repack_mask, 6, sizeof(cl_ulong), &mask_nb1_padded));
+            CL_CHECK(clSetKernelArg(repack_mask, 7, sizeof(cl_ulong), &mask_nb2_padded));
+            CL_CHECK(clSetKernelArg(repack_mask, 8, sizeof(cl_ulong), &mask_nb3_padded));
+
+            size_t repack_mask_gws[3] = {(size_t)n_kv, (size_t)mask->ne[1], (size_t)mask_ne2 * (size_t)mask->ne[3]};
+            backend_ctx->enqueue_ndrange_kernel(repack_mask, 3, repack_mask_gws, NULL, dst);
+        }
+
+        // use image 1d buffer for matrix Mask (padded row stride)
+        cl_image_format img_fmt_mask_1d = { CL_RGBA, CL_HALF_FLOAT};
+        cl_image_desc img_desc_mask_1d;
+        memset(&img_desc_mask_1d, 0, sizeof(img_desc_mask_1d));
+        img_desc_mask_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_mask_1d.image_width = mask_nb_padded / 2 / 4;
+        img_desc_mask_1d.buffer = mem_matrixMask_padded;
+        mem_tex_matrixMask_1dbuf = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_mask_1d, &img_desc_mask_1d, NULL, &err);
+        CL_CHECK(err);
+    }
+
+    // WMM QK uses repacked 3D images.
+    // Q image: rows, heads, packed depth.
+    cl_image_format img_fmt_3d = { CL_RGBA, CL_HALF_FLOAT };
+    cl_image_desc   img_desc_3d;
+
+    memset(&img_desc_3d, 0, sizeof(img_desc_3d));
+    img_desc_3d.image_type   = CL_MEM_OBJECT_IMAGE3D;
+    img_desc_3d.image_width  = (size_t)n_q;
+    img_desc_3d.image_height = (size_t)n_batch * (size_t)n_head;
+    img_desc_3d.image_depth  = (size_t)d_head_q / 4;
+    cl_mem img_q_wmm = NULL;
+    img_q_wmm = clCreateImage(context, CL_MEM_READ_WRITE, &img_fmt_3d, &img_desc_3d, NULL, &err);
+    CL_CHECK(err);
+
+    {
+        cl_kernel repack_q = backend_ctx->fa.kernel_repack_q_for_wmm;
+        CL_CHECK(clSetKernelArg(repack_q, 0, sizeof(cl_mem),   &mem_matrixQ));
+        CL_CHECK(clSetKernelArg(repack_q, 1, sizeof(cl_ulong), &q_nb1));
+        CL_CHECK(clSetKernelArg(repack_q, 2, sizeof(cl_ulong), &q_nb2));
+        CL_CHECK(clSetKernelArg(repack_q, 3, sizeof(cl_ulong), &q_nb3));
+        CL_CHECK(clSetKernelArg(repack_q, 4, sizeof(int),      &n_head));
+        CL_CHECK(clSetKernelArg(repack_q, 5, sizeof(cl_mem),   &img_q_wmm));
+
+        size_t repack_q_gws[3] = {(size_t)d_head_q / 4, (size_t)n_q, (size_t)n_batch * (size_t)n_head};
+        backend_ctx->enqueue_ndrange_kernel(repack_q, 3, repack_q_gws, NULL, dst);
+    }
+
+    // K image: columns, row groups, KV heads.
+    const size_t n_kv_row4 = ((size_t)n_kv + 3) / 4;
+
+    memset(&img_desc_3d, 0, sizeof(img_desc_3d));
+    img_desc_3d.image_type   = CL_MEM_OBJECT_IMAGE3D;
+    img_desc_3d.image_width  = (size_t)d_head_q;
+    img_desc_3d.image_height = n_kv_row4;
+    img_desc_3d.image_depth  = (size_t)n_batch * (size_t)n_head_kv;
+    cl_mem img_k_wmm = NULL;
+    img_k_wmm = clCreateImage(context, CL_MEM_READ_WRITE, &img_fmt_3d, &img_desc_3d, NULL, &err);
+    CL_CHECK(err);
+
+    {
+        cl_kernel repack_k = backend_ctx->fa.kernel_repack_k_for_wmm;
+        CL_CHECK(clSetKernelArg(repack_k, 0, sizeof(cl_mem),   &mem_matrixK));
+        CL_CHECK(clSetKernelArg(repack_k, 1, sizeof(cl_ulong), &k_nb1));
+        CL_CHECK(clSetKernelArg(repack_k, 2, sizeof(cl_ulong), &k_nb2));
+        CL_CHECK(clSetKernelArg(repack_k, 3, sizeof(cl_ulong), &k_nb3));
+        CL_CHECK(clSetKernelArg(repack_k, 4, sizeof(int),      &n_head_kv));
+        CL_CHECK(clSetKernelArg(repack_k, 5, sizeof(int),      &n_kv));
+        CL_CHECK(clSetKernelArg(repack_k, 6, sizeof(cl_mem),   &img_k_wmm));
+
+        size_t repack_k_gws[3] = {(size_t)d_head_q, n_kv_row4, (size_t)n_batch * (size_t)n_head_kv};
+        backend_ctx->enqueue_ndrange_kernel(repack_k, 3, repack_k_gws, NULL, dst);
+    }
+
+    // V image: kv-rows (contracted), packed head-dim groups, KV heads.
+    memset(&img_desc_3d, 0, sizeof(img_desc_3d));
+    img_desc_3d.image_type   = CL_MEM_OBJECT_IMAGE3D;
+    img_desc_3d.image_width  = (size_t)n_kv;
+    img_desc_3d.image_height = (size_t)d_head_v / 4;
+    img_desc_3d.image_depth  = (size_t)n_batch * (size_t)n_head_kv;
+    cl_mem img_v_wmm = NULL;
+    img_v_wmm = clCreateImage(context, CL_MEM_READ_WRITE, &img_fmt_3d, &img_desc_3d, NULL, &err);
+    CL_CHECK(err);
+
+    {
+        cl_kernel repack_v = backend_ctx->fa.kernel_repack_v_for_wmm;
+        CL_CHECK(clSetKernelArg(repack_v, 0, sizeof(cl_mem),   &mem_matrixV));
+        CL_CHECK(clSetKernelArg(repack_v, 1, sizeof(cl_ulong), &v_nb1));
+        CL_CHECK(clSetKernelArg(repack_v, 2, sizeof(cl_ulong), &v_nb2));
+        CL_CHECK(clSetKernelArg(repack_v, 3, sizeof(cl_ulong), &v_nb3));
+        CL_CHECK(clSetKernelArg(repack_v, 4, sizeof(int),      &n_head_kv));
+        CL_CHECK(clSetKernelArg(repack_v, 5, sizeof(cl_mem),   &img_v_wmm));
+
+        size_t repack_v_gws[3] = {(size_t)d_head_v / 4, (size_t)n_kv, (size_t)n_batch * (size_t)n_head_kv};
+        backend_ctx->enqueue_ndrange_kernel(repack_v, 3, repack_v_gws, NULL, dst);
+    }
+
+    cl_int enable_mask = (extra_mask) ? 1 : 0;
+    mask_buffer = extra_mask ? mem_tex_matrixMask_1dbuf : mem_tex_mask_fallback_1dbuf;
+
+    cl_mem mem_sinksBuf = NULL;
+    cl_mem mem_tex_sinks_1dbuf = NULL;
+    cl_int enable_sinks = (sinks_buffer != NULL) ? 1 : 0;
+    if (enable_sinks) {
+        region.origin = offset_sinks;
+        region.size = ggml_nbytes(sinks);
+        mem_sinksBuf = clCreateSubBuffer(extra_sinks->data_device, CL_MEM_READ_ONLY, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_image_format img_fmt_sinks_1d = { CL_R, CL_FLOAT };
+        cl_image_desc img_desc_sinks_1d;
+        memset(&img_desc_sinks_1d, 0, sizeof(img_desc_sinks_1d));
+        img_desc_sinks_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_sinks_1d.image_width = (size_t)n_head;
+        img_desc_sinks_1d.buffer = mem_sinksBuf;
+        mem_tex_sinks_1dbuf = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_sinks_1d, &img_desc_sinks_1d, NULL, &err);
+        CL_CHECK(err);
+    } else {
+        // The image obj cannot be null so we back with buffer of size 1 and use matrixK to back because it always exists
+        cl_image_format img_fmt_sinks_fallback = { CL_R, CL_FLOAT };
+        cl_image_desc img_desc_sinks_fallback;
+        memset(&img_desc_sinks_fallback, 0, sizeof(img_desc_sinks_fallback));
+        img_desc_sinks_fallback.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_sinks_fallback.image_width = 1;
+        img_desc_sinks_fallback.buffer = mem_matrixK;
+        mem_tex_sinks_1dbuf = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_sinks_fallback, &img_desc_sinks_fallback, NULL, &err);
+        CL_CHECK(err);
+    }
+
+    cl_uint arg = 0;
+
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_mem),   &mem_tex_matrixO_1dbuf));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &n_q));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &n_kv));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &is_causal));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &n_head));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &q_nb1));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &q_nb2));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &q_nb3));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &k_nb1));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &k_nb2));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &k_nb3));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &v_nb1));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &v_nb2));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &v_nb3));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &o_nb1));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &o_nb2));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &o_nb3));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &n_head_log2_val));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(float),    &logit_softcap));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &n_head_kv));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_mem),   &mask_buffer));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &enable_mask));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &mask_nb1_padded));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &mask_nb2_padded));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_ulong), &mask_nb3_padded));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &mask_ne2));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &mask_ne3));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_mem),   &mem_tex_sinks_1dbuf));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &enable_sinks));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_mem),   &img_q_wmm));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_mem),   &img_k_wmm));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(cl_mem),   &img_v_wmm));
+    CL_CHECK(clSetKernelArg(kernel, arg++, sizeof(int),      &d_head_q));
+
+    size_t global_work_size[3], local_work_size[3];
+
+    const int n_waves_v = d_head_q / 64;
+
+    local_work_size[0] = 64;
+    local_work_size[1] = n_waves_v;
+    local_work_size[2] = 1;
+
+    global_work_size[0] = 64;
+    global_work_size[1] = ((n_q + 64 - 1) / 64) * n_waves_v;
+    global_work_size[2] = n_batch * n_head;
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+    CL_CHECK(clReleaseMemObject(mem_tex_matrixO_1dbuf));
+    CL_CHECK(clReleaseMemObject(img_q_wmm));
+    CL_CHECK(clReleaseMemObject(img_k_wmm));
+    CL_CHECK(clReleaseMemObject(img_v_wmm));
+
+    if (mem_tex_matrixMask_1dbuf) {
+        CL_CHECK(clReleaseMemObject(mem_tex_matrixMask_1dbuf));
+    }
+    if (mem_matrixMask) {
+        CL_CHECK(clReleaseMemObject(mem_matrixMask));
+    }
+    if (mem_matrixMask_padded) {
+        CL_CHECK(clReleaseMemObject(mem_matrixMask_padded));
+    }
+    if (mem_tex_sinks_1dbuf) {
+        CL_CHECK(clReleaseMemObject(mem_tex_sinks_1dbuf));
+    }
+    if (mem_sinksBuf) {
+        CL_CHECK(clReleaseMemObject(mem_sinksBuf));
+    }
+    CL_CHECK(clReleaseMemObject(mem_matrixQ));
+    CL_CHECK(clReleaseMemObject(mem_matrixK));
+    CL_CHECK(clReleaseMemObject(mem_matrixV));
+    CL_CHECK(clReleaseMemObject(mem_matrixO));
+}
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
 static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
    const ggml_tensor * v = dst->src[2];
    const ggml_tensor * mask = dst->src[3];
@@ -17164,6 +17728,14 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
    const bool is_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 && v->type == GGML_TYPE_Q8_0;
    const bool is_q4_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q4_0 && v->type == GGML_TYPE_Q4_0;

+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    if (use_fa_bin_kernels_prefill(backend_ctx, q, k, v)) {
+        // We support the prefill path of flash attn with a specialized d_head = 64/128/256
+        ggml_cl_flash_attn_prefill_bin(backend, q, k, dst);
+        return;
+    }
+#endif
+
    if (is_f16) {
        ggml_opencl_ensure_fa_variant(backend_ctx, d_head_q, d_head_v, FA_VARIANT_F16);
    } else if (is_mixed) {
@@ -21112,6 +21684,145 @@ static void ggml_cl_mul_mat_q4_k_f32_adreno(ggml_backend_t backend, const ggml_t
 #endif
 }

+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+static void ggml_cl_mul_mat_q6_K_f32_adreno_ila(ggml_backend_t backend, const ggml_tensor * src0,
+                                                const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+
+    const int ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int           err;
+    cl_buffer_region region;
+    cl_image_format  img_fmt;
+    cl_image_desc    img_desc;
+
+    const int M = ne01;
+    const int N = ne1;
+    const int K = ne00;
+
+    if (ne1 == 1) {
+        cl_mem b_sub_buf  = nullptr;
+        cl_mem b_img      = nullptr;
+
+        region.origin = offset1;
+        region.size   = (size_t)K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        img_fmt = { CL_RGBA, CL_FLOAT };
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = (size_t)K * N / 4;
+        img_desc.buffer      = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_q6_k_f32_32b_trans;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0_q6_K->ql_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q6_K->qh_img));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q6_K->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0_q6_K->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne01));
+
+        size_t local_work_size[3]  = { 64, 8, 1 };
+        size_t global_work_size[3] = { (size_t)ne01, 8, 1 };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+    } else {
+        const int gemm_tile_n = 64;
+        int N_pad = CEIL_DIV(N, gemm_tile_n) * gemm_tile_n;
+
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_padded  = nullptr;
+        cl_mem b_buf     = nullptr;
+        if (N_pad == N) {
+            region.origin = offset1;
+            region.size   = (size_t)K * N * sizeof(float);
+            CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+            b_buf = b_sub_buf;
+        } else {
+            CL_CHECK((b_padded = clCreateBuffer(context, CL_MEM_READ_WRITE, (size_t)K * N_pad * sizeof(float), NULL, &err), err));
+            const float zero = 0.0f;
+            CL_CHECK(clEnqueueFillBuffer(backend_ctx->queue, b_padded, &zero, sizeof(zero), 0, (size_t)K * N_pad * sizeof(float), 0, NULL, NULL));
+            CL_CHECK(clEnqueueCopyBuffer(backend_ctx->queue, extra1->data_device, b_padded, offset1, 0, (size_t)K * N * sizeof(float), 0, NULL, NULL));
+            b_buf = b_padded;
+        }
+
+        img_fmt = { CL_R, CL_FLOAT };
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = (size_t)K * N_pad;
+        img_desc.buffer      = b_buf;
+        cl_mem b_img;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        region.origin = offsetd;
+        region.size   = (size_t)M * N * sizeof(float);
+        cl_mem d_sub_buf;
+        CL_CHECK((d_sub_buf = clCreateSubBuffer(extrad->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+        img_fmt = { CL_R, CL_FLOAT };
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = (size_t)M * N;
+        img_desc.buffer      = d_sub_buf;
+        cl_mem d_img;
+        CL_CHECK((d_img = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),  &extra0_q6_K->ql_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),  &extra0_q6_K->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),  &extra0_q6_K->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),  &extra0_q6_K->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),  &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),  &d_img));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uint), &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_uint), &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),     &N));
+
+        size_t local_work_size[3]  = { 64, 2, 2 };
+        size_t m_tiles = (size_t)CEIL_DIV(M, 64);
+        size_t global_work_size[3] = { 64, m_tiles, (size_t)CEIL_DIV(N_pad, gemm_tile_n) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_img));
+        if (b_sub_buf) {
+            CL_CHECK(clReleaseMemObject(b_sub_buf));
+        }
+        if (b_padded) {
+            CL_CHECK(clReleaseMemObject(b_padded));
+        }
+        CL_CHECK(clReleaseMemObject(d_img));
+        CL_CHECK(clReleaseMemObject(d_sub_buf));
+    }
+}
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
 static void ggml_cl_mul_mat_q6_K_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    GGML_ASSERT(src0);
@@ -21160,6 +21871,20 @@ static void ggml_cl_mul_mat_q6_K_f32_adreno(ggml_backend_t backend, const ggml_t
    // (the #1 MTP bottleneck; mc3 above can't, it reads the noshuffle layout).
    const bool use_q6k_tiled_mc = q6k_mc3 && (ne1 == 3) && (ne01 >= 32768) && use_q6k_tiled(backend_ctx, src0);

+    const bool use_bin = use_q6_k_bin_kernels(backend_ctx, src0);
+
+    if (use_bin) {
+        if (use_q6k_mc3 || use_q6k_tiled_mc) {
+            static bool warned = false;
+            if (!warned) {
+                GGML_LOG_WARN("ggml_opencl: GGML_OPENCL_Q6K_MC3 is bypassed by Q6_K binary kernels\n");
+                warned = true;
+            }
+        }
+        ggml_cl_mul_mat_q6_K_f32_adreno_ila(backend, src0, src1, dst);
+        return;
+    }
+
    if (ne1 == 1 || use_q6k_mc3 || use_q6k_tiled_mc) {
        cl_mem ql_img = nullptr;
        cl_mem qh_img = nullptr;
@@ -0,0 +1,92 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+__kernel void kernel_repack_mask_for_wmm(
+    const global half* mask_buf,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    global half* mask_buf_padded,
+    const ulong mask_nb1_padded,
+    const ulong mask_nb2_padded,
+    const ulong mask_nb3_padded
+) {
+    int col   = get_global_id(0);   // 0 .. n_kv
+    int row   = get_global_id(1);   // 0 .. n_q
+    int slice = get_global_id(2);   // 0 .. (n_head * n_batch)
+
+    int head_idx  = slice % mask_ne2;
+    int batch_idx = slice / mask_ne2;
+
+    ulong src_off = (ulong)batch_idx * mask_nb3 + (ulong)head_idx * mask_nb2 + (ulong)row * mask_nb1;
+    ulong dst_off = (ulong)batch_idx * mask_nb3_padded + (ulong)head_idx * mask_nb2_padded + (ulong)row * mask_nb1_padded;
+
+    mask_buf_padded[dst_off / 2 + col] = mask_buf[src_off / 2 + col];
+}
+
+__kernel void kernel_repack_q_for_wmm(
+    const global float* q_buf,
+    const ulong q_nb1,
+    const ulong q_nb2,
+    const ulong q_nb3,
+    const int n_head,
+    __write_only image3d_t img_q_wmm
+) {
+    int k4    = get_global_id(0);
+    int row   = get_global_id(1);
+    int slice = get_global_id(2);
+    int batch_idx = slice / n_head;
+    int head_idx  = slice % n_head;
+
+
+    ulong elem_off = (batch_idx * q_nb3 + head_idx * q_nb2 + row * q_nb1) / 4 + (ulong)k4 * 4;
+    float4 v = vload4(elem_off / 4, q_buf);
+
+    write_imageh(img_q_wmm, (int4)(row, slice, k4, 0), convert_half4(v));
+}
+
+__kernel void kernel_repack_k_for_wmm(
+    const global half* k_buf,
+    const ulong k_nb1,
+    const ulong k_nb2,
+    const ulong k_nb3,
+    const int n_head_kv,
+    const int n_kv,
+    __write_only image3d_t img_k_wmm
+) {
+    int kk    = get_global_id(0);
+    int row4  = get_global_id(1);
+    int slice = get_global_id(2);
+    int batch_idx   = slice / n_head_kv;
+    int head_kv_idx = slice % n_head_kv;
+
+    ulong base = batch_idx * k_nb3 + head_kv_idx * k_nb2;
+    int row0 = row4 * 4;
+    half4 v;
+    v.x = (row0 + 0 < n_kv) ? k_buf[(base + (ulong)(row0 + 0) * k_nb1) / 2 + kk] : (half)0;
+    v.y = (row0 + 1 < n_kv) ? k_buf[(base + (ulong)(row0 + 1) * k_nb1) / 2 + kk] : (half)0;
+    v.z = (row0 + 2 < n_kv) ? k_buf[(base + (ulong)(row0 + 2) * k_nb1) / 2 + kk] : (half)0;
+    v.w = (row0 + 3 < n_kv) ? k_buf[(base + (ulong)(row0 + 3) * k_nb1) / 2 + kk] : (half)0;
+
+    write_imageh(img_k_wmm, (int4)(kk, row4, slice, 0), v);
+}
+
+__kernel void kernel_repack_v_for_wmm(
+      const global half* v_buf,
+      const ulong v_nb1,
+      const ulong v_nb2,
+      const ulong v_nb3,
+      const int n_head_kv,
+      __write_only image3d_t img_v_wmm
+) {
+    int hdim4   = get_global_id(0);   // now fastest — walks contiguous memory
+    int row     = get_global_id(1);
+    int slice   = get_global_id(2);
+    int batch_idx   = slice / n_head_kv;
+    int head_kv_idx = slice % n_head_kv;
+
+    ulong row_off = batch_idx * v_nb3 + head_kv_idx * v_nb2 + (ulong)row * v_nb1;
+    half4 v = vload4((row_off / 2 + (ulong)hdim4 * 4) / 4, v_buf);
+
+    write_imageh(img_v_wmm, (int4)(row, hdim4, slice, 0), v);
+}
@@ -0,0 +1,128 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#define QK_K 256
+#define N_SIMDGROUP 8
+#define SIMDGROUP_WIDTH 64
+
+static inline float8 q6_k_to_fp32_packed8(ushort2 ql8, ushort qh8, float d_scale) {
+    float8 fp32x8;
+    fp32x8.s0 = ((float)(( ql8.s0 & 0x000F)        | ((uint)((qh8      ) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s1 = ((float)((( ql8.s0 >> 4) & 0x000F) | ((uint)((qh8 >> 2) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s2 = ((float)((( ql8.s0 >> 8) & 0x000F) | ((uint)((qh8 >> 4) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s3 = ((float)((( ql8.s0 >> 12)& 0x000F) | ((uint)((qh8 >> 6) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s4 = ((float)(( ql8.s1 & 0x000F)        | ((uint)((qh8 >> 8) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s5 = ((float)((( ql8.s1 >> 4) & 0x000F) | ((uint)((qh8 >>10) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s6 = ((float)((( ql8.s1 >> 8) & 0x000F) | ((uint)((qh8 >>12) & 0x3) << 4)) - 32.f) * d_scale;
+    fp32x8.s7 = ((float)((( ql8.s1 >> 12)& 0x000F) | ((uint)((qh8 >>14) & 0x3) << 4)) - 32.f) * d_scale;
+    return fp32x8;
+}
+
+__attribute__((qcom_reqd_sub_group_size("half")))
+__kernel void kernel_gemv_noshuffle_q6_k_f32_32b_trans(
+    __read_only image1d_buffer_t src0_ql,
+    __read_only image1d_buffer_t src0_qh,
+    __global char *         src0_s,
+    __global half *         src0_d,
+    __read_only image1d_buffer_t src1,
+    __global float *        dst,
+    ulong                   offsetd,
+    int                     ne00,
+    int                     ne01
+) {
+    uint i01  = get_global_id(0);
+    uint sgid = get_local_id(1);
+    uint slid = get_sub_group_local_id();
+
+    int num_superblocks = ne00 / QK_K;
+    int num_subblocks   = ne00 / 32;    // 2 sub-blocks of 16 processed per iter below
+    int scales_per_row   = num_superblocks * 16;
+
+    __private float sum = 0.0f;
+
+    // Loop over 32-element groups (2 sub-blocks of 16 each), N_SIMDGROUP groups per iter.
+    for (uint ib = sgid; ib < num_subblocks; ib += N_SIMDGROUP) {
+        uint sb = ib / 8;   // super-block index
+        uint j  = ib % 8;   // 32-element group within super-block (0..7)
+
+        // Load d for this super-block.
+        half d_val = src0_d[sb * ne01 + i01];
+
+        // Load 2 sub-block scales (int8), one per 16 elements.
+        global const char * sc = src0_s + i01 * scales_per_row + sb * 16;
+        float scale0 = (float)d_val * (float)sc[j * 2];
+        float scale1 = (float)d_val * (float)sc[j * 2 + 1];
+
+        // Load 4 uints of ql (32 elements, 4-bit each = 128 bits), column-major stride ne01.
+        uint ql_base = (ib * 4) * ne01 + i01;
+        uint4 regQL;
+        regQL.s0 = read_imageui(src0_ql, ql_base).x;
+        regQL.s1 = read_imageui(src0_ql, ql_base + ne01).x;
+        regQL.s2 = read_imageui(src0_ql, ql_base + ne01 * 2).x;
+        regQL.s3 = read_imageui(src0_ql, ql_base + ne01 * 3).x;
+
+        // Load 2 uints of qh (32 elements, 2-bit each = 64 bits), column-major stride ne01.
+        uint qh_base = (ib * 2) * ne01 + i01;
+        uint2 regQH;
+        regQH.s0 = read_imageui(src0_qh, qh_base).x;
+        regQH.s1 = read_imageui(src0_qh, qh_base + ne01).x;
+
+        // Load activations: 32 floats = 8 float4s.
+        uint y_offset = ib * 8;
+
+        float4 y_local = (slid < 8) ? read_imagef(src1, (y_offset + slid)) : (float4)0.0f;
+        float4 y0 = sub_group_broadcast(y_local, 0);
+        float4 y1 = sub_group_broadcast(y_local, 1);
+        float4 y2 = sub_group_broadcast(y_local, 2);
+        float4 y3 = sub_group_broadcast(y_local, 3);
+        float4 y4v = sub_group_broadcast(y_local, 4);
+        float4 y5 = sub_group_broadcast(y_local, 5);
+        float4 y6 = sub_group_broadcast(y_local, 6);
+        float4 y7 = sub_group_broadcast(y_local, 7);
+
+        // Dequantize elements 0..7 (scale0).
+        float8 fp32x8 = q6_k_to_fp32_packed8(as_ushort2(regQL.s0), (ushort)(regQH.s0 & 0xFFFF), scale0);
+
+        float4 acc = y0 * fp32x8.lo;
+        acc += y1 * fp32x8.hi;
+
+        // Dequantize elements 8..15 (scale0).
+        fp32x8 = q6_k_to_fp32_packed8(as_ushort2(regQL.s1), (ushort)(regQH.s0 >> 16), scale0);
+
+        acc += y2 * fp32x8.lo;
+        acc += y3 * fp32x8.hi;
+
+        // Dequantize elements 16..23 (scale1).
+        fp32x8 = q6_k_to_fp32_packed8(as_ushort2(regQL.s2), (ushort)(regQH.s1 & 0xFFFF), scale1);
+
+        acc += y4v * fp32x8.lo;
+        acc += y5 * fp32x8.hi;
+
+        // Dequantize elements 24..31 (scale1).
+        fp32x8 = q6_k_to_fp32_packed8(as_ushort2(regQL.s3), (ushort)(regQH.s1 >> 16), scale1);
+
+        acc += y6 * fp32x8.lo;
+        acc += y7 * fp32x8.hi;
+
+        sum += ((acc.s0 + acc.s1) + (acc.s2 + acc.s3));
+    }
+
+    // reduction in local memory, assumes #subgroups=4
+    __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)];
+    if (sgid > 0) {
+        reduceLM[SIMDGROUP_WIDTH * (sgid - 1) + slid] = sum;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgid == 0) {
+        for (uint i = 0; i < N_SIMDGROUP - 1; ++i) {
+            sum += reduceLM[SIMDGROUP_WIDTH * i + slid];
+        }
+    }
+
+    // 1 output per thread in subgroup 0
+    if (sgid == 0) {
+        dst = dst + (offsetd >> 2);
+        dst[i01] = sum;
+    }
+}
@@ -245,7 +245,7 @@ void GgmlOvDecoder::set_input_output() {
            if (src->op == GGML_OP_VIEW) {
                // Traverse upward through nested VIEW operations
                std::remove_reference_t<decltype(current_node_info.node_inputs_views[src_name])> view_chain;
-                auto current = src;
+                auto * current = src;

                while (current != nullptr) {
                    auto current_name = get_tensor_ov_name(m_cgraph, current);
@@ -612,9 +612,8 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr
                if (node->src[1]->view_src != nullptr) {
                    if (node->src[3] != nullptr) {
                        return 4;  // decoder self-attention
-                    } else {
-                        return 5;  // cross-attention or encoder self-attention
-                    };
+                    }
+                    return 5;      // cross-attention or encoder self-attention
                }
                break;
            default:
@@ -736,8 +735,7 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr

    bool rope_seen = false;
    for (int i = 0; i < cgraph->n_nodes; i++) {
-        auto * node = cgraph->nodes[i];
-        std::string name = std::string(node->name);
+        ggml_tensor * node = cgraph->nodes[i];
        const int attention_pattern_case = get_attention_pattern_case(node);
        if (attention_pattern_case != -1) {
            ggml_tensor * cache_k_permute = nullptr;
@@ -948,7 +946,6 @@ ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op,
    if (m_naive) {
        return input != nullptr ? ov::PartialShape{get_shape(input)} : ov::PartialShape{get_shape(op)};
    }
-    auto name = std::string(input->name);
    ov::PartialShape input_shape;

    if (is_inp_tok(input, op) || is_inp_pos(input, op)) {
@@ -1474,7 +1471,7 @@ std::shared_ptr<ov::Node> GgmlOvDecoder::create_weight_node(ggml_tensor * tensor
 void GgmlOvDecoder::dump_cgraph(const ggml_cgraph * cgraph, std::string & filename) {
    std::ofstream file(filename);
    if (!file.is_open()) {
-        std::cerr << "Failed to open file" << std::endl;
+        std::cerr << "Failed to open file" << '\n';
        return;
    }

@@ -1580,11 +1577,11 @@ void print_tensor_address_map(const ggml_cgraph * cgraph) {
        }
    }
    for (const auto & pair : address_map) {
-        std::cout << "Address: " << pair.first << std::endl;
+        std::cout << "Address: " << pair.first << '\n';
        for (const auto & name : pair.second) {
            std::cout << name << " ; ";
        }
-        std::cout << std::endl << std::endl;
+        std::cout << "\n\n";
    }
 }

@@ -2226,7 +2223,7 @@ void GgmlOvDecoder::compute_node_dynamic_dims() {
                    std::cout << ", ";
                }
            }
-            std::cout << "]" << std::endl;
+            std::cout << "]" << '\n';
            // print the src name & shape with the dynamic dim for debugging
            for (int j = 0; j < GGML_MAX_SRC; j++) {
                ggml_tensor * src = node->src[j];
@@ -2245,9 +2242,9 @@ void GgmlOvDecoder::compute_node_dynamic_dims() {
                        std::cout << ", ";
                    }
                }
-                std::cout << "]" << std::endl;
+                std::cout << "]" << '\n';
            }
-            std::cout << std::endl;
+            std::cout << '\n';
        }
    }
 }
@@ -354,41 +354,41 @@ public:

    void update_io(ggml_cgraph * cgraph);

-    inline static bool is_inp_tok(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_inp_tok(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] && op->src[0]->op == GGML_OP_NONE;
    }

-    inline static bool is_inp_pos(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_inp_pos(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_ROPE && tensor == op->src[1];
    }

    // IMROPE packs 4 stacked position planes (t/h/w/e) into inp_pos, each of length
    // n_tokens; other modes carry a single position per token.
-    inline static int get_inp_pos_n_planes(const ggml_tensor * op) {
+    static int get_inp_pos_n_planes(const ggml_tensor * op) {
        return op->op_params[2] == GGML_ROPE_TYPE_IMROPE ? 4 : 1;
    }

-    inline static bool is_inp_emb(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_inp_emb(const ggml_tensor * tensor, const ggml_tensor * op) {
        return tensor->op == GGML_OP_GET_ROWS && op->op == GGML_OP_RMS_NORM;
    }

-    inline static bool is_inp_mask(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_inp_mask(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_CPY || (op->op == GGML_OP_FLASH_ATTN_EXT && tensor == op->src[3]) ||
               (op->op == GGML_OP_SOFT_MAX && tensor == op->src[1]);
    }

-    inline static bool is_inp_mean(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_inp_mean(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_MUL_MAT && tensor == op->src[1] && tensor->op == GGML_OP_NONE &&
               (tensor->flags & GGML_TENSOR_FLAG_INPUT) && tensor->type == GGML_TYPE_F32 &&
               op->src[0] != nullptr && op->src[0]->op != GGML_OP_NONE;
    }

-    inline static bool is_rope_freqs_weight(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_rope_freqs_weight(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_ROPE && tensor == op->src[2];
    }

    // also returns true for cache_s and cache_r in SSM/DeltaNet models
-    inline static bool is_kvcache(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_kvcache(const ggml_tensor * tensor, const ggml_tensor * op) {
        if (tensor == nullptr) {
            return false;
        }
@@ -396,14 +396,14 @@ public:
               (op != nullptr && op->op == GGML_OP_SET_ROWS && op->src[2] == tensor);
    }

-    inline static bool is_conv_state_writeback(const ggml_tensor * node) {
+    static bool is_conv_state_writeback(const ggml_tensor * node) {
        return node->op == GGML_OP_CPY && node->view_src != nullptr && is_kvcache(node->view_src, nullptr) &&
               node->src[0] != nullptr && node->src[0]->op == GGML_OP_VIEW && node->src[0]->src[0] != nullptr &&
               node->src[0]->src[0]->op == GGML_OP_CONCAT && node->src[1] != nullptr &&
               node->src[1]->op == GGML_OP_VIEW && node->src[1]->view_src == node->view_src;
    }

-    inline static bool is_kv_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_kv_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_SET_ROWS && op->src[1] == tensor;
    }

@@ -411,13 +411,13 @@ public:
        return m_model_params.swa_mask != nullptr && tensor == m_model_params.swa_mask;
    }

-    inline static bool is_output_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_output_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] && op->src[0]->op != GGML_OP_NONE &&
               op->src[1]->op == GGML_OP_NONE;
    }

    // the state permutation index input used in SSM/DeltaNet models (inp->s_copy in llama-graph.cpp)
-    inline static bool is_inp_s_copy(const ggml_tensor * tensor, const ggml_tensor * op) {
+    static bool is_inp_s_copy(const ggml_tensor * tensor, const ggml_tensor * op) {
        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] &&
               op->src[0]->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY;
    }
@@ -481,5 +481,3 @@ private:
 };

 void print_tensor_address_map(const ggml_cgraph * cgraph);
-
-std::optional<int> extract_layer_from_name(const std::string & name);
@@ -472,10 +472,6 @@ ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_ten

    switch (tensor->type) {
    case GGML_TYPE_MXFP4:
-        layout.is_u4 = true;
-        layout.is_symmetric = true;
-        break;
-
    case GGML_TYPE_Q4_0:
        layout.is_u4 = true;
        layout.is_symmetric = true;
@@ -28,12 +28,7 @@
 #include <string>
 #include <vector>

-#ifndef _WIN32
-#    include <sys/mman.h>
-#    include <unistd.h>
-#endif
-
-#if defined(_WIN32)
+#ifdef _WIN32
 #    define WIN32_LEAN_AND_MEAN
 #    ifndef NOMINMAX
 #        define NOMINMAX
@@ -61,6 +56,7 @@
 // - CPU repack buffer: tensor->extra stores tensor_traits with repacked data
 // =====================================================

+namespace {
 // Buffer context that manages per-tensor allocations (no contiguous buffer for weights)
 struct ggml_backend_openvino_buffer_context {
    int device;
@@ -199,6 +195,7 @@ struct ggml_backend_openvino_buffer_type_context {
    int device;
    std::string name;
 };
+}  // namespace

 // =====================================================
 // Host weight-buffer release (GGML_OPENVINO_RELEASE_WEIGHTS)
@@ -258,14 +255,16 @@ void ggml_openvino_release_weight_buffers() {
    for (const auto & b : reg.buffers) {
        // Align down/up to page boundaries so madvise only drops whole pages
        // fully owned by this buffer.
-        const long page = sysconf(_SC_PAGESIZE);
-        uintptr_t start = reinterpret_cast<uintptr_t>(b.first);
-        uintptr_t end = start + b.second;
-        uintptr_t astart = (start + page - 1) & ~(uintptr_t) (page - 1);
-        uintptr_t aend = end & ~(uintptr_t) (page - 1);
-        if (aend > astart) {
-            if (madvise(reinterpret_cast<void *>(astart), aend - astart, MADV_DONTNEED) == 0) {
-                total += aend - astart;
+        const size_t page = (size_t) sysconf(_SC_PAGESIZE);
+        const uintptr_t ustart = reinterpret_cast<uintptr_t>(b.first);
+        const size_t offset_to_page = (page - (ustart & (page - 1))) & (page - 1);
+        if (b.second > offset_to_page) {
+            const size_t aligned_len = (b.second - offset_to_page) & ~(page - 1);
+            if (aligned_len > 0) {
+                char * astart = static_cast<char *>(b.first) + offset_to_page;
+                if (madvise(astart, aligned_len, MADV_DONTNEED) == 0) {
+                    total += aligned_len;
+                }
            }
        }
    }
@@ -876,11 +875,13 @@ GGML_BACKEND_API bool ggml_backend_is_openvino(ggml_backend_t backend) {
    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_openvino_guid());
 }

+namespace {
 struct ggml_backend_openvino_device_context {
    int device;
    std::string name;
    std::string description;
 };
+}

 static const char * ggml_backend_openvino_device_get_name(ggml_backend_dev_t dev) {
    ggml_backend_openvino_device_context * ctx = (ggml_backend_openvino_device_context *) dev->context;
@@ -1588,9 +1589,11 @@ static const struct ggml_backend_device_i ggml_backend_openvino_device_interface
    /* .event_synchronize    = */ NULL,
 };

+namespace {
 struct ggml_backend_openvino_reg_context {
    std::vector<ggml_backend_dev_t> devices;
 };
+}

 static const char * ggml_backend_openvino_reg_get_name(ggml_backend_reg_t reg) {
    return GGML_OPENVINO_NAME;
@@ -34,6 +34,15 @@
 #include <string>
 #include <vector>

+// From <openvino>/src/common/transformations/include/transformations/utils/utils.hpp
+namespace ov::op::util {
+// From <openvino>/src/common/transformations/include/transformations/utils/utils.hpp
+bool get_single_value(const std::shared_ptr<ov::op::v0::Constant> & const_node,
+                      float & value,
+                      bool check_value_range = true);
+}  // namespace ov::op::util
+
+namespace {
 void unpack_32_4(const uint8_t * data, uint8_t * dst) {
    std::fill_n(dst, 16, 0);
    for (int j = 0; j < 16; ++j) {
@@ -48,11 +57,11 @@ void unpack_32_4(const uint8_t * data, uint8_t * dst) {
    }
 }

-static constexpr size_t MXFP4_BLOCK_SIZE = 32;
-static constexpr size_t MXFP4_BLOCK_QS_SIZE = MXFP4_BLOCK_SIZE / 2;
-static constexpr size_t MXFP4_BLOCK_BYTES = sizeof(uint8_t) + MXFP4_BLOCK_QS_SIZE;
+constexpr size_t MXFP4_BLOCK_SIZE = 32;
+constexpr size_t MXFP4_BLOCK_QS_SIZE = MXFP4_BLOCK_SIZE / 2;
+constexpr size_t MXFP4_BLOCK_BYTES = sizeof(uint8_t) + MXFP4_BLOCK_QS_SIZE;

-static void pack_32_mxfp4_for_openvino(const uint8_t * data, uint8_t * dst) {
+void pack_32_mxfp4_for_openvino(const uint8_t * data, uint8_t * dst) {
    for (int j = 0; j < static_cast<int>(MXFP4_BLOCK_QS_SIZE); j += 2) {
        const uint8_t v0 = data[j] & 0x0F;
        const uint8_t v1 = (data[j + 1] & 0x0F) << 4;
@@ -419,7 +428,7 @@ void extract_q6_k_data(const ggml_tensor * tensor,
    }
 }

-static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t * d, uint8_t * m) {
+inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t * d, uint8_t * m) {
    if (j < 4) {
        *d = q[j] & 63;
        *m = q[j + 4] & 63;
@@ -514,9 +523,9 @@ void extract_q5_k_data(const ggml_tensor * tensor,
 ov::Output<ov::Node> make_int8_weights(ov::Tensor & weight,
                                       ov::Tensor & scales,
                                       ov::Tensor & zp,
-                                       size_t group_size,
-                                       bool use_bias,
-                                       bool for_gather_matmul) {
+                                       size_t group_size = GGML_QUANTIZATION_GROUP_SIZE,
+                                       bool use_bias = false,
+                                       bool for_gather_matmul = false) {
    ov::Shape orig_shape = weight.get_shape();
    bool is_signed = (weight.get_element_type() == ov::element::i8);  // Symmetric: signed weights, no ZP

@@ -611,13 +620,24 @@ ov::Output<ov::Node> make_int8_weights(ov::Tensor & weight,
    return std::make_shared<ov::op::v0::Convert>(result, ov::element::f32);
 }

+// If for_gather_matmul is true, the weight tensor may be N-D (e.g. 3D MoE expert weights
+// [n_expert, rows, cols]). The dequantization chain (Convert->[Subtract]->Multiply) is built as
+// usual but left in f16 (no final Convert to f32) -- ov::pass::MarkDequantization (registered in
+// translate_session.cpp) marks the chain so it survives model-build-time ConstantFolding -- see
+// make_int8_weights.cpp/make_int4_weights.cpp. mul_mat_id.cpp constructs ov::op::internal::GatherMatmul
+// directly from the resulting f16 dequant chain.
+//
+// When use_bias is true (explicitly, or implicitly because for_gather_matmul is true), the zp
+// tensor is expected to hold an exact f16 bias value (rather than a rounded integer zero point);
+// it is converted in place into an exact zero_point = -bias/scale and consumed via Subtract, not
+// Add, so the chain still matches OpenVINO's Convert->Subtract->Multiply decompression pattern.
 // See make_int8_weights for the meaning of for_gather_matmul.
 ov::Output<ov::Node> make_int4_weights(ov::Tensor & weight,
                                       ov::Tensor & scales,
                                       ov::Tensor & zp,
-                                       size_t group_size,
-                                       bool use_bias,
-                                       bool for_gather_matmul) {
+                                       size_t group_size = GGML_QUANTIZATION_GROUP_SIZE,
+                                       bool use_bias = false,
+                                       bool for_gather_matmul = false) {
    ov::Shape orig_weight_shape = weight.get_shape();
    bool is_signed = (weight.get_element_type() == ov::element::i4);  // Symmetric: signed weights, no ZP

@@ -746,357 +766,6 @@ ov::Output<ov::Node> make_mxfp4_moe_packed_weights(ov::Tensor & weight) {
    return weights_node;
 }

-// Extract quantized weights from tensor and create weight subgraph
-std::shared_ptr<ov::Node> extract_quantized_weights(const ggml_tensor * tensor,
-                                                    const void * data,
-                                                    ov::Tensor & weights,
-                                                    ov::Tensor & scales,
-                                                    ov::Tensor & zp,
-                                                    bool use_bias) {
-    // Create a temporary tensor for extraction functions that read from tensor->data
-    ggml_tensor temp_tensor = *tensor;
-    temp_tensor.data = const_cast<void *>(data);
-
-    if (tensor->type == GGML_TYPE_MXFP4) {
-        extract_mxfp4_data(&temp_tensor, weights, scales);
-        auto result = make_mxfp4_weights(weights, scales).get_node_shared_ptr();
-        result->set_friendly_name(tensor->name);
-        return result;
-    }
-
-    // Determine block size based on tensor type
-    int64_t weights_per_block;
-    bool is_u4;
-    switch (tensor->type) {
-    case GGML_TYPE_Q4_0:
-    case GGML_TYPE_Q4_1:
-    case GGML_TYPE_Q4_K:
-        is_u4 = true;
-        weights_per_block = 32;
-        break;
-    case GGML_TYPE_Q8_0:
-    case GGML_TYPE_Q5_1:
-    case GGML_TYPE_Q5_K:
-        is_u4 = false;
-        weights_per_block = 32;
-        break;
-    case GGML_TYPE_Q6_K:
-        is_u4 = false;
-        weights_per_block = 16;
-        break;
-    default:
-        throw std::runtime_error("Unsupported quantized type for extraction: " +
-                                 std::string(ggml_type_name(tensor->type)));
-    }
-
-    // 3D MoE expert weights (for_gather_matmul) always use the exact f16 zero-point extraction
-    // (see make_int8_weights/make_int4_weights) rather than the rounded integer zero point --
-    // round(min/scale) error is what corrupts Q4_K/Q5_1 experts, and the f16-zp form still fuses
-    // into GatherMatmulCompressed since it stays a Subtract, not an Add.
-    const bool for_gather_matmul = tensor->ne[2] > 1;
-    use_bias = use_bias || for_gather_matmul;
-
-    // Extract quantized data
-    switch (tensor->type) {
-    case GGML_TYPE_Q4_0:
-        extract_q4_0_data(&temp_tensor, weights, scales, zp);
-        break;
-    case GGML_TYPE_Q4_1:
-        extract_q4_1_data(&temp_tensor, weights, scales, zp, use_bias);
-        break;
-    case GGML_TYPE_Q4_K:
-        extract_q4_k_data(&temp_tensor, weights, scales, zp, use_bias);
-        break;
-    case GGML_TYPE_Q5_1:
-        extract_q5_1_data(&temp_tensor, weights, scales, zp, use_bias);
-        break;
-    case GGML_TYPE_Q8_0:
-        extract_q8_0_data(&temp_tensor, weights, scales, zp);
-        break;
-    case GGML_TYPE_Q6_K:
-        extract_q6_k_data(&temp_tensor, weights, scales, zp);
-        break;
-    case GGML_TYPE_Q5_K:
-        extract_q5_k_data(&temp_tensor, weights, scales, zp, use_bias);
-        break;
-    default:
-        throw std::runtime_error("Unsupported quantized type: " + std::string(ggml_type_name(tensor->type)));
-    }
-
-    // Create the OpenVINO weight subgraph. 3D expert weights (MoE) are routed through the
-    // GatherMatmul-oriented path: dequantized in f16, with constant folding disabled on the chain.
-    ov::Output<ov::Node> weight_node;
-    if (is_u4) {
-        weight_node = make_int4_weights(weights, scales, zp, weights_per_block, use_bias, for_gather_matmul);
-    } else {
-        weight_node = make_int8_weights(weights, scales, zp, weights_per_block, use_bias, for_gather_matmul);
-    }
-
-    auto result = weight_node.get_node_shared_ptr();
-    result->set_friendly_name(tensor->name);
-    return result;
-}
-
-// Requantize weights to target format, writing to provided buffers
-std::shared_ptr<ov::Node> requantize_to_buffers(const ggml_tensor * tensor,
-                                                const void * data,
-                                                ExtraQuantType requant_type,
-                                                int64_t block_size,
-                                                ov::Tensor & weights,
-                                                ov::Tensor & scales,
-                                                ov::Tensor & zp) {
-    int64_t n_elements = ggml_nelements(tensor);
-    const int64_t ne0 = tensor->ne[0];                 // elements per row
-    const int64_t n_rows = n_elements / ne0;
-    const auto * type_traits = ggml_get_type_traits(tensor->type);
-    const size_t src_row_bytes = ggml_row_size(tensor->type, ne0);
-
-    bool is_u4 = (requant_type == ExtraQuantType::Q4_0_C || requant_type == ExtraQuantType::Q4_0_128 ||
-                  requant_type == ExtraQuantType::Q4_0_64 || requant_type == ExtraQuantType::Q4_1_64);
-
-    // Streaming dequant (opt-in via GGML_OPENVINO_REDUCE_COMPILE_MEM or
-    // GGML_OPENVINO_MEMORY_OPTIMIZE): instead of
-    // materializing the full n_elements F32 array (e.g. ~1 GB for token_embd), dequantize
-    // a chunk of complete rows into a small scratch and quantize/convert it straight into
-    // the output buffers, capping the transient F32 footprint at CHUNK_ROWS*ne0 floats.
-    //
-    // Only valid (and only used) for the Q8_0_C / Q8_1_C / F16 targets whose block size
-    // divides a row (channel-wise _C uses block_size == ne0) so no target block straddles
-    // a row boundary, and Q8/F16 have no cross-block packing. The u4 (Q4_0) path packs two
-    // weights per byte with running zp ORs that assume a single whole-array call, so it is
-    // never streamed. When the flag is off, behavior is identical to the original
-    // full-materialization path.
-    const bool stream_requant = ggml_openvino_reduce_compile_mem_enabled() && !is_u4 &&
-                                !(block_size > 0 && ne0 % block_size != 0);
-
-    if (!stream_requant) {
-        // Full materialization (original behavior): dequantize the whole tensor to F32,
-        // then convert/quantize in one call.
-        std::vector<float> weights_f32(n_elements);
-        type_traits->to_float(data, weights_f32.data(), n_elements);
-        if (requant_type == ExtraQuantType::F16) {
-            ggml_get_type_traits(GGML_TYPE_F16)->from_float_ref(weights_f32.data(), weights.data(), n_elements);
-            auto result = std::make_shared<ov::op::v0::Constant>(weights);
-            result->set_friendly_name(tensor->name);
-            return result;
-        }
-        if (requant_type == ExtraQuantType::Q4_1_64) {
-            quantize_q4_1_asym(weights_f32.data(), weights, scales, zp, n_elements, block_size);
-        } else if (is_u4) {
-            quantize_q4_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
-        } else if (requant_type == ExtraQuantType::Q8_1_C) {
-            quantize_q8_1(weights_f32.data(), weights, scales, zp, n_elements, block_size);
-        } else {
-            quantize_q8_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
-        }
-    } else {
-        // Streaming path for Q8_0_C / Q8_1_C / F16 (covers token_embd, output.weight,
-        // and per-layer Q6_K/Q5_K requant — the large transient cases).
-        const int64_t CHUNK_ROWS = std::min<int64_t>(n_rows, 256);
-        std::vector<float> scratch(CHUNK_ROWS * ne0);
-        // F16 destination: 2 bytes/element, advanced per chunk by r0*ne0 elements.
-        auto * f16_base = static_cast<uint8_t *>(weights.data());
-        for (int64_t r0 = 0; r0 < n_rows; r0 += CHUNK_ROWS) {
-            const int64_t rows = std::min(CHUNK_ROWS, n_rows - r0);
-            const int64_t elems = rows * ne0;
-            const auto * src = static_cast<const uint8_t *>(data) + r0 * src_row_bytes;
-            type_traits->to_float(src, scratch.data(), elems);
-
-            if (requant_type == ExtraQuantType::F16) {
-                ggml_get_type_traits(GGML_TYPE_F16)
-                    ->from_float_ref(scratch.data(), f16_base + (r0 * ne0) * sizeof(uint16_t), elems);
-            } else {
-                const int64_t block_offset = (r0 * ne0) / block_size;
-                if (requant_type == ExtraQuantType::Q8_1_C) {
-                    quantize_q8_1(scratch.data(), weights, scales, zp, elems, block_size, block_offset);
-                } else {
-                    quantize_q8_0(scratch.data(), weights, scales, zp, elems, block_size, block_offset);
-                }
-            }
-        }
-        if (requant_type == ExtraQuantType::F16) {
-            auto result = std::make_shared<ov::op::v0::Constant>(weights);
-            result->set_friendly_name(tensor->name);
-            return result;
-        }
-    }
-
-    // Create the OpenVINO weight subgraph
-    ov::Output<ov::Node> weight_node;
-    if (is_u4) {
-        weight_node = make_int4_weights(weights, scales, zp, block_size);
-    } else {
-        weight_node = make_int8_weights(weights, scales, zp, block_size);
-    }
-
-    auto result = weight_node.get_node_shared_ptr();
-    result->set_friendly_name(tensor->name);
-    return result;
-}
-
-OvWeight process_weight_tensor(const ggml_tensor * tensor, const void * data, void * output_base_ptr, bool use_bias) {
-    GGML_ASSERT(tensor != nullptr);
-    GGML_ASSERT(data != nullptr);
-
-    OvWeight result;
-
-    // Get shape for weights: [rows, cols], or [n_expert, rows, cols] for 3D MoE expert weights.
-    ov::Shape node_shape = (tensor->ne[2] > 1) ?
-                               ov::Shape{static_cast<size_t>(tensor->ne[2]), static_cast<size_t>(tensor->ne[1]),
-                                         static_cast<size_t>(tensor->ne[0])} :
-                               ov::Shape{static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
-
-    // Handle F16/F32/BF16 weights
-    if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
-        ov::element::Type element_type;
-        switch (tensor->type) {
-        case GGML_TYPE_F32:
-            element_type = ov::element::f32;
-            break;
-        case GGML_TYPE_F16:
-            element_type = ov::element::f16;
-            break;
-        case GGML_TYPE_BF16:
-            element_type = ov::element::bf16;
-            break;
-        default:
-            OPENVINO_THROW("Unexpected tensor type in F16/F32/BF16 path");
-        }
-
-        if (output_base_ptr && output_base_ptr != data) {
-            // Using external buffer - copy data and create shared-memory constant
-            size_t tensor_bytes = ggml_nbytes(tensor);
-            memcpy(output_base_ptr, data, tensor_bytes);
-            result.weights = ov::Tensor(element_type, node_shape, output_base_ptr);
-        } else {
-            result.weights = ov::Tensor(element_type, node_shape, data);
-        }
-        result.weight_node = std::make_shared<ov::op::v0::Constant>(result.weights);
-        return result;
-    }
-
-    // Handle quantized weights
-    if (!ggml_is_quantized(tensor->type)) {
-        OPENVINO_THROW("Unsupported weight tensor type: ", ggml_type_name(tensor->type));
-    }
-
-    result.layout = ggml_openvino_get_extracted_layout(tensor, use_bias);
-    const auto & layout = result.layout;
-    if (layout.total_size == 0) {
-        OPENVINO_THROW("Unsupported quantized type: ", ggml_type_name(tensor->type));
-    }
-
-    // 3D MoE expert weights (for_gather_matmul) always use the exact f16 zero-point path (see
-    // extract_quantized_weights) -- must be kept in sync with the "use_bias || for_gather_matmul"
-    // check in ggml_openvino_get_extracted_layout, which sizes/offsets the zp slot accordingly.
-    // Requantized tensors (layout.is_requant) are handled by requantize_to_buffers instead, whose
-    // zp sizing/type is unaffected by for_gather_matmul, so they are excluded here.
-    const bool for_gather_matmul = tensor->ne[2] > 1;
-    const bool zp_is_f16 = !layout.is_requant && (use_bias || for_gather_matmul);
-
-    const bool is_3d_mxfp4_moe = tensor->type == GGML_TYPE_MXFP4 && (tensor->ne[2] > 1 || tensor->ne[3] > 1);
-    if (is_3d_mxfp4_moe) {
-        ov::Shape packed_shape = {static_cast<size_t>(tensor->ne[3]),
-                                  static_cast<size_t>(tensor->ne[2]),
-                                  static_cast<size_t>(tensor->ne[1]),
-                                  static_cast<size_t>(tensor->ne[0] / MXFP4_BLOCK_SIZE),
-                                  MXFP4_BLOCK_BYTES};
-        const size_t tensor_bytes = ggml_nbytes(tensor);
-        if (output_base_ptr) {
-            auto * buf_base = static_cast<uint8_t *>(output_base_ptr);
-            memcpy(buf_base + layout.weights_offset, data, tensor_bytes);
-            result.weights = ov::Tensor(ov::element::u8, packed_shape, buf_base + layout.weights_offset);
-        } else {
-            result.weights = ov::Tensor(ov::element::u8, packed_shape);
-            memcpy(result.weights.data(), data, tensor_bytes);
-        }
-        result.weight_node = make_mxfp4_moe_packed_weights(result.weights).get_node_shared_ptr();
-        result.weight_node->set_friendly_name(tensor->name);
-        return result;
-    }
-
-    if (use_bias) {
-        OPENVINO_ASSERT(!layout.is_requant,
-                        "use_bias is only used for test-backend-ops, which should not have requantization");
-        // bias node will be created on the fly and not use backend buffer
-        output_base_ptr = nullptr;
-    }
-
-    // F16 requant path - no separate scales/zp needed in result
-    if (layout.is_requant && layout.requant_type.has_value() && layout.requant_type.value() == ExtraQuantType::F16) {
-        if (output_base_ptr) {
-            result.weights = ov::Tensor(ov::element::f16, node_shape,
-                                        static_cast<uint8_t *>(output_base_ptr) + layout.weights_offset);
-        } else {
-            result.weights = ov::Tensor(ov::element::f16, node_shape);
-        }
-        ov::Tensor dummy_scales, dummy_zp;  // Not used for F16
-        result.weight_node =
-            requantize_to_buffers(tensor, data, ExtraQuantType::F16, 0, result.weights, dummy_scales, dummy_zp);
-        return result;
-    }
-
-    // Quantized path (normal extraction or quantized requant)
-    // Create weight/scale/zp tensors - shared between both paths
-    // For symmetric quantization, use signed types (i4/i8) and no ZP tensor
-    ov::element::Type weight_type = tensor->type == GGML_TYPE_MXFP4 ?
-                                        ov::element::f4e2m1 :
-                                        (layout.is_symmetric ? (layout.is_u4 ? ov::element::i4 : ov::element::i8) :
-                                                               (layout.is_u4 ? ov::element::u4 : ov::element::u8));
-    ov::Shape scale_shape = node_shape;
-    scale_shape.back() /= layout.weights_per_block;
-
-    if (tensor->type == GGML_TYPE_MXFP4) {
-        if (tensor->ne[2] == 1 && tensor->ne[3] == 1) {
-            node_shape = {static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
-        } else {
-            node_shape.clear();
-            for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
-                node_shape.push_back(static_cast<size_t>(tensor->ne[i]));
-            }
-        }
-
-        scale_shape = node_shape;
-        scale_shape.back() /= layout.weights_per_block;
-    }
-
-    if (output_base_ptr) {
-        uint8_t * buf_base = static_cast<uint8_t *>(output_base_ptr);
-        result.weights = ov::Tensor(weight_type, node_shape, buf_base + layout.weights_offset);
-        const ov::element::Type scale_type = tensor->type == GGML_TYPE_MXFP4 ? ov::element::f8e8m0 : ov::element::f16;
-        result.scales = ov::Tensor(scale_type, scale_shape, buf_base + layout.scales_offset);
-        if (!layout.is_symmetric) {
-            ov::element::Type zp_type =
-                zp_is_f16 ? ov::element::f16 : (layout.is_u4 ? ov::element::u4 : ov::element::u8);
-            result.zp = ov::Tensor(zp_type, scale_shape, buf_base + layout.zp_offset);
-        }
-        // else: result.zp remains default-constructed (empty) for symmetric
-    } else {
-        result.weights = ov::Tensor(weight_type, node_shape);
-        const ov::element::Type scale_type = tensor->type == GGML_TYPE_MXFP4 ? ov::element::f8e8m0 : ov::element::f16;
-        result.scales = ov::Tensor(scale_type, scale_shape);
-        if (!layout.is_symmetric) {
-            if (zp_is_f16) {
-                result.zp = ov::Tensor(ov::element::f16, scale_shape);
-            } else {
-                ov::element::Type zp_type = layout.is_u4 ? ov::element::u4 : ov::element::u8;
-                result.zp = ov::Tensor(zp_type, scale_shape);
-            }
-        }
-        // else: result.zp remains default-constructed (empty) for symmetric
-    }
-
-    if (layout.is_requant && layout.requant_type.has_value()) {
-        result.weight_node = requantize_to_buffers(tensor, data, layout.requant_type.value(), layout.weights_per_block,
-                                                   result.weights, result.scales, result.zp);
-    } else {
-        result.weight_node =
-            extract_quantized_weights(tensor, data, result.weights, result.scales, result.zp, use_bias);
-    }
-
-    return result;
-}
-
 void quantize_q4_0(const float * x,
                   ov::Tensor & weights_arr,
                   ov::Tensor & scales_arr,
@@ -1252,7 +921,7 @@ void quantize_q8_0(const float * x,
                   ov::Tensor & zp_arr,
                   int64_t k,
                   int64_t qk,
-                   int64_t block_offset) {
+                   int64_t block_offset = 0) {
    assert(k % qk == 0);
    const int nb = k / qk;

@@ -1308,7 +977,7 @@ void quantize_q8_1(const float * x,
                   ov::Tensor & zp_arr,
                   int64_t k,
                   int64_t qk,
-                   int64_t block_offset) {
+                   int64_t block_offset = 0) {
    assert(k % qk == 0);
    const int nb = k / qk;

@@ -1339,3 +1008,366 @@ void quantize_q8_1(const float * x,
        }
    }
 }
+
+// Extract quantized weights from tensor and create weight subgraph
+// If weights/scales/zp are provided (non-empty), uses them as output buffers
+// Otherwise allocates new ov::Tensors internally
+// Returns the weight node (make_int4_weights or make_int8_weights result)
+std::shared_ptr<ov::Node> extract_quantized_weights(const ggml_tensor * tensor,
+                                                    const void * data,  // Source data pointer (may differ from tensor->data)
+                                                    ov::Tensor & weights,
+                                                    ov::Tensor & scales,
+                                                    ov::Tensor & zp,
+                                                    // Use an exact f16 zero point (vs. a rounded integer one); always
+                                                    // used for for_gather_matmul (3D MoE expert) weights regardless of
+                                                    // this flag, and also settable explicitly for test-backend-ops.
+                                                    bool use_bias = false) {
+    // Create a temporary tensor for extraction functions that read from tensor->data
+    ggml_tensor temp_tensor = *tensor;
+    temp_tensor.data = const_cast<void *>(data);
+
+    if (tensor->type == GGML_TYPE_MXFP4) {
+        extract_mxfp4_data(&temp_tensor, weights, scales);
+        auto result = make_mxfp4_weights(weights, scales).get_node_shared_ptr();
+        result->set_friendly_name(tensor->name);
+        return result;
+    }
+
+    // Determine block size based on tensor type
+    int64_t weights_per_block;
+    bool is_u4;
+    switch (tensor->type) {
+    case GGML_TYPE_Q4_0:
+    case GGML_TYPE_Q4_1:
+    case GGML_TYPE_Q4_K:
+        is_u4 = true;
+        weights_per_block = 32;
+        break;
+    case GGML_TYPE_Q8_0:
+    case GGML_TYPE_Q5_1:
+    case GGML_TYPE_Q5_K:
+        is_u4 = false;
+        weights_per_block = 32;
+        break;
+    case GGML_TYPE_Q6_K:
+        is_u4 = false;
+        weights_per_block = 16;
+        break;
+    default:
+        throw std::runtime_error("Unsupported quantized type for extraction: " +
+                                 std::string(ggml_type_name(tensor->type)));
+    }
+
+    // 3D MoE expert weights (for_gather_matmul) always use the exact f16 zero-point extraction
+    // (see make_int8_weights/make_int4_weights) rather than the rounded integer zero point --
+    // round(min/scale) error is what corrupts Q4_K/Q5_1 experts, and the f16-zp form still fuses
+    // into GatherMatmulCompressed since it stays a Subtract, not an Add.
+    const bool for_gather_matmul = tensor->ne[2] > 1;
+    use_bias = use_bias || for_gather_matmul;
+
+    // Extract quantized data
+    switch (tensor->type) {
+    case GGML_TYPE_Q4_0:
+        extract_q4_0_data(&temp_tensor, weights, scales, zp);
+        break;
+    case GGML_TYPE_Q4_1:
+        extract_q4_1_data(&temp_tensor, weights, scales, zp, use_bias);
+        break;
+    case GGML_TYPE_Q4_K:
+        extract_q4_k_data(&temp_tensor, weights, scales, zp, use_bias);
+        break;
+    case GGML_TYPE_Q5_1:
+        extract_q5_1_data(&temp_tensor, weights, scales, zp, use_bias);
+        break;
+    case GGML_TYPE_Q8_0:
+        extract_q8_0_data(&temp_tensor, weights, scales, zp);
+        break;
+    case GGML_TYPE_Q6_K:
+        extract_q6_k_data(&temp_tensor, weights, scales, zp);
+        break;
+    case GGML_TYPE_Q5_K:
+        extract_q5_k_data(&temp_tensor, weights, scales, zp, use_bias);
+        break;
+    default:
+        throw std::runtime_error("Unsupported quantized type: " + std::string(ggml_type_name(tensor->type)));
+    }
+
+    // Create the OpenVINO weight subgraph. 3D expert weights (MoE) are routed through the
+    // GatherMatmul-oriented path: dequantized in f16, with constant folding disabled on the chain.
+    ov::Output<ov::Node> weight_node;
+    if (is_u4) {
+        weight_node = make_int4_weights(weights, scales, zp, weights_per_block, use_bias, for_gather_matmul);
+    } else {
+        weight_node = make_int8_weights(weights, scales, zp, weights_per_block, use_bias, for_gather_matmul);
+    }
+
+    auto result = weight_node.get_node_shared_ptr();
+    result->set_friendly_name(tensor->name);
+    return result;
+}
+
+// Requantize weights from tensor to target format, writing to provided buffers
+// For F16 target, only weights buffer is used (scales/zp ignored)
+// Returns the weight node
+std::shared_ptr<ov::Node> requantize_to_buffers(const ggml_tensor * tensor,
+                                                const void * data,  // Source data pointer
+                                                ExtraQuantType requant_type,
+                                                int64_t block_size,
+                                                ov::Tensor & weights,
+                                                ov::Tensor & scales,
+                                                ov::Tensor & zp) {
+    int64_t n_elements = ggml_nelements(tensor);
+    const int64_t ne0 = tensor->ne[0];                 // elements per row
+    const int64_t n_rows = n_elements / ne0;
+    const auto * type_traits = ggml_get_type_traits(tensor->type);
+    const size_t src_row_bytes = ggml_row_size(tensor->type, ne0);
+
+    bool is_u4 = (requant_type == ExtraQuantType::Q4_0_C || requant_type == ExtraQuantType::Q4_0_128 ||
+                  requant_type == ExtraQuantType::Q4_0_64 || requant_type == ExtraQuantType::Q4_1_64);
+
+    // Streaming dequant (opt-in via GGML_OPENVINO_REDUCE_COMPILE_MEM or
+    // GGML_OPENVINO_MEMORY_OPTIMIZE): instead of
+    // materializing the full n_elements F32 array (e.g. ~1 GB for token_embd), dequantize
+    // a chunk of complete rows into a small scratch and quantize/convert it straight into
+    // the output buffers, capping the transient F32 footprint at CHUNK_ROWS*ne0 floats.
+    //
+    // Only valid (and only used) for the Q8_0_C / Q8_1_C / F16 targets whose block size
+    // divides a row (channel-wise _C uses block_size == ne0) so no target block straddles
+    // a row boundary, and Q8/F16 have no cross-block packing. The u4 (Q4_0) path packs two
+    // weights per byte with running zp ORs that assume a single whole-array call, so it is
+    // never streamed. When the flag is off, behavior is identical to the original
+    // full-materialization path.
+    const bool stream_requant = ggml_openvino_reduce_compile_mem_enabled() && !is_u4 &&
+                                !(block_size > 0 && ne0 % block_size != 0);
+
+    if (!stream_requant) {
+        // Full materialization (original behavior): dequantize the whole tensor to F32,
+        // then convert/quantize in one call.
+        std::vector<float> weights_f32(n_elements);
+        type_traits->to_float(data, weights_f32.data(), n_elements);
+        if (requant_type == ExtraQuantType::F16) {
+            ggml_get_type_traits(GGML_TYPE_F16)->from_float_ref(weights_f32.data(), weights.data(), n_elements);
+            auto result = std::make_shared<ov::op::v0::Constant>(weights);
+            result->set_friendly_name(tensor->name);
+            return result;
+        }
+        if (requant_type == ExtraQuantType::Q4_1_64) {
+            quantize_q4_1_asym(weights_f32.data(), weights, scales, zp, n_elements, block_size);
+        } else if (is_u4) {
+            quantize_q4_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
+        } else if (requant_type == ExtraQuantType::Q8_1_C) {
+            quantize_q8_1(weights_f32.data(), weights, scales, zp, n_elements, block_size);
+        } else {
+            quantize_q8_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
+        }
+    } else {
+        // Streaming path for Q8_0_C / Q8_1_C / F16 (covers token_embd, output.weight,
+        // and per-layer Q6_K/Q5_K requant — the large transient cases).
+        const int64_t CHUNK_ROWS = std::min<int64_t>(n_rows, 256);
+        std::vector<float> scratch(CHUNK_ROWS * ne0);
+        // F16 destination: 2 bytes/element, advanced per chunk by r0*ne0 elements.
+        auto * f16_base = static_cast<uint8_t *>(weights.data());
+        for (int64_t r0 = 0; r0 < n_rows; r0 += CHUNK_ROWS) {
+            const int64_t rows = std::min(CHUNK_ROWS, n_rows - r0);
+            const int64_t elems = rows * ne0;
+            const auto * src = static_cast<const uint8_t *>(data) + r0 * src_row_bytes;
+            type_traits->to_float(src, scratch.data(), elems);
+
+            if (requant_type == ExtraQuantType::F16) {
+                ggml_get_type_traits(GGML_TYPE_F16)
+                    ->from_float_ref(scratch.data(), f16_base + (r0 * ne0) * sizeof(uint16_t), elems);
+            } else {
+                const int64_t block_offset = (r0 * ne0) / block_size;
+                if (requant_type == ExtraQuantType::Q8_1_C) {
+                    quantize_q8_1(scratch.data(), weights, scales, zp, elems, block_size, block_offset);
+                } else {
+                    quantize_q8_0(scratch.data(), weights, scales, zp, elems, block_size, block_offset);
+                }
+            }
+        }
+        if (requant_type == ExtraQuantType::F16) {
+            auto result = std::make_shared<ov::op::v0::Constant>(weights);
+            result->set_friendly_name(tensor->name);
+            return result;
+        }
+    }
+
+    // Create the OpenVINO weight subgraph
+    ov::Output<ov::Node> weight_node;
+    if (is_u4) {
+        weight_node = make_int4_weights(weights, scales, zp, block_size);
+    } else {
+        weight_node = make_int8_weights(weights, scales, zp, block_size);
+    }
+
+    auto result = weight_node.get_node_shared_ptr();
+    result->set_friendly_name(tensor->name);
+    return result;
+}
+}  // namespace
+
+OvWeight process_weight_tensor(const ggml_tensor * tensor, const void * data, void * output_base_ptr, bool use_bias) {
+    GGML_ASSERT(tensor != nullptr);
+    GGML_ASSERT(data != nullptr);
+
+    OvWeight result;
+
+    // Get shape for weights: [rows, cols], or [n_expert, rows, cols] for 3D MoE expert weights.
+    ov::Shape node_shape = (tensor->ne[2] > 1) ?
+                               ov::Shape{static_cast<size_t>(tensor->ne[2]), static_cast<size_t>(tensor->ne[1]),
+                                         static_cast<size_t>(tensor->ne[0])} :
+                               ov::Shape{static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
+
+    // Handle F16/F32/BF16 weights
+    if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
+        ov::element::Type element_type;
+        switch (tensor->type) {
+        case GGML_TYPE_F32:
+            element_type = ov::element::f32;
+            break;
+        case GGML_TYPE_F16:
+            element_type = ov::element::f16;
+            break;
+        case GGML_TYPE_BF16:
+            element_type = ov::element::bf16;
+            break;
+        default:
+            OPENVINO_THROW("Unexpected tensor type in F16/F32/BF16 path");
+        }
+
+        if (output_base_ptr && output_base_ptr != data) {
+            // Using external buffer - copy data and create shared-memory constant
+            size_t tensor_bytes = ggml_nbytes(tensor);
+            memcpy(output_base_ptr, data, tensor_bytes);
+            result.weights = ov::Tensor(element_type, node_shape, output_base_ptr);
+        } else {
+            result.weights = ov::Tensor(element_type, node_shape, data);
+        }
+        result.weight_node = std::make_shared<ov::op::v0::Constant>(result.weights);
+        return result;
+    }
+
+    // Handle quantized weights
+    if (!ggml_is_quantized(tensor->type)) {
+        OPENVINO_THROW("Unsupported weight tensor type: ", ggml_type_name(tensor->type));
+    }
+
+    result.layout = ggml_openvino_get_extracted_layout(tensor, use_bias);
+    const auto & layout = result.layout;
+    if (layout.total_size == 0) {
+        OPENVINO_THROW("Unsupported quantized type: ", ggml_type_name(tensor->type));
+    }
+
+    // 3D MoE expert weights (for_gather_matmul) always use the exact f16 zero-point path (see
+    // extract_quantized_weights) -- must be kept in sync with the "use_bias || for_gather_matmul"
+    // check in ggml_openvino_get_extracted_layout, which sizes/offsets the zp slot accordingly.
+    // Requantized tensors (layout.is_requant) are handled by requantize_to_buffers instead, whose
+    // zp sizing/type is unaffected by for_gather_matmul, so they are excluded here.
+    const bool for_gather_matmul = tensor->ne[2] > 1;
+    const bool zp_is_f16 = !layout.is_requant && (use_bias || for_gather_matmul);
+
+    const bool is_3d_mxfp4_moe = tensor->type == GGML_TYPE_MXFP4 && (tensor->ne[2] > 1 || tensor->ne[3] > 1);
+    if (is_3d_mxfp4_moe) {
+        ov::Shape packed_shape = {static_cast<size_t>(tensor->ne[3]),
+                                  static_cast<size_t>(tensor->ne[2]),
+                                  static_cast<size_t>(tensor->ne[1]),
+                                  static_cast<size_t>(tensor->ne[0] / MXFP4_BLOCK_SIZE),
+                                  MXFP4_BLOCK_BYTES};
+        const size_t tensor_bytes = ggml_nbytes(tensor);
+        if (output_base_ptr) {
+            auto * buf_base = static_cast<uint8_t *>(output_base_ptr);
+            memcpy(buf_base + layout.weights_offset, data, tensor_bytes);
+            result.weights = ov::Tensor(ov::element::u8, packed_shape, buf_base + layout.weights_offset);
+        } else {
+            result.weights = ov::Tensor(ov::element::u8, packed_shape);
+            memcpy(result.weights.data(), data, tensor_bytes);
+        }
+        result.weight_node = make_mxfp4_moe_packed_weights(result.weights).get_node_shared_ptr();
+        result.weight_node->set_friendly_name(tensor->name);
+        return result;
+    }
+
+    if (use_bias) {
+        OPENVINO_ASSERT(!layout.is_requant,
+                        "use_bias is only used for test-backend-ops, which should not have requantization");
+        // bias node will be created on the fly and not use backend buffer
+        output_base_ptr = nullptr;
+    }
+
+    // F16 requant path - no separate scales/zp needed in result
+    if (layout.is_requant && layout.requant_type.has_value() && layout.requant_type.value() == ExtraQuantType::F16) {
+        if (output_base_ptr) {
+            result.weights = ov::Tensor(ov::element::f16, node_shape,
+                                        static_cast<uint8_t *>(output_base_ptr) + layout.weights_offset);
+        } else {
+            result.weights = ov::Tensor(ov::element::f16, node_shape);
+        }
+        // Not used for F16:
+        ov::Tensor dummy_scales;
+        ov::Tensor dummy_zp;
+        result.weight_node =
+            requantize_to_buffers(tensor, data, ExtraQuantType::F16, 0, result.weights, dummy_scales, dummy_zp);
+        return result;
+    }
+
+    // Quantized path (normal extraction or quantized requant)
+    // Create weight/scale/zp tensors - shared between both paths
+    // For symmetric quantization, use signed types (i4/i8) and no ZP tensor
+    ov::element::Type weight_type;
+    if (tensor->type == GGML_TYPE_MXFP4) {
+        weight_type = ov::element::f4e2m1;
+    } else if (layout.is_symmetric) {
+        weight_type = layout.is_u4 ? ov::element::i4 : ov::element::i8;
+    } else {
+        weight_type = layout.is_u4 ? ov::element::u4 : ov::element::u8;
+    }
+    ov::Shape scale_shape = node_shape;
+    scale_shape.back() /= layout.weights_per_block;
+
+    if (tensor->type == GGML_TYPE_MXFP4) {
+        if (tensor->ne[2] == 1 && tensor->ne[3] == 1) {
+            node_shape = {static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
+        } else {
+            node_shape.clear();
+            for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
+                node_shape.push_back(static_cast<size_t>(tensor->ne[i]));
+            }
+        }
+
+        scale_shape = node_shape;
+        scale_shape.back() /= layout.weights_per_block;
+    }
+
+    const ov::element::Type scale_type = tensor->type == GGML_TYPE_MXFP4 ? ov::element::f8e8m0 : ov::element::f16;
+    ov::element::Type zp_type = layout.is_u4 ? ov::element::u4 : ov::element::u8;
+    if (zp_is_f16) {
+        zp_type = ov::element::f16;
+    }
+
+    if (output_base_ptr) {
+        uint8_t * buf_base = static_cast<uint8_t *>(output_base_ptr);
+        result.weights = ov::Tensor(weight_type, node_shape, buf_base + layout.weights_offset);
+        result.scales = ov::Tensor(scale_type, scale_shape, buf_base + layout.scales_offset);
+        if (!layout.is_symmetric) {
+            result.zp = ov::Tensor(zp_type, scale_shape, buf_base + layout.zp_offset);
+        }
+        // else: result.zp remains default-constructed (empty) for symmetric
+    } else {
+        result.weights = ov::Tensor(weight_type, node_shape);
+        result.scales = ov::Tensor(scale_type, scale_shape);
+        if (!layout.is_symmetric) {
+            result.zp = ov::Tensor(zp_type, scale_shape);
+        }
+        // else: result.zp remains default-constructed (empty) for symmetric
+    }
+
+    if (layout.is_requant && layout.requant_type.has_value()) {
+        result.weight_node = requantize_to_buffers(tensor, data, layout.requant_type.value(), layout.weights_per_block,
+                                                   result.weights, result.scales, result.zp);
+    } else {
+        result.weight_node =
+            extract_quantized_weights(tensor, data, result.weights, result.scales, result.zp, use_bias);
+    }
+
+    return result;
+}
@@ -2,112 +2,12 @@
 #include "ggml-openvino-extra.h"  // For ExtraQuantType
 #include "ggml.h"

-#include <cstdint>
-#include <openvino/op/constant.hpp>
 #include <openvino/core/node_output.hpp>
+#include <openvino/op/constant.hpp>
 #include <openvino/runtime/tensor.hpp>

-void unpack_32_4(const uint8_t * data, uint8_t * dst);
-
-void extract_q4_0_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr);
-
-void extract_q4_1_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr,
-                       bool use_bias = false);
-
-void extract_q5_1_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr,
-                       bool use_bias = false);
-
-void extract_q8_0_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr);
-
-void unpack_256_4(const uint8_t * data, uint8_t * dst);
-
-void extract_q4_k_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr,
-                       bool use_bias = false);
-
-void extract_q5_k_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr,
-                       bool use_bias = false);
-
-void extract_q6_k_data(const ggml_tensor * tensor,
-                       ov::Tensor & weights_arr,
-                       ov::Tensor & scales_arr,
-                       ov::Tensor & zp_arr);
-
-void extract_mxfp4_data(const ggml_tensor * tensor, ov::Tensor & weights_arr, ov::Tensor & scales_arr);
-
 static constexpr size_t GGML_QUANTIZATION_GROUP_SIZE = 32;

-// If for_gather_matmul is true, the weight tensor may be N-D (e.g. 3D MoE expert weights
-// [n_expert, rows, cols]). The dequantization chain (Convert->[Subtract]->Multiply) is built as
-// usual but left in f16 (no final Convert to f32) -- ov::pass::MarkDequantization (registered in
-// translate_session.cpp) marks the chain so it survives model-build-time ConstantFolding -- see
-// make_int8_weights.cpp/make_int4_weights.cpp. mul_mat_id.cpp constructs ov::op::internal::GatherMatmul
-// directly from the resulting f16 dequant chain.
-//
-// When use_bias is true (explicitly, or implicitly because for_gather_matmul is true), the zp
-// tensor is expected to hold an exact f16 bias value (rather than a rounded integer zero point);
-// it is converted in place into an exact zero_point = -bias/scale and consumed via Subtract, not
-// Add, so the chain still matches OpenVINO's Convert->Subtract->Multiply decompression pattern.
-ov::Output<ov::Node> make_int8_weights(ov::Tensor & weight,
-                                       ov::Tensor & scales,
-                                       ov::Tensor & zp,
-                                       size_t group_size = GGML_QUANTIZATION_GROUP_SIZE,
-                                       bool use_bias = false,
-                                       bool for_gather_matmul = false);
-
-ov::Output<ov::Node> make_int4_weights(ov::Tensor & weight,
-                                       ov::Tensor & scales,
-                                       ov::Tensor & zp,
-                                       size_t group_size = GGML_QUANTIZATION_GROUP_SIZE,
-                                       bool use_bias = false,
-                                       bool for_gather_matmul = false);
-
-ov::Output<ov::Node> make_mxfp4_weights(ov::Tensor & weight, ov::Tensor & scales);
-
-ov::Output<ov::Node> make_mxfp4_moe_packed_weights(ov::Tensor & weight);
-
-// Extract quantized weights from tensor and create weight subgraph
-// If weights/scales/zp are provided (non-empty), uses them as output buffers
-// Otherwise allocates new ov::Tensors internally
-// Returns the weight node (make_int4_weights or make_int8_weights result)
-std::shared_ptr<ov::Node> extract_quantized_weights(
-    const ggml_tensor * tensor,
-    const void * data,  // Source data pointer (may differ from tensor->data)
-    ov::Tensor & weights,
-    ov::Tensor & scales,
-    ov::Tensor & zp,
-    bool use_bias = false);  // Use an exact f16 zero point (vs. a rounded integer one); always
-                             // used for for_gather_matmul (3D MoE expert) weights regardless of
-                             // this flag, and also settable explicitly for test-backend-ops.
-
-// Requantize weights from tensor to target format, writing to provided buffers
-// For F16 target, only weights buffer is used (scales/zp ignored)
-// Returns the weight node
-std::shared_ptr<ov::Node> requantize_to_buffers(const ggml_tensor * tensor,
-                                                const void * data,  // Source data pointer
-                                                ExtraQuantType requant_type,
-                                                int64_t block_size,
-                                                ov::Tensor & weights,
-                                                ov::Tensor & scales,
-                                                ov::Tensor & zp);
-
 inline const char * extra_quant_type_name(ExtraQuantType t) {
    switch (t) {
    case ExtraQuantType::F16:
@@ -156,41 +56,3 @@ OvWeight process_weight_tensor(
                                       // always used for for_gather_matmul (3D MoE expert) weights
                                       // regardless of this flag, and also settable explicitly for
                                       // test-backend-ops.
-
-void quantize_q4_0(const float * x,
-                   ov::Tensor & weights_arr,
-                   ov::Tensor & scales_arr,
-                   ov::Tensor & zp_arr,
-                   int64_t k,
-                   int64_t qk);
-void quantize_q8_1(const float * x,
-                   ov::Tensor & weights_arr,
-                   ov::Tensor & scales_arr,
-                   ov::Tensor & zp_arr,
-                   int64_t k,
-                   int64_t qk,
-                   int64_t block_offset = 0);
-void quantize_q4_1_asym(const float * x,
-                        ov::Tensor & weights_arr,
-                        ov::Tensor & scales_arr,
-                        ov::Tensor & zp_arr,
-                        int64_t k,
-                        int64_t qk);
-void quantize_q8_0(const float * x,
-                   ov::Tensor & weights_arr,
-                   ov::Tensor & scales_arr,
-                   ov::Tensor & zp_arr,
-                   int64_t k,
-                   int64_t qk,
-                   int64_t block_offset = 0);
-
-namespace ov {
-namespace op {
-namespace util {
-// From <openvino>/src/common/transformations/include/transformations/utils/utils.hpp
-bool get_single_value(const std::shared_ptr<ov::op::v0::Constant> & const_node,
-                      float & value,
-                      bool check_value_range = true);
-}  // namespace util
-}  // namespace op
-}  // namespace ov
@@ -237,7 +237,8 @@ bool ggml_openvino_model_cache_verify_manifest(const std::string & path,
    if (!f.is_open()) {
        return false;
    }
-    std::string tag, val;
+    std::string tag;
+    std::string val;
    // header: fingerprint
    if (!(f >> tag >> val) || tag != "fingerprint" || val != hex64(fingerprint)) {
        return false;
@@ -12,7 +12,6 @@ namespace ggml {

 class FrontEnd {
 public:
-    using Ptr = std::shared_ptr<FrontEnd>;
    FrontEnd();

    static std::shared_ptr<Model> convert(const InputModel::Ptr & model, bool naive = false);
@@ -20,7 +20,7 @@ namespace op {
 static ov::Output<ov::Node> reshape_add_id_input_to_2d(const ov::Output<ov::Node> & input,
                                                       const ov::PartialShape & input_shape,
                                                       const std::vector<int> & dims) {
-    const auto actual_shape = input.get_partial_shape();
+    const auto & actual_shape = input.get_partial_shape();
    if (actual_shape.rank().is_static() && actual_shape.rank().get_length() == 2) {
        return input;
    }
@@ -3,12 +3,9 @@
 #include "../op_table.h"
 #include "../utils.h"

-#include <climits>
-#include <cstdint>
 #include <memory>
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/slice.hpp>
-#include <vector>

 namespace ov {
 namespace frontend {
@@ -195,7 +195,9 @@ OutputVector translate_flash_attn_ext(const NodeContext & context) {
    auto tile_kv = [&](int64_t n_heads, int64_t n_heads_kv, int64_t hs, ov::Output<Node> kv) {
        int64_t f = n_heads / n_heads_kv;
        if (f > 1 && n_heads_kv > 1) {
-            ov::Output<ov::Node> kv_broadcast_shape, kv_unsqueezed, new_kv_shape;
+            ov::Output<ov::Node> kv_broadcast_shape;
+            ov::Output<ov::Node> kv_unsqueezed;
+            ov::Output<ov::Node> new_kv_shape;
            auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, Shape{}, {2});
            kv_unsqueezed = std::make_shared<ov::op::v0::Unsqueeze>(kv, unsqueeze_axes);

@@ -196,7 +196,7 @@ static OutputVector translate_gated_delta_net_ref(const NodeContext & context) {
    }

    // Merge batch and head dims: [B*H_v, T, S_v]
-    auto merge_bh = [&](ov::Output<ov::Node> x, int64_t last_dim) {
+    auto merge_bh = [&](const ov::Output<ov::Node> & x, int64_t last_dim) {
        auto shape = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{B * H_v, T, last_dim});
        return std::make_shared<ov::op::v1::Reshape>(x, shape, false);
    };
@@ -1,7 +1,6 @@
 #include "../node_context.h"
 #include "../op_table.h"
 #include "../utils.h"
-#include "ggml-impl.h"

 #include <cstddef>
 #include <memory>
@@ -42,7 +42,7 @@ ov::Output<ov::Node> slice_axis(const ov::Output<ov::Node> & input, int64_t axis

 ov::Output<ov::Node> static_shape_dims_or_shapeof(const ov::Output<ov::Node> & input,
                                                  const std::vector<int> & dims) {
-    const auto partial_shape = input.get_partial_shape();
+    const auto & partial_shape = input.get_partial_shape();
    if (partial_shape.is_static()) {
        std::vector<int64_t> values;
        values.reserve(dims.size());
@@ -8,6 +8,7 @@
 #include <openvino/op/pad.hpp>
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/shape_of.hpp>
+#include <utility>
 #include <vector>

 namespace ov {
@@ -20,7 +21,7 @@ namespace {
 ov::Output<ov::Node> translate_circular_pad(ov::Output<ov::Node> input,
                                            const std::array<int32_t, 8> & pads,
                                            const ov::Shape & input_shape) {
-    ov::Output<ov::Node> result = input;
+    ov::Output<ov::Node> result = std::move(input);

    const std::array<int32_t, 4> pads_begin = {pads[6], pads[4], pads[2], pads[0]};
    const std::array<int32_t, 4> pads_end = {pads[7], pads[5], pads[3], pads[1]};
@@ -1,7 +1,6 @@
 #include "../node_context.h"
 #include "../op_table.h"
 #include "../utils.h"
-#include "ggml.h"

 #include <memory>
 #include <openvino/op/broadcast.hpp>
@@ -25,9 +25,7 @@ OutputVector translate_rms_norm(const NodeContext & context) {
    auto op_case = context.get_op_case();

    ov::Output<ov::Node> input_node;
-    if (op_case == 1) {
-        input_node = process_view_input_new(context, 0);
-    } else if (op_case == 2) {
+    if (op_case == 2) {
        auto ssm_state_size = context.get_ssm_state_size();
        // The GDN op packs [attn | new_state] along the row axis; the state occupies the last
        // ssm_state_size * n_seqs rows. Slice it off (scaling by the active sequence count) to keep
@@ -7,7 +7,6 @@
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/shape_of.hpp>
 #include <openvino/op/slice.hpp>
-#include <set>

 namespace ov {
 namespace frontend {
@@ -153,7 +152,8 @@ OutputVector translate_view(const NodeContext & context) {
        return {input};
    }

-    int64_t src_elems = 1, dst_elems = 1;
+    int64_t src_elems = 1;
+    int64_t dst_elems = 1;
    for (int64_t i = 0; i < src_shape.rank().get_length(); ++i) {
        if (src_shape[i].is_dynamic()) {
            return {input};
@@ -84,7 +84,7 @@ bool KVStateSeqAxis::run_on_model(const std::shared_ptr<ov::Model> & model) {
        // Readers still expect seq at dim 1. A reader that is itself the inverse
        // Transpose wanted seq at dim 2 all along, so drop it; give anything else the
        // inverse Transpose so its input is unchanged.
-        for (auto & reader : readers) {
+        for (const auto & reader : readers) {
            auto * node = reader.get_node();
            if (ov::is_type<ov::op::v6::Assign>(node)) {
                continue;
@@ -344,7 +344,7 @@ std::shared_ptr<Model> TranslateSession::translate_graph(const frontend::InputMo
        }
    };

-    auto node_visitor = [&](std::shared_ptr<GgmlDecoder> decoder, int node_idx) {
+    auto node_visitor = [&](const std::shared_ptr<GgmlDecoder> & decoder, int node_idx) {
        auto converted_outputs = translate_node(decoder, node_idx);
        if (converted_outputs.empty()) {
            return;
@@ -1,7 +1,5 @@
 #include "utils.h"

-#include "ggml-impl.h"
-
 #include <cmath>
 #include <cstddef>
 #include <ctime>
@@ -28,13 +26,6 @@ namespace ov {
 namespace frontend {
 namespace ggml {

-std::string getCurrentTime() {
-    std::time_t now = std::time(nullptr);
-    char buf[100];
-    std::strftime(buf, sizeof(buf), "%Y-%m-%d %H:%M:%S", std::localtime(&now));
-    return buf;
-}
-
 void num_inputs_check(const NodeContext & context, size_t min_inputs, size_t max_inputs) {
    auto input_size = context.get_input_size();
    FRONT_END_OP_CONVERSION_CHECK(input_size >= min_inputs, "Got less inputs than expected");
@@ -82,7 +73,7 @@ namespace {
 ov::Output<ov::Node> rope_yarn_ramp_mix(int n_dims, const float corr_dims[2], float ext_factor) {
    int half_n_dims = n_dims / 2;
    std::vector<float> dim_ids_vec(half_n_dims);
-    std::iota(dim_ids_vec.begin(), dim_ids_vec.end(), 0);
+    std::iota(dim_ids_vec.begin(), dim_ids_vec.end(), 0.0f);
    auto dim_ids = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, (size_t) half_n_dims}, dim_ids_vec);
    auto corr_low = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {corr_dims[0]});
    auto corr_high = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {corr_dims[1]});
@@ -551,6 +542,7 @@ ov::Output<ov::Node> process_view_input_new(const NodeContext & context, int inp

                if (tail_begin >= 0 && tail_end <= tail_src_elems) {
                    std::vector<int64_t> flat_shape;
+                    flat_shape.reserve(slice_dim);
                    for (int i = 0; i < slice_dim; ++i) {
                        flat_shape.push_back(static_cast<int64_t>(view_src_ggml_shape[i]));
                    }
@@ -14,8 +14,6 @@ namespace ggml {

 std::string getCurrentTime();

-void dump_ov_model(std::shared_ptr<ov::Model> model);
-
 void num_inputs_check(const NodeContext & context, size_t min_inputs, size_t max_inputs);

 int non_cont_dim(std::vector<size_t> ne, std::vector<size_t> nb);
@@ -142,9 +142,6 @@ struct ov_runtime_context {

 enum ggml_status ov_graph_compute(struct ggml_cgraph * cgraph, ggml_backend_t backend);

-enum ggml_status ov_graph_compute_dynamic(struct ggml_cgraph * cgraph, std::shared_ptr<ov_runtime_context> r_ctx);
-enum ggml_status ov_graph_compute_static(struct ggml_cgraph * cgraph, std::shared_ptr<ov_runtime_context> r_ctx);
-
 size_t checksum(const void * data, size_t size);

 bool save_ggml_tensor_data_to_txt(const ggml_tensor * tensor, const std::string & file_path);
@@ -185,18 +182,6 @@ int64_t get_inp_pos_n_tokens(struct ggml_cgraph * cgraph, const ggml_tensor * in

 bool get_is_prefill(struct ggml_cgraph * cgraph, const ggml_tensor * inp_pos);

-ov::Tensor get_ov_input_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder, const std::string & param_name);
-ov::Tensor get_ov_input_tensor_static_decode(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
-                                             const std::string & param_name);
-ov::Tensor get_ov_input_tensor_static_prefill(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
-                                              const std::string & param_name,
-                                              int chunk_index);
-
-ov::Tensor create_ov_output_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
-                                   std::shared_ptr<ov::InferRequest> infer_request,
-                                   int output_index,
-                                   const ggml_tensor * ggml_tensor);
-
 bool is_naive(struct ggml_cgraph * cgraph);

 /**
@@ -205,9 +190,3 @@ bool is_naive(struct ggml_cgraph * cgraph);
 * @return true if the graph is identified as split; otherwise false.
 */
 bool is_model_splitted(struct ggml_cgraph * cgraph);
-
-enum ggml_status naive_compute(struct ggml_cgraph * cgraph,
-                               ov::Core & core,
-                               const std::string & device,
-                               const ov::AnyMap & config,
-                               ov_compiled_model_cache & cache);
@@ -208,5 +208,53 @@ bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializ
        return true;
    }

+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_SSM_CONV && ops.begin()[1] == GGML_OP_UNARY &&
+        unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {
+        const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];
+        const ggml_tensor * silu     = cgraph->nodes[node_idx + 1];
+
+        if (ggml_get_unary_op(silu) != unary_ops.begin()[0]) {
+            return false;
+        }
+        if (ssm_conv->type != GGML_TYPE_F32 || silu->type != GGML_TYPE_F32) {
+            return false;
+        }
+        // the fused kernel writes the SiLU output with dense strides, so it must be contiguous
+        if (!ggml_is_contiguous(silu)) {
+            return false;
+        }
+
+        return true;
+    }
+
+    if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SSM_CONV && ops.begin()[1] == GGML_OP_ADD &&
+        ops.begin()[2] == GGML_OP_UNARY && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {
+        const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];
+        const ggml_tensor * add      = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * silu     = cgraph->nodes[node_idx + 2];
+
+        if (ggml_get_unary_op(silu) != unary_ops.begin()[0]) {
+            return false;
+        }
+        if (ssm_conv->type != GGML_TYPE_F32 || add->type != GGML_TYPE_F32 || silu->type != GGML_TYPE_F32) {
+            return false;
+        }
+        // the fused kernel writes the SiLU output with dense strides, so it must be contiguous
+        if (!ggml_is_contiguous(silu)) {
+            return false;
+        }
+
+        // ADD must consume ssm_conv's output and broadcast a 1-D channel-wise bias
+        const ggml_tensor * bias = (add->src[0] == ssm_conv) ? add->src[1] : add->src[0];
+        if (bias->type != GGML_TYPE_F32 || !ggml_is_contiguous(bias)) {
+            return false;
+        }
+        if (ggml_nelements(bias) != ssm_conv->ne[0] || bias->ne[0] != ssm_conv->ne[0]) {
+            return false;
+        }
+
+        return true;
+    }
+
    return false;
 }
@@ -1011,12 +1011,25 @@ static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_ty
    GGML_UNUSED(buft);
 }

+bool is_bmg_g31_arch(int device) {
+    return ggml_sycl_info().devices[device].hw_info.arch == gpu_arch::intel_gpu_bmg_g31;
+}
+
 static size_t ggml_backend_sycl_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
    size_t max_alloc_size = dpct::get_current_device().get_max_mem_alloc_size();
    if (g_ggml_sycl_host_pinned_mem_2g) {
        return std::min(max_alloc_size, (size_t) 2LL*1024*1024*1024);
    } else {
+        ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
+        int device = ctx->device;
+        if(is_bmg_g31_arch(device)) {
+            //Todo, it's workaround for BMG-G31, which has a known issue with large allocations.
+            //The max alloc size is reduced to 60% of the reported max alloc size.
+            //remove it after https://github.com/intel/compute-runtime/issues/998 is fixed.
+            max_alloc_size = max_alloc_size*0.6;
+        }
        return max_alloc_size;
+
    }
    GGML_UNUSED(buft);
 }
@@ -1484,7 +1497,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface
    /* .is_host          = */ ggml_backend_sycl_split_buffer_type_is_host,
 };

-ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) {
+ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type([[maybe_unused]] int main_device, const float * tensor_split) {
    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_split_buffer_type\n");

    static std::mutex mutex;
@@ -3243,7 +3256,7 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor * ds
    SYCL_CHECK(0);
 }

-static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
+static void ggml_sycl_set_peer_access(const int n_tokens, [[maybe_unused]] int main_device) {
    static bool peer_access_enabled = false;

    const bool enable_peer_access = n_tokens <= GGML_SYCL_PEER_MAX_BATCH_SIZE;
@@ -6034,6 +6047,20 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
            }
        }

+        if (node->op == GGML_OP_SSM_CONV &&
+            ggml_sycl_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_ADD, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
+            ggml_sycl_ssm_conv_fused(*sycl_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2]);
+            i += 2;
+            continue;
+        }
+
+        if (node->op == GGML_OP_SSM_CONV &&
+            ggml_sycl_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
+            ggml_sycl_ssm_conv_fused(*sycl_ctx, node, nullptr, cgraph->nodes[i + 1]);
+            i++;
+            continue;
+        }
+
        if (node->op == GGML_OP_MUL_MAT && ggml_sycl_mul_mat_glu_mmvq_fused(*sycl_ctx, cgraph, i)) {
            i += 2;
            continue;
@@ -1,11 +1,71 @@
 #include "ssm_conv.hpp"
 #include "common.hpp"
+#include "element_wise.hpp"

 #include <cstdio>

 using namespace sycl;

-static void kernel_ssm_conv(
+// One output element of the conv. DC is d_conv as a compile-time constant (0 keeps the
+// runtime loop); unfused callers pass literal false/nullptr so the epilogue folds away.
+template <int DC>
+static __dpct_inline__ void ssm_conv_element(
+    size_t idx,
+    const float *src_data,
+    const float *weights,
+    float *dst_data,
+    int d_conv,
+    int d_inner,
+    int n_t,
+    int src_stride_inner,
+    int src_stride_seq,
+    int dst_stride_token,
+    int dst_stride_seq,
+    bool apply_silu,
+    const float *bias
+) {
+    // src is token-contiguous per channel, dst is channel-contiguous per token,
+    // so indexing token-fastest coalesces the d_conv loads.
+    const int token   = static_cast<int>(idx % n_t);
+    const int channel = static_cast<int>((idx / n_t) % d_inner);
+    const int seq     = static_cast<int>(idx / (static_cast<size_t>(n_t) * static_cast<size_t>(d_inner)));
+
+    const float *s = src_data
+        + static_cast<size_t>(seq) * static_cast<size_t>(src_stride_seq)
+        + static_cast<size_t>(channel) * static_cast<size_t>(src_stride_inner)
+        + static_cast<size_t>(token);
+
+    const float *c = weights + static_cast<size_t>(channel) * static_cast<size_t>(d_conv);
+
+    float sumf = 0.0f;
+    if constexpr (DC > 0) {
+#pragma unroll
+        for (int i0 = 0; i0 < DC; ++i0) {
+            sumf += s[i0] * c[i0];
+        }
+    } else {
+        for (int i0 = 0; i0 < d_conv; ++i0) {
+            sumf += s[i0] * c[i0];
+        }
+    }
+
+    // fused bias add: the ADD node broadcasts a 1-D channel bias over tokens
+    if (bias != nullptr) {
+        sumf += bias[channel];
+    }
+
+    const size_t dst_idx =
+        static_cast<size_t>(seq) * static_cast<size_t>(dst_stride_seq) +
+        static_cast<size_t>(token) * static_cast<size_t>(dst_stride_token) +
+        static_cast<size_t>(channel);
+
+    dst_data[dst_idx] = apply_silu ? op_silu(sumf) : sumf;
+}
+
+// FUSED=false keeps apply_silu/bias out of the kernel capture list, so the unfused launch
+// takes the pre-fusion argument list; matters at n_t == 1, where the op is launch-bound.
+template <int DC, bool FUSED>
+static void kernel_ssm_conv_impl(
    queue &q,
    const float *src_data,
    const float *weights,
@@ -18,7 +78,9 @@ static void kernel_ssm_conv(
    int src_stride_inner,
    int src_stride_seq,
    int dst_stride_token,
-    int dst_stride_seq
+    int dst_stride_seq,
+    bool apply_silu,
+    const float *bias
 ) {
    const size_t total_work = static_cast<size_t>(d_inner) * static_cast<size_t>(n_t) * static_cast<size_t>(n_s);
    const size_t work_group_size = 256;
@@ -27,53 +89,199 @@ static void kernel_ssm_conv(
    const range<1> global_range(num_work_groups * work_group_size);
    const range<1> local_range(work_group_size);

-    q.submit([&](handler &h) {
-        h.parallel_for(
-            nd_range<1>(global_range, local_range),
-            [=](nd_item<1> item) {
-                const size_t idx = item.get_global_id(0);
-                if (idx >= total_work) {
-                    return;
+    if constexpr (FUSED) {
+        q.submit([&](handler &h) {
+            h.parallel_for(
+                nd_range<1>(global_range, local_range),
+                [=](nd_item<1> item) {
+                    const size_t idx = item.get_global_id(0);
+                    if (idx >= total_work) {
+                        return;
+                    }
+
+                    ssm_conv_element<DC>(idx, src_data, weights, dst_data, d_conv, d_inner, n_t,
+                                         src_stride_inner, src_stride_seq, dst_stride_token,
+                                         dst_stride_seq, apply_silu, bias);
                }
+            );
+        });
+    } else {
+        GGML_UNUSED(apply_silu);
+        GGML_UNUSED(bias);

-                // src has the tokens of one channel contiguous, dst has the channels of one
-                // token contiguous, so either the loads or the store must be strided. Indexing
-                // token-fastest coalesces the d_conv loads, which measured faster except for
-                // short, cache-resident rows.
-                const int token   = static_cast<int>(idx % n_t);
-                const int channel = static_cast<int>((idx / n_t) % d_inner);
-                const int seq     = static_cast<int>(idx / (static_cast<size_t>(n_t) * static_cast<size_t>(d_inner)));
+        q.submit([&](handler &h) {
+            h.parallel_for(
+                nd_range<1>(global_range, local_range),
+                [=](nd_item<1> item) {
+                    const size_t idx = item.get_global_id(0);
+                    if (idx >= total_work) {
+                        return;
+                    }

-                const float *s = src_data
-                    + static_cast<size_t>(seq) * static_cast<size_t>(src_stride_seq)
-                    + static_cast<size_t>(channel) * static_cast<size_t>(src_stride_inner)
-                    + static_cast<size_t>(token);
-
-                const float *c = weights + static_cast<size_t>(channel) * static_cast<size_t>(d_conv);
-
-                float sumf = 0.0f;
-                for (int i0 = 0; i0 < d_conv; ++i0) {
-                    sumf += s[i0] * c[i0];
+                    ssm_conv_element<DC>(idx, src_data, weights, dst_data, d_conv, d_inner, n_t,
+                                         src_stride_inner, src_stride_seq, dst_stride_token,
+                                         dst_stride_seq, false, nullptr);
                }
-
-                const size_t dst_idx =
-                    static_cast<size_t>(seq) * static_cast<size_t>(dst_stride_seq) +
-                    static_cast<size_t>(token) * static_cast<size_t>(dst_stride_token) +
-                    static_cast<size_t>(channel);
-
-                dst_data[dst_idx] = sumf;
-            }
-        );
-    });
+            );
+        });
+    }
 }

-inline void ggml_sycl_op_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+// SLM transpose tile: coalesces both the loads and the stores. The +1 pad makes the row
+// stride 33, coprime with 32 banks, so both phases are bank-conflict-free.
+template <int DC, int TT, int TC, int WG>
+static __dpct_inline__ void ssm_conv_tile(
+    nd_item<1> it, local_accessor<float, 1> tile, const float *src_data, const float *weights,
+    float *dst_data, int n_t, int nt_tiles, int nc_tiles, int src_stride_inner,
+    int src_stride_seq, int dst_stride_token, int dst_stride_seq, bool apply_silu,
+    const float *bias
+) {
+    const int    lid = static_cast<int>(it.get_local_id(0));
+    const size_t g   = it.get_group(0);
+    const int    tt  = static_cast<int>(g % nt_tiles);
+    const int    ct  = static_cast<int>((g / nt_tiles) % nc_tiles);
+    const int    seq = static_cast<int>(g / (static_cast<size_t>(nt_tiles) * nc_tiles));
+    const int    t0 = tt * TT, c0 = ct * TC;
+
+    const int ti = lid % TT;
+    const int cj = lid / TT;
+#pragma unroll
+    for (int r = 0; r < TC / (WG / TT); ++r) {
+        const int c   = cj + r * (WG / TT);
+        const int tok = t0 + ti;
+        float sumf = 0.0f;
+        if (tok < n_t) {
+            const float *s = src_data + static_cast<size_t>(seq) * src_stride_seq
+                           + static_cast<size_t>(c0 + c) * src_stride_inner + tok;
+            const float *cw = weights + static_cast<size_t>(c0 + c) * DC;
+#pragma unroll
+            for (int i = 0; i < DC; ++i) sumf += s[i] * cw[i];
+            if (bias != nullptr) sumf += bias[c0 + c];
+            if (apply_silu) sumf = op_silu(sumf);
+        }
+        tile[c * (TT + 1) + ti] = sumf;
+    }
+    it.barrier(access::fence_space::local_space);
+
+    const int cc = lid % TC;
+    const int tj = lid / TC;
+#pragma unroll
+    for (int r = 0; r < TT / (WG / TC); ++r) {
+        const int t   = tj + r * (WG / TC);
+        const int tok = t0 + t;
+        if (tok < n_t) {
+            dst_data[static_cast<size_t>(seq) * dst_stride_seq
+                     + static_cast<size_t>(tok) * dst_stride_token + c0 + cc]
+                = tile[cc * (TT + 1) + t];
+        }
+    }
+}
+
+// Same FUSED split as kernel_ssm_conv_impl. The fused instantiation keeps the runtime
+// apply_silu/bias branches: at n_t >= 32 they are amortized over the whole tile.
+template <int DC, bool FUSED>
+static void kernel_ssm_conv_tiled(
+    queue &q, const float *src_data, const float *weights, float *dst_data,
+    int d_inner, int n_t, int n_s, int src_stride_inner, int src_stride_seq,
+    int dst_stride_token, int dst_stride_seq, bool apply_silu, const float *bias
+) {
+    constexpr int TT = 32, TC = 32, WG = 256;
+    const int nt_tiles = (n_t + TT - 1) / TT;
+    const int nc_tiles = d_inner / TC;
+    const size_t groups = static_cast<size_t>(nt_tiles) * nc_tiles * n_s;
+
+    if constexpr (FUSED) {
+        q.submit([&](handler &h) {
+            local_accessor<float, 1> tile(range<1>(TC * (TT + 1)), h);
+            h.parallel_for(nd_range<1>(range<1>(groups * WG), range<1>(WG)), [=](nd_item<1> it) {
+                ssm_conv_tile<DC, TT, TC, WG>(it, tile, src_data, weights, dst_data, n_t, nt_tiles,
+                                              nc_tiles, src_stride_inner, src_stride_seq,
+                                              dst_stride_token, dst_stride_seq, apply_silu, bias);
+            });
+        });
+    } else {
+        GGML_UNUSED(apply_silu);
+        GGML_UNUSED(bias);
+
+        q.submit([&](handler &h) {
+            local_accessor<float, 1> tile(range<1>(TC * (TT + 1)), h);
+            h.parallel_for(nd_range<1>(range<1>(groups * WG), range<1>(WG)), [=](nd_item<1> it) {
+                ssm_conv_tile<DC, TT, TC, WG>(it, tile, src_data, weights, dst_data, n_t, nt_tiles,
+                                              nc_tiles, src_stride_inner, src_stride_seq,
+                                              dst_stride_token, dst_stride_seq, false, nullptr);
+            });
+        });
+    }
+}
+
+static void kernel_ssm_conv(
+    queue &q,
+    const float *src_data,
+    const float *weights,
+    float *dst_data,
+    int d_conv,
+    int d_inner,
+    int n_t,
+    int n_s,
+    int ncs,
+    int src_stride_inner,
+    int src_stride_seq,
+    int dst_stride_token,
+    int dst_stride_seq,
+    bool apply_silu,
+    const float *bias
+) {
+    // Only the fused instantiations carry apply_silu/bias as kernel arguments; the plain
+    // ssm_conv launch keeps the argument list it had before the fusion landed.
+    const bool fused = apply_silu || bias != nullptr;
+
+    // d_inner must be a multiple of 32 so the channel tiles are exact; the transpose is only
+    // worth it for n_t >= 32. d_conv == 4 is the only window with a DC-specialized kernel.
+    if (d_conv == 4 && n_t >= 32 && (d_inner % 32) == 0) {
+        if (fused) {
+            kernel_ssm_conv_tiled<4, true>(q, src_data, weights, dst_data, d_inner, n_t, n_s,
+                                           src_stride_inner, src_stride_seq, dst_stride_token,
+                                           dst_stride_seq, apply_silu, bias);
+        } else {
+            kernel_ssm_conv_tiled<4, false>(q, src_data, weights, dst_data, d_inner, n_t, n_s,
+                                            src_stride_inner, src_stride_seq, dst_stride_token,
+                                            dst_stride_seq, apply_silu, bias);
+        }
+        return;
+    }
+
+    if (d_conv == 4) {
+        if (fused) {
+            kernel_ssm_conv_impl<4, true>(q, src_data, weights, dst_data, d_conv, d_inner, n_t, n_s,
+                                          ncs, src_stride_inner, src_stride_seq, dst_stride_token,
+                                          dst_stride_seq, apply_silu, bias);
+        } else {
+            kernel_ssm_conv_impl<4, false>(q, src_data, weights, dst_data, d_conv, d_inner, n_t, n_s,
+                                           ncs, src_stride_inner, src_stride_seq, dst_stride_token,
+                                           dst_stride_seq, apply_silu, bias);
+        }
+        return;
+    }
+
+    if (fused) {
+        kernel_ssm_conv_impl<0, true>(q, src_data, weights, dst_data, d_conv, d_inner, n_t, n_s,
+                                      ncs, src_stride_inner, src_stride_seq, dst_stride_token,
+                                      dst_stride_seq, apply_silu, bias);
+    } else {
+        kernel_ssm_conv_impl<0, false>(q, src_data, weights, dst_data, d_conv, d_inner, n_t, n_s,
+                                       ncs, src_stride_inner, src_stride_seq, dst_stride_token,
+                                       dst_stride_seq, apply_silu, bias);
+    }
+}
+
+inline void ggml_sycl_op_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst = nullptr, const float * bias = nullptr) {
    ggml_tensor * src0 = dst->src[0];
    ggml_tensor * src1 = dst->src[1];

    GGML_ASSERT(src0->type == GGML_TYPE_F32);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);
    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(bias == nullptr || silu_dst != nullptr);

    const int d_conv   = src1->ne[0];
    const int ncs      = src0->ne[0];
@@ -104,7 +312,8 @@ inline void ggml_sycl_op_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor *

        const float *src_data = static_cast<const float *>(src0->data);
        const float *weights  = static_cast<const float *>(src1->data);
-        float *dst_data       = static_cast<float *>(dst->data);
+        const bool apply_silu = silu_dst != nullptr;
+        float *dst_data       = static_cast<float *>((silu_dst ? silu_dst : dst)->data);

        GGML_ASSERT(src_data && weights && dst_data);

@@ -121,7 +330,9 @@ inline void ggml_sycl_op_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor *
            src_stride_inner,
            src_stride_seq,
            dst_stride_token,
-            dst_stride_seq
+            dst_stride_seq,
+            apply_silu,
+            bias
        );

    } catch (const std::exception &e) {
@@ -134,3 +345,17 @@ void ggml_sycl_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
    ggml_sycl_op_ssm_conv(ctx, dst);
 }
+
+// Fused ssm_conv + ADD + SiLU: write silu(conv(x) + b) straight into silu_dst, eliding the
+// standalone SiLU launch and its HBM round-trip of the conv output.
+void ggml_sycl_ssm_conv_fused(ggml_backend_sycl_context & ctx, ggml_tensor * dst, ggml_tensor * add, ggml_tensor * silu_dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    GGML_ASSERT(silu_dst && ggml_are_same_shape(dst, silu_dst) && silu_dst->type == GGML_TYPE_F32);
+    // the fused kernel reads only the ADD's bias operand; the ADD result is never written
+    const float * bias = nullptr;
+    if (add != nullptr) {
+        const ggml_tensor * bias_t = (add->src[0] == dst) ? add->src[1] : add->src[0];
+        bias = static_cast<const float *>(bias_t->data);
+    }
+    ggml_sycl_op_ssm_conv(ctx, dst, silu_dst, bias);
+}
@@ -3,3 +3,4 @@
 #include "common.hpp"

 void ggml_sycl_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_ssm_conv_fused(ggml_backend_sycl_context & ctx, ggml_tensor * dst, ggml_tensor * add, ggml_tensor * silu_dst);
@@ -62,6 +62,11 @@ if (Vulkan_FOUND)
    ggml_add_backend_library(ggml-vulkan
                             ggml-vulkan.cpp
                             ../../include/ggml-vulkan.h
+                             ggml-vulkan-types.h
+                             ggml-vulkan-push-constants.h
+                             ggml-vulkan-common.h
+                             ggml-vulkan-buffers.cpp
+                             ggml-vulkan-debug.cpp
                            )

    set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
@@ -0,0 +1,783 @@
+#include "ggml-vulkan-common.h"
+
+ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_vk_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_vk_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_vk_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_vk_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+static std::vector<uint32_t> ggml_vk_find_memory_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
+    std::vector<uint32_t> indices;
+
+    for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
+        vk::MemoryType memory_type = mem_props->memoryTypes[i];
+        if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
+            (flags & memory_type.propertyFlags) == flags &&
+            mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
+            indices.push_back(i);
+        }
+    }
+    return indices;
+}
+
+static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std::initializer_list<vk::MemoryPropertyFlags> & req_flags_list,
+                                       void *import_ptr = nullptr) {
+    VK_LOG_DEBUG("ggml_vk_create_buffer(" << device->name << ", " << size << ", " << to_string(req_flags_list.begin()[0]) << ", " << to_string(req_flags_list.begin()[req_flags_list.size()-1]) << ")");
+    if (size > device->max_buffer_size) {
+        throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device buffer size limit");
+    }
+
+    vk_buffer buf = std::make_shared<vk_buffer_struct>();
+
+    if (size == 0) {
+        buf->size = 0;
+        return buf;
+    }
+
+    vk::BufferUsageFlags usage_flags = vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc | vk::BufferUsageFlagBits::eTransferDst;
+    vk::MemoryAllocateFlags mem_flags {};
+    if (device->buffer_device_address) {
+        usage_flags |= vk::BufferUsageFlagBits::eShaderDeviceAddress;
+        mem_flags |= vk::MemoryAllocateFlagBits::eDeviceAddress;
+    }
+
+    vk::BufferCreateInfo buffer_create_info{
+        vk::BufferCreateFlags(),
+        size,
+        usage_flags,
+        vk::SharingMode::eExclusive,
+        0,
+        nullptr,
+    };
+
+    vk::ExternalMemoryBufferCreateInfo external_memory_bci;
+    if (import_ptr) {
+        external_memory_bci.handleTypes = vk::ExternalMemoryHandleTypeFlagBits::eHostAllocationEXT;
+        buffer_create_info.setPNext(&external_memory_bci);
+    }
+
+    buf->buffer = device->device.createBuffer(buffer_create_info);
+
+    vk::MemoryRequirements mem_req = device->device.getBufferMemoryRequirements(buf->buffer);
+
+    vk::PhysicalDeviceMemoryProperties mem_props = device->physical_device.getMemoryProperties();
+
+    const vk::MemoryPriorityAllocateInfoEXT mem_priority_info { 1.0f };
+
+    vk::MemoryAllocateFlagsInfo mem_flags_info { mem_flags };
+
+    if (device->memory_priority) {
+        mem_flags_info.setPNext(&mem_priority_info);
+    }
+
+    if (import_ptr) {
+        vk::MemoryHostPointerPropertiesEXT host_pointer_props;
+        try {
+            host_pointer_props = device->device.getMemoryHostPointerPropertiesEXT(vk::ExternalMemoryHandleTypeFlagBits::eHostAllocationEXT, import_ptr);
+        } catch (vk::SystemError& e) {
+            GGML_LOG_WARN("ggml_vulkan: Failed getMemoryHostPointerPropertiesEXT (%s)\n", e.what());
+            device->device.destroyBuffer(buf->buffer);
+            return {};
+        }
+        vk::PhysicalDeviceMemoryProperties mem_props = device->physical_device.getMemoryProperties();
+
+        uint32_t memory_type_idx;
+        vk::MemoryPropertyFlags property_flags = *req_flags_list.begin();
+        for (memory_type_idx = 0; memory_type_idx < 32; ++memory_type_idx) {
+            if (!(host_pointer_props.memoryTypeBits & (1u << memory_type_idx))) {
+                continue;
+            }
+            if (!(mem_req.memoryTypeBits & (1u << memory_type_idx))) {
+                continue;
+            }
+
+            vk::MemoryType memory_type = mem_props.memoryTypes[memory_type_idx];
+            // check for visible+coherent+cached. Other flags (e.g. devicelocal) are allowed
+            if ((memory_type.propertyFlags & property_flags) == property_flags) {
+                property_flags = memory_type.propertyFlags;
+                break;
+            }
+        }
+        if (memory_type_idx == 32) {
+            GGML_LOG_WARN("ggml_vulkan: Memory type for host allocation not found\n");
+            device->device.destroyBuffer(buf->buffer);
+            return {};
+        }
+
+        buf->memory_property_flags = mem_props.memoryTypes[memory_type_idx].propertyFlags;
+        try {
+            vk::ImportMemoryHostPointerInfoEXT import_info;
+            import_info.handleType = vk::ExternalMemoryHandleTypeFlagBits::eHostAllocationEXT;
+            import_info.pHostPointer = import_ptr;
+            import_info.setPNext(&mem_flags_info);
+            buf->device_memory = device->device.allocateMemory({ size, memory_type_idx, &import_info });
+        } catch (const vk::SystemError& e) {
+        }
+    } else {
+        for (auto it = req_flags_list.begin(); it != req_flags_list.end(); it++) {
+            const auto & req_flags = *it;
+
+            const std::vector<uint32_t> memory_type_indices = ggml_vk_find_memory_properties(&mem_props, &mem_req, req_flags);
+
+            if (memory_type_indices.empty()) {
+                continue;
+            }
+
+            bool done = false;
+
+            for (auto mtype_it = memory_type_indices.begin(); mtype_it != memory_type_indices.end(); mtype_it++) {
+                try {
+                    buf->device_memory = device->device.allocateMemory({ mem_req.size, *mtype_it, &mem_flags_info });
+                    buf->memory_property_flags = mem_props.memoryTypes[*mtype_it].propertyFlags;
+                    done = true;
+                    break;
+                } catch (const vk::SystemError& e) {
+                    // loop and retry
+                    // during last attempt throw the exception
+                    if (it + 1 == req_flags_list.end() && mtype_it + 1 == memory_type_indices.end()) {
+                        device->device.destroyBuffer(buf->buffer);
+                        throw e;
+                    }
+                }
+            }
+
+            if (done) {
+                break;
+            }
+        }
+    }
+
+    if (!buf->device_memory) {
+        device->device.destroyBuffer(buf->buffer);
+        throw vk::OutOfDeviceMemoryError("No suitable memory type found");
+    }
+
+    buf->ptr = nullptr;
+
+    if (import_ptr) {
+        buf->ptr = import_ptr;
+    } else {
+        if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+            buf->ptr = device->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE);
+        }
+    }
+
+    device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0);
+
+    buf->device = device;
+    buf->size = size;
+
+    if (device->buffer_device_address) {
+        const vk::BufferDeviceAddressInfo addressInfo(buf->buffer);
+        buf->bda_addr = device->device.getBufferAddress(addressInfo);
+    }
+
+    device->memory_logger->log_allocation(buf, size);
+
+    return buf;
+}
+
+vk_buffer ggml_vk_create_buffer_check(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags) {
+    try {
+        return ggml_vk_create_buffer(device, size, {req_flags, fallback_flags});
+    } catch (const vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        throw e;
+    }
+}
+
+vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
+    vk_buffer buf;
+    try {
+        if (device->prefer_host_memory) {
+            buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
+                                                       vk::MemoryPropertyFlagBits::eDeviceLocal});
+        } else if (device->uma) {
+            // On UMA, prefer host-visible memory so direct tensor borrowing works.
+            // If unavailable, fall back to device-local memory.
+            buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
+                                                       vk::MemoryPropertyFlagBits::eDeviceLocal,
+                                                       vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
+        } else if (device->disable_host_visible_vidmem) {
+            if (device->allow_sysmem_fallback) {
+                buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal,
+                                                           vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
+            } else {
+                buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal});
+            }
+        } else {
+            // use rebar if available, otherwise fallback to device only visible memory
+            if (device->allow_sysmem_fallback) {
+                buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
+                                                           vk::MemoryPropertyFlagBits::eDeviceLocal,
+                                                           vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
+            } else {
+                buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
+                                                           vk::MemoryPropertyFlagBits::eDeviceLocal});
+            }
+        }
+    } catch (const vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        throw e;
+    }
+
+    return buf;
+}
+
+void ggml_vk_destroy_buffer(vk_buffer& buf) {
+    if (buf == nullptr) {
+        return;
+    }
+
+    if (buf->device != nullptr) {
+        buf->device->memory_logger->log_deallocation(buf);
+    }
+
+    buf.reset();
+}
+
+void * ggml_vk_host_malloc(vk_device& device, size_t size) {
+    VK_LOG_MEMORY("ggml_vk_host_malloc(" << size << ")");
+    vk_buffer buf = ggml_vk_create_buffer(device, size,
+        {vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+         vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
+
+    if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
+        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
+            size/1024.0/1024.0);
+        device->device.freeMemory(buf->device_memory);
+        device->device.destroyBuffer(buf->buffer);
+        return nullptr;
+    }
+
+    std::lock_guard<std::shared_mutex> guard(device->pinned_memory_mutex);
+    device->pinned_memory.push_back(std::make_tuple(buf->ptr, size, buf));
+
+    return buf->ptr;
+}
+
+void ggml_vk_host_free(vk_device& device, void* ptr) {
+    if (ptr == nullptr) {
+        return;
+    }
+    VK_LOG_MEMORY("ggml_vk_host_free(" << ptr << ")");
+    std::lock_guard<std::shared_mutex> guard(device->pinned_memory_mutex);
+
+    vk_buffer buf;
+    size_t index;
+    for (size_t i = 0; i < device->pinned_memory.size(); i++) {
+        const uint8_t* addr = (const uint8_t*) std::get<0>(device->pinned_memory[i]);
+        const uint8_t* endr = addr + std::get<1>(device->pinned_memory[i]);
+        if (ptr >= addr && ptr < endr) {
+            buf = std::get<2>(device->pinned_memory[i]);
+            index = i;
+            break;
+        }
+    }
+    if (buf == nullptr) {
+        fprintf(stderr, "WARNING: failed to free pinned memory: memory not in map\n");
+        return;
+    }
+
+    ggml_vk_destroy_buffer(buf);
+
+    device->pinned_memory.erase(device->pinned_memory.begin() + index);
+}
+
+void ggml_vk_host_get(const vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
+    std::shared_lock<std::shared_mutex> guard(device->pinned_memory_mutex);
+    buf = nullptr;
+    buf_offset = 0;
+    for (size_t i = 0; i < device->pinned_memory.size(); i++) {
+        const uint8_t* addr = (const uint8_t*) std::get<0>(device->pinned_memory[i]);
+        const uint8_t* endr = addr + std::get<1>(device->pinned_memory[i]);
+        if (ptr >= addr && ptr < endr) {
+            buf = std::get<2>(device->pinned_memory[i]);
+            buf_offset = ((const uint8_t *)ptr) - addr;
+            break;
+        }
+    }
+}
+
+void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) {
+    if (device->sync_staging == nullptr || device->sync_staging->size < size) {
+        VK_LOG_MEMORY("ggml_vk_ensure_sync_staging_buffer(" << size << ")");
+        ggml_vk_destroy_buffer(device->sync_staging);
+        device->sync_staging = ggml_vk_create_buffer_check(device, size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+    }
+}
+
+void ggml_vk_ensure_sync_staging_buffer(ggml_backend_vk_context * ctx, size_t size) {
+    if (ctx->sync_staging == nullptr || ctx->sync_staging->size < size) {
+        VK_LOG_MEMORY("ggml_vk_ensure_sync_staging_buffer(" << size << ")");
+        ggml_vk_destroy_buffer(ctx->sync_staging);
+        ctx->sync_staging = ggml_vk_create_buffer_check(ctx->device, size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+    }
+}
+
+static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context& subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_nc_async(" << tensor << ")");
+    GGML_ASSERT(!ggml_is_contiguous(tensor));
+    // Buffer is already mapped
+    if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+        std::cerr << "ggml_vulkan: buffer_write_nc_async dst buffer is host_visible. Use synchronous write." << std::endl;
+        GGML_ABORT("fatal error");
+    }
+    // Check if src is pinned memory
+    vk_buffer buf = nullptr;
+    size_t buf_offset = 0;
+    ggml_vk_host_get(ctx->device, tensor->data, buf, buf_offset);
+
+    const uint64_t ne0 = tensor->ne[0];
+    const uint64_t ne1 = tensor->ne[1];
+    const uint64_t ne2 = tensor->ne[2];
+    const uint64_t ne3 = tensor->ne[3];
+    const uint64_t nb0 = tensor->nb[0];
+    const uint64_t nb1 = tensor->nb[1];
+    const uint64_t nb2 = tensor->nb[2];
+    const uint64_t nb3 = tensor->nb[3];
+    const ggml_type type = tensor->type;
+    const uint64_t ts = ggml_type_size(type);
+    const uint64_t bs = ggml_blck_size(type);
+
+    const uint64_t dstnb0 = ts;
+    const uint64_t dstnb1 = dstnb0*(ne0/bs);
+    const uint64_t dstnb2 = dstnb1*ne1;
+    const uint64_t dstnb3 = dstnb2*ne2;
+
+    const uint64_t ne = ggml_nelements(tensor);
+
+    if (buf != nullptr) {
+        // Memory is pinned, use as staging buffer
+        std::vector<vk::BufferCopy> slices;
+
+        for (uint64_t i3 = 0; i3 < ne3; i3++) {
+            for (uint64_t i2 = 0; i2 < ne2; i2++) {
+                // Find longest contiguous slice
+                if (ne1*nb1 == dstnb2) {
+                    slices.push_back({ buf_offset + i3*nb3 + i2*nb2, offset + i3*dstnb3 + i2*dstnb2, dstnb2 });
+                } else {
+                    for (uint64_t i1 = 0; i1 < ne1; i1++) {
+                        if (ne0*nb0/bs == dstnb1) {
+                            slices.push_back({ buf_offset + i3*nb3 + i2*nb2 + i1*nb1, offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, dstnb1 });
+                        } else {
+                            const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
+                            const uint64_t d_off = offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
+                            for (uint64_t i0 = 0; i0 < ne0; i0++) {
+                                slices.push_back({ s_off + i0*nb0, d_off + i0*dstnb0, dstnb0 });
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        ggml_vk_sync_buffers(ctx, subctx);
+        subctx->s->buffer->buf.copyBuffer(buf->buffer, dst->buffer, slices);
+        return;
+    }
+
+    if (!sync_staging) {
+        GGML_ABORT("Asynchronous write to non-pinned memory not supported");
+    }
+
+    // Staging buffer required
+    vk_buffer& staging = ctx->device->sync_staging;
+    const uint64_t copy_size = ts*ne/bs;
+    ggml_vk_ensure_sync_staging_buffer(ctx->device, copy_size);
+    VkBufferCopy buf_copy{ 0, offset, copy_size };
+
+    ggml_vk_sync_buffers(ctx, subctx);
+    vkCmdCopyBuffer(subctx->s->buffer->buf, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
+
+    for (uint64_t i3 = 0; i3 < ne3; i3++) {
+        for (uint64_t i2 = 0; i2 < ne2; i2++) {
+            // Find longest contiguous slice
+            if (ne1*nb1 == dstnb2) {
+                deferred_memcpy((uint8_t *)staging->ptr + i3*dstnb3 + i2*dstnb2, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2, dstnb2, &subctx->in_memcpys);
+            } else {
+                for (uint64_t i1 = 0; i1 < ne1; i1++) {
+                    if (ne0*nb0/bs == dstnb1) {
+                        deferred_memcpy((uint8_t *)staging->ptr + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2 + i1*nb1, dstnb1, &subctx->in_memcpys);
+                    } else {
+                        const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
+                        const uint64_t d_off = i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
+                        for (uint64_t i0 = 0; i0 < ne0; i0++) {
+                            deferred_memcpy((uint8_t *)staging->ptr + d_off + i0*dstnb0, (const uint8_t *) tensor->data + s_off + i0*nb0, dstnb0, &subctx->in_memcpys);
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+bool ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_2d_async(" << width << ", " << height << ")");
+    // Check if src is pinned memory
+    vk_buffer buf = nullptr;
+    size_t buf_offset = 0;
+    ggml_vk_host_get(dst->device, src, buf, buf_offset);
+
+    if (buf != nullptr) {
+        // Memory is pinned, use as staging buffer
+        std::vector<vk::BufferCopy> slices(1);
+        if (width == spitch && width == dpitch) {
+            // Only do single write if stride is equal
+            slices[0].srcOffset = buf_offset;
+            slices[0].dstOffset = offset;
+            slices[0].size = width * height;
+        } else {
+            slices.resize(height);
+            for (size_t i = 0; i < height; i++) {
+                slices[i].srcOffset = buf_offset + i * spitch;
+                slices[i].dstOffset = offset + i * dpitch;
+                slices[i].size = width;
+            }
+        }
+
+        ggml_vk_sync_buffers(nullptr, subctx);
+        subctx->s->buffer->buf.copyBuffer(buf->buffer, dst->buffer, slices);
+        return true;
+    }
+    VK_LOG_DEBUG("STAGING");
+
+    if (!sync_staging) {
+        // copy was not handled caller needs to fall back
+        return false;
+    }
+
+    // Staging buffer required
+    const size_t staging_size = width * height;
+    ggml_vk_ensure_sync_staging_buffer(dst->device, staging_size);
+
+    vk_buffer& staging_buffer = dst->device->sync_staging;
+
+    std::vector<vk::BufferCopy> slices(1);
+    if (width == dpitch) {
+        slices[0].srcOffset = 0;
+        slices[0].dstOffset = offset;
+        slices[0].size = staging_size;
+    } else {
+        slices.resize(height);
+        for (size_t i = 0; i < height; i++) {
+            slices[i].srcOffset = i * width;
+            slices[i].dstOffset = offset + i * dpitch;
+            slices[i].size = width;
+        }
+    }
+
+    ggml_vk_sync_buffers(nullptr, subctx);
+    subctx->s->buffer->buf.copyBuffer(staging_buffer->buffer, dst->buffer, slices);
+
+    if (width == spitch) {
+        deferred_memcpy((uint8_t *)staging_buffer->ptr, src, staging_size, &subctx->in_memcpys);
+    } else {
+        for (size_t i = 0; i < height; i++) {
+            deferred_memcpy((uint8_t *)staging_buffer->ptr + i * width, (const uint8_t *) src + i * spitch, width, &subctx->in_memcpys);
+        }
+    }
+    return true;
+}
+
+bool ggml_vk_buffer_write_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, bool sync_staging) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_async(" << size << ")");
+    return ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, size, size, size, 1, sync_staging);
+}
+
+void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t dpitch, size_t width, size_t height) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_2d(" << width << ", " << height << ")");
+    // Buffer is already mapped
+    if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+        GGML_ASSERT(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
+
+        if (width == spitch && width == dpitch) {
+            memcpy((uint8_t *)dst->ptr + offset, src, width * height);
+        } else {
+            for (size_t i = 0; i < height; i++) {
+                memcpy((uint8_t *)dst->ptr + offset + i * dpitch, (const uint8_t *) src + i * spitch, width);
+            }
+        }
+    } else {
+        std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
+
+        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue->cmd_pool);
+        ggml_vk_ctx_begin(dst->device, subctx);
+        bool ret = ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, dpitch, width, height, true);
+        GGML_ASSERT(ret);
+        ggml_vk_ctx_end(subctx);
+
+        for (auto& cpy : subctx->in_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+
+        for (auto& mset : subctx->memsets) {
+            memset(mset.dst, mset.val, mset.n);
+        }
+
+        ggml_vk_submit(subctx, dst->device->fence);
+        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences", dst->device);
+        dst->device->device.resetFences({ dst->device->fence });
+        ggml_vk_queue_command_pools_cleanup(dst->device);
+    }
+}
+
+void ggml_vk_buffer_write(vk_buffer& dst, size_t offset, const void * src, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write(" << size << ")");
+    ggml_vk_buffer_write_2d(dst, offset, src, size, size, size, 1);
+}
+
+bool ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging) {
+    VK_LOG_DEBUG("ggml_vk_buffer_read_2d_async(offset=" << offset << ", width=" << width << ", height=" << height << ")");
+    GGML_ASSERT(width > 0);
+    GGML_ASSERT(height > 0);
+    GGML_ASSERT(src != nullptr);
+
+    // TODO: staging_offset is not used
+
+    // Check if dst is pinned memory
+    vk_buffer buf = nullptr;
+    size_t buf_offset = 0;
+    ggml_vk_host_get(src->device, dst, buf, buf_offset);
+
+    std::vector<vk::BufferCopy> slices(1);
+    if (width == spitch && width == dpitch) {
+        // Only do single write if stride is equal
+        slices[0].srcOffset = offset;
+        slices[0].dstOffset = buf_offset;
+        slices[0].size = width * height;
+    } else {
+        slices.resize(height);
+        for (size_t i = 0; i < height; i++) {
+            slices[i].srcOffset = offset + i * spitch;
+            slices[i].dstOffset = buf_offset + i * dpitch;
+            slices[i].size = width;
+        }
+    }
+
+    if (buf != nullptr) {
+        // Memory is pinned, use as staging buffer
+        ggml_vk_sync_buffers(nullptr, subctx);
+        subctx->s->buffer->buf.copyBuffer(src->buffer, buf->buffer, slices);
+
+        return true;
+    }
+    VK_LOG_DEBUG("STAGING");
+
+    if (!sync_staging) {
+        // copy was not handled caller needs to fall back
+        return false;
+    }
+
+    // Fall back to staging buffer
+    const size_t staging_size = width * height;
+    ggml_vk_ensure_sync_staging_buffer(src->device, staging_size);
+
+    vk_buffer& staging_buffer = src->device->sync_staging;
+
+    std::vector<vk::BufferCopy> staging_slices(1);
+    if (width == spitch) {
+        staging_slices[0].srcOffset = offset;
+        staging_slices[0].dstOffset = 0;
+        staging_slices[0].size = staging_size;
+    } else {
+        staging_slices.resize(height);
+        for (size_t i = 0; i < height; i++) {
+            staging_slices[i].srcOffset = offset + i * spitch;
+            staging_slices[i].dstOffset = i * width;
+            staging_slices[i].size = width;
+        }
+    }
+
+    ggml_vk_sync_buffers(nullptr, subctx);
+    subctx->s->buffer->buf.copyBuffer(src->buffer, staging_buffer->buffer, staging_slices);
+
+    if (width == dpitch) {
+        deferred_memcpy(dst, staging_buffer->ptr, staging_size, &subctx->out_memcpys);
+    } else {
+        for (size_t i = 0; i < height; i++) {
+            deferred_memcpy((uint8_t *) dst + i * dpitch, (const uint8_t *) staging_buffer->ptr + i * width, width, &subctx->out_memcpys);
+        }
+    }
+    return true;
+}
+
+static bool ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, bool sync_staging = false) {
+    return ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, size, size, size, 1, sync_staging);
+}
+
+void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height) {
+    VK_LOG_DEBUG("ggml_vk_buffer_read_2d(" << src->buffer << ", " << offset << ", " << width << ", " << height << ")");
+
+    // If the device is not an UMA device the memory is host-accessible through rebar. While writing
+    // through PCIe is sufficient fast reading back data from PCIe is slower than going through
+    // the HW device to host copy path.
+    if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) {
+        GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
+
+        std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->compute_queue->cmd_pool);
+        ggml_vk_ctx_begin(src->device, subctx);
+        subctx->s->buffer->buf.pipelineBarrier(
+            vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer,
+            vk::PipelineStageFlagBits::eHost,
+            {},
+            { { vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferWrite,
+                vk::AccessFlagBits::eHostRead } },
+            {}, {});
+        ggml_vk_ctx_end(subctx);
+        ggml_vk_submit(subctx, src->device->fence);
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX),
+                 "vk_buffer_read_2d uma waitForFences", src->device);
+        src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
+
+        if (width == spitch && width == dpitch) {
+            memcpy(dst, (const uint8_t *) src->ptr + offset, width * height);
+        } else {
+            for (size_t i = 0; i < height; i++) {
+                memcpy((uint8_t *) dst + i * dpitch, (const uint8_t *) src->ptr + offset + i * spitch, width);
+            }
+        }
+    } else {
+        std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
+
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue->cmd_pool);
+        ggml_vk_ctx_begin(src->device, subctx);
+        bool ret = ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, spitch, dpitch, width, height, true);
+        GGML_ASSERT(ret);
+        ggml_vk_ctx_end(subctx);
+
+        ggml_vk_submit(subctx, src->device->fence);
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read_2d waitForFences", src->device);
+        src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
+
+        for (auto& cpy : subctx->out_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+    }
+}
+
+void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")");
+    ggml_vk_buffer_read_2d(src, offset, dst, size, size, size, 1);
+}
+
+void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_copy_async(" << size << ")");
+    // Make sure both buffers are on same device
+    GGML_ASSERT(src->device == dst->device);
+
+    VkBufferCopy bc{ src_offset, dst_offset, size };
+
+    vkCmdCopyBuffer(ctx->s->buffer->buf, (VkBuffer)src->buffer, (VkBuffer)dst->buffer, 1, &bc);
+}
+
+void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
+    if (src->device == dst->device) {
+        std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
+        VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")");
+        // Copy within the device
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue->cmd_pool);
+        ggml_vk_ctx_begin(src->device, subctx);
+        ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
+        ggml_vk_ctx_end(subctx);
+        ggml_vk_submit(subctx, src->device->fence);
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences", src->device);
+        src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
+    } else {
+        VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")");
+        // Copy device to device
+        ggml_vk_ensure_sync_staging_buffer(src->device, size);
+
+        // Copy to src staging buffer
+        ggml_vk_buffer_copy(src->device->sync_staging, 0, src, src_offset, size);
+        // Copy to dst buffer
+        ggml_vk_buffer_write(dst, dst_offset, src->device->sync_staging->ptr, size);
+    }
+}
+
+void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_memset_async(" << offset << ", " << c << ", " << size << ")");
+
+    if (dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible &&
+        dst->device->uma) {
+        deferred_memset((uint8_t*)dst->ptr + offset, c, size, &ctx->memsets);
+        return;
+    }
+
+    // Fall back to GPU fillBuffer for non-UMA or non-host-visible buffers
+    ctx->s->buffer->buf.fillBuffer(dst->buffer, offset, size, c);
+}
+
+void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")");
+
+    if (dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible &&
+        dst->device->uma) {
+        memset((uint8_t*)dst->ptr + offset, c, size);
+        return;
+    }
+
+    std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
+    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue->cmd_pool);
+    ggml_vk_ctx_begin(dst->device, subctx);
+    subctx->s->buffer->buf.fillBuffer(dst->buffer, offset, size, c);
+    ggml_vk_ctx_end(subctx);
+
+    ggml_vk_submit(subctx, dst->device->fence);
+    VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences", dst->device);
+    dst->device->device.resetFences({ dst->device->fence });
+    ggml_vk_queue_command_pools_cleanup(dst->device);
+}
+
+ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_vk_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_vk_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_vk_buffer_init_tensor,
+    /* .memset_tensor   = */ ggml_backend_vk_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_vk_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_vk_buffer_get_tensor,
+    /* .set_tensor_2d   = */ ggml_backend_vk_buffer_set_tensor_2d,
+    /* .get_tensor_2d   = */ ggml_backend_vk_buffer_get_tensor_2d,
+    /* .cpy_tensor      = */ ggml_backend_vk_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_vk_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+vk_buffer ggml_vk_buffer_from_host_ptr(vk_device & device, void * ptr, size_t size) {
+    if (!device->external_memory_host) {
+        return {};
+    }
+
+    uintptr_t uptr = reinterpret_cast<uintptr_t>(ptr);
+    if (uptr & (device->min_imported_host_pointer_alignment - 1)) {
+        return {};
+    }
+    if (size & (device->min_imported_host_pointer_alignment - 1)) {
+        return {};
+    }
+
+    const vk::MemoryPropertyFlags property_flags = vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached;
+
+    vk_buffer buf {};
+    try {
+        buf = ggml_vk_create_buffer(device, size, { property_flags }, ptr);
+    } catch (vk::SystemError& e) {
+        GGML_LOG_WARN("ggml_vulkan: Failed ggml_vk_create_buffer (%s)\n", e.what());
+    }
+
+    return buf;
+}
+
@@ -0,0 +1,282 @@
+#pragma once
+#include "ggml-vulkan-push-constants.h"
+
+// shared globals
+extern ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface;
+extern bool vk_memory_logger_enabled;
+extern bool vk_perf_logger_enabled;
+extern bool vk_perf_logger_concurrent;
+extern bool vk_enable_sync_logger;
+extern uint32_t vk_perf_logger_frequency;
+extern std::string vk_pipeline_stats_filter;
+extern void * const vk_ptr_base;
+extern vk_instance_t vk_instance;
+extern ggml_backend_buffer_i ggml_backend_vk_buffer_interface;
+
+// instance
+vk_device ggml_vk_get_device(size_t idx);
+DispatchLoaderDynamic & ggml_vk_default_dispatcher();
+void ggml_vk_instance_init();
+void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx);
+int ggml_vk_get_device_count();
+void ggml_vk_get_device_description(int device, char * description, size_t description_size);
+bool ggml_vk_instance_layer_settings_available();
+bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
+bool ggml_vk_instance_debug_utils_ext_available(const std::vector<vk::ExtensionProperties> & instance_extensions);
+bool ggml_vk_device_is_supported(const vk::PhysicalDevice & vkdev);
+bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch);
+uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev);
+bool ggml_vk_intel_windows_driver_in_range(uint32_t driver_version, uint32_t lower_major, uint32_t lower_minor, uint32_t upper_major, uint32_t upper_minor);
+
+// shaders
+void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
+vk_fa_tuning_params get_fa_tuning_params(const vk_device& device, uint32_t hsk, uint32_t hsv, uint32_t n_rows, uint32_t n_kv, ggml_type k_type, ggml_type v_type, bool f32acc);
+vk_fa_pipeline_state get_fa_pipeline_state(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool aligned, bool f32acc, bool use_mask, bool use_mask_opt, bool use_logit_softcap, bool use_sparse, ggml_type k_type, ggml_type v_type);
+uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_device_architecture &arch);
+void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested = nullptr);
+bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type, ggml_type v_type);
+bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type k_type = GGML_TYPE_F16, ggml_type v_type = GGML_TYPE_F16);
+
+// buffers
+vk_buffer ggml_vk_create_buffer_check(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0));
+vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size);
+void ggml_vk_destroy_buffer(vk_buffer& buf);
+void * ggml_vk_host_malloc(vk_device& device, size_t size);
+void ggml_vk_host_free(vk_device& device, void* ptr);
+void ggml_vk_host_get(const vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset);
+void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size);
+void ggml_vk_ensure_sync_staging_buffer(ggml_backend_vk_context * ctx, size_t size);
+bool ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false);
+bool ggml_vk_buffer_write_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, bool sync_staging = false);
+void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t dpitch, size_t width, size_t height);
+void ggml_vk_buffer_write(vk_buffer& dst, size_t offset, const void * src, size_t size);
+bool ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false);
+void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height);
+void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size);
+void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size);
+void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size);
+void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t offset, uint32_t c, size_t size);
+void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size);
+vk_buffer ggml_vk_buffer_from_host_ptr(vk_device & device, void * ptr, size_t size);
+
+// pipelines
+uint64_t vk_tensor_offset(const ggml_tensor * tensor);
+uint32_t get_misalign_bytes(const ggml_backend_vk_context * ctx, const ggml_tensor * t);
+void ggml_vk_wait_for_fence(ggml_backend_vk_context * ctx);
+void ggml_pipeline_request_descriptor_sets(ggml_backend_vk_context *ctx, vk_pipeline& pipeline, uint32_t n);
+void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
+void ggml_vk_submit(vk_context& ctx, vk::Fence fence);
+uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyProperties>& queue_family_props, const vk::QueueFlags& required, const vk::QueueFlags& avoid, int32_t compute_index, uint32_t min_num_queues);
+std::unique_ptr<vk_queue> ggml_vk_create_queue(vk_device& device, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags, bool transfer_only);
+std::unique_ptr<vk_queue> ggml_vk_create_aliased_queue(vk_device& device, const std::unique_ptr<vk_queue>& source);
+vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_command_pool& p);
+vk_context ggml_vk_create_temporary_context(vk_command_pool& p);
+void ggml_vk_command_pool_cleanup(vk_device& device, vk_command_pool& p);
+void ggml_vk_queue_command_pools_cleanup(vk_device& device);
+vk_subbuffer ggml_vk_subbuffer(const ggml_backend_vk_context* ctx, const vk_buffer& buf, size_t offset = 0);
+void ggml_vk_sync_buffers(ggml_backend_vk_context* ctx, vk_context& subctx);
+void ggml_vk_set_event(vk_context& ctx, vk::Event& event);
+void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events);
+vk_subbuffer ggml_vk_tensor_subbuffer(const ggml_backend_vk_context * ctx, const ggml_tensor * tensor, bool allow_misalign = false);
+void ggml_vk_cmd_label_begin(vk::CommandBuffer buf, const char * name);
+void ggml_vk_ctx_end(vk_context& ctx);
+void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx);
+vk_context ggml_vk_get_compute_ctx(ggml_backend_vk_context * ctx);
+vk_context ggml_vk_get_transfer_ctx(ggml_backend_vk_context * ctx);
+bool ggml_vk_submit_transfer_ctx(ggml_backend_vk_context * ctx);
+size_t ggml_vk_align_size(size_t width, size_t align);
+void deferred_memcpy(void * dst, const void * src, size_t size, std::vector<vk_staging_memcpy>* memcpys = nullptr);
+void deferred_memset(void * dst, uint32_t val, size_t size, std::vector<vk_staging_memset>* memsets = nullptr);
+
+// matmul
+vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type type);
+void ggml_vk_matmul(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline, vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3, uint32_t padded_n);
+bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor);
+vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src, const ggml_tensor * dst, ggml_type to);
+vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type);
+void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, const vk_subbuffer & in, const vk_subbuffer & out, uint32_t ne);
+void ggml_vk_dsv4_hc_comb(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * mixes, const ggml_tensor * scale, const ggml_tensor * base, ggml_tensor * dst);
+void ggml_vk_dsv4_hc_pre(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * x, const ggml_tensor * weights, ggml_tensor * dst);
+void ggml_vk_dsv4_hc_post(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * x, const ggml_tensor * residual, const ggml_tensor * post, const ggml_tensor * comb, ggml_tensor * dst);
+void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx);
+bool ggml_vk_use_mul_mat_vec_id(const struct ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx);
+
+// flash-attn
+void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, const ggml_tensor * sinks, ggml_tensor * dst);
+
+// operators
+void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor, const vk_subbuffer & in, const vk_subbuffer & out);
+bool ggml_vk_can_use_fwht(const ggml_backend_vk_context * ctx, const ggml_tensor * src1, const ggml_tensor * dst);
+void ggml_vk_fwht(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src, ggml_tensor * dst);
+void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_get_rows_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_out_prod(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+int ggml_vk_unary_mul_op_index(ggml_unary_op op);
+void ggml_vk_unary_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst);
+void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_gated_linear_attn(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_lightning_indexer(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_ssm_conv(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst);
+void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_add1(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_arange(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_fill(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_log(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_tri(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_diag(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_pad_reflect_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+uint32_t ggml_vk_rms_partials_size(ggml_backend_vk_context * ctx, const ggml_tensor *node);
+void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx, float * op_params);
+void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_xielu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst);
+void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_cgraph * cgraph, int node_idx, bool backprop);
+void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_topk_qsa(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_cumsum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_cross_entropy_loss(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_cross_entropy_loss_back(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst);
+void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_solve_tri(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_im2col_3d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_col2im_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_snake_dispatch_fused(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_pool_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_conv_3d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst);
+
+// graph
+void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_context subctx);
+bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool last_node, bool almost_ready, bool submit);
+void ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool almost_ready);
+void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx);
+void ggml_vk_cleanup(ggml_backend_vk_context * ctx);
+void ggml_vk_synchronize(ggml_backend_vk_context * ctx);
+bool ggml_vk_is_empty(ggml_tensor * node);
+bool ggml_vk_can_fuse(const ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops);
+bool ggml_vk_can_fuse_ssm_conv(const ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx, int num_extra);
+bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx, topk_moe_mode mode);
+bool ggml_vk_can_fuse_topk_qsa(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx);
+bool ggml_vk_can_fuse_rope_set_rows(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx);
+bool ggml_vk_can_fuse_rms_norm_set_rows(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx);
+bool ggml_vk_can_fuse_snake(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx);
+bool ggml_vk_tensors_overlap(const ggml_tensor * a, const ggml_tensor * b, bool elementwise);
+bool ggml_vk_can_fuse_rms_norm_mul_rope(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx);
+uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx);
+void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph * graph, struct ggml_backend_graph_optimize_params * params);
+
+// backend
+bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer);
+void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer);
+void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer);
+enum ggml_status ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor);
+void ggml_backend_vk_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size);
+void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+void ggml_backend_vk_buffer_set_tensor_2d(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size);
+void ggml_backend_vk_buffer_get_tensor_2d(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst);
+void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value);
+const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
+ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
+size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
+size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft);
+size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor);
+void ggml_backend_vk_free(ggml_backend_t backend);
+ggml_backend_reg_t ggml_backend_vk_reg();
+
+// debug
+int64_t ggml_vk_get_op_batch_size(const ggml_tensor * op);
+
+// ggml-vulkan.cpp (residual)
+bool ggml_vk_lightning_indexer_k_type_supported(ggml_type type);
+void ggml_vk_print_device_fault_info(const vk_device& device);
+uint64_t ggml_vk_get_node_flops(const ggml_tensor * node);
+void ggml_vk_print_node_list(const ggml_cgraph * cgraph, int start, int end);
+void ggml_vk_print_device_lost_info(const vk_device& device);
+size_t ggml_vk_tensor_buffer_offset(const ggml_backend_vk_context * ctx, const ggml_tensor * t);
+size_t ggml_vk_descriptor_offset(size_t tensor_offset, size_t alignment, size_t type_size);
+uint32_t ggml_vk_concat_unit_size(ggml_type type);
+bool ggml_vk_concat_supported(const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * dst);
+
+template <typename T>
+inline void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, const T &push_constants, std::array<uint32_t, 3> elements) {
+    const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]);
+    const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]);
+    const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
+    VK_LOG_DEBUG("ggml_vk_dispatch_pipeline(" << pipeline->name << ", {";
+    for (auto& buffer : descriptor_buffer_infos) {
+        std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.range << "), ";
+    }
+    std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))");
+    GGML_ASSERT(wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] &&
+                wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] &&
+                wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+    GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size());
+    GGML_ASSERT(descriptor_buffer_infos.size() <= MAX_PARAMETER_COUNT);
+    GGML_ASSERT(pipeline->parameter_count == descriptor_buffer_infos.size());
+    GGML_ASSERT(pipeline->push_constant_size == push_constant_size(push_constants));
+
+    vk::DescriptorSet& descriptor_set = ctx->descriptor_sets[ctx->descriptor_set_idx++];
+    vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
+    ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});
+
+    subctx->s->buffer->buf.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size(push_constants), push_constant_data(push_constants));
+    subctx->s->buffer->buf.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline);
+    subctx->s->buffer->buf.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                pipeline->layout,
+                                0,
+                                { descriptor_set },
+                                {});
+    {
+        ggml_vk_debug_label dbg(subctx, pipeline->name, wg0, wg1, wg2);
+        subctx->s->buffer->buf.dispatch(wg0, wg1, wg2);
+    }
+}
+
@@ -27,6 +27,8 @@ layout (push_constant) uniform parameter {
    uint inner_end;
 } p;

+shared int s;
+
 void argsort(bool needs_bounds_check, const uint row) {
    // bitonic sort
    int col = int(gl_GlobalInvocationID.x);
@@ -35,6 +37,12 @@ void argsort(bool needs_bounds_check, const uint row) {
    const uint row_offset = row * p.ncols;
    uint idx_offset = row * p.ncols_padded;

+    // workaround for NV driver/compiler bug - dummy use of shared memory
+    if (gl_LocalInvocationIndex == 0) {
+        s = 0;
+        col += s;
+    }
+
    bool need_barrier = false;

    // initialize indices
@@ -30,9 +30,14 @@ layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {uint data_a[];};
 layout (binding = 1) writeonly buffer D {uint data_d[];};

-shared uint vals[BLOCK_SIZE];
-shared uint offsets[BLOCK_SIZE];
-shared uint cursors[BLOCK_SIZE];
+// Upper bound on n_experts for the hoisted row-id path. Must match the limit in
+// ggml_vk_mul_mat_id_q_f16 (hoist_row_ids). The non-hoisted reduction below only
+// needs BLOCK_SIZE entries.
+#define MAX_EXPERTS 1024
+
+shared uint vals[MAX_EXPERTS];
+shared uint offsets[MAX_EXPERTS];
+shared uint cursors[MAX_EXPERTS];

 // data_d layout when p.hoist_row_ids is set:
 //   [0,              n_experts)   per-expert row count
@@ -46,8 +51,8 @@ void main() {
    const uint tid = gl_LocalInvocationID.x;

    if (p.hoist_row_ids != 0) {
-        if (tid < p.n_experts) {
-            vals[tid] = 0;
+        for (uint e = tid; e < p.n_experts; e += BLOCK_SIZE) {
+            vals[e] = 0;
        }
        barrier();

@@ -7,8 +7,13 @@
 //
 //   dst[i0, idst, it] = x[i0, it]*post[idst, it]
 //                     + sum_isrc residual[i0, isrc, it]*comb[idst, isrc, it]
+//
+// HAS_COMB == 0: identity mixing, each stream keeps its own residual:
+//
+//   dst[i0, idst, it] = x[i0, it]*post[idst, it] + residual[i0, idst, it]

 layout(constant_id = 0) const uint BLOCK_SIZE = 256;
+layout(constant_id = 1) const uint HAS_COMB = 1;

 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

@@ -48,7 +53,7 @@ void main() {
    if (tid < hc) {
        post_s[tid] = data_p[p_offset + tid * nbp0 + it * nbp1];
    }
-    if (tid < hc * hc) {
+    if (HAS_COMB == 1 && tid < hc * hc) {
        const uint idst = tid & 3;
        const uint isrc = tid >> 2;
        comb_s[tid] = data_c[c_offset + idst * nbc0 + isrc * nbc1 + it * nbc2];
@@ -74,9 +79,13 @@ void main() {
    [[unroll]]
    for (uint idst = 0; idst < hc; ++idst) {
        float result = xv * post_s[idst];
-        [[unroll]]
-        for (uint isrc = 0; isrc < hc; ++isrc) {
-            result = fma(r[isrc], comb_s[idst + hc * isrc], result);
+        if (HAS_COMB == 1) {
+            [[unroll]]
+            for (uint isrc = 0; isrc < hc; ++isrc) {
+                result = fma(r[isrc], comb_s[idst + hc * isrc], result);
+            }
+        } else {
+            result += r[idst];
        }
        data_d[d_offset + i0 * nbd0 + idst * nbd1 + it * nbd2] = result;
    }
@@ -4,9 +4,14 @@

 // Collapse the hc residual streams of a token into one, weighted per stream:
 //
-//   dst[i0, it] = sum_ih x[i0, ih, it] * weights[ih, it]
+//   dst[i0, it] = scale * sum_ih x[i0, ih, it] * weights[ih, it]
+//
+// GATED: weights is a per-element gate [n_embd, hc, n_tokens], applied as sigmoid:
+//
+//   dst[i0, it] = scale * sum_ih x[i0, ih, it] * sigmoid(gate[i0, ih, it])

 layout(constant_id = 0) const uint BLOCK_SIZE = 256;
+layout(constant_id = 1) const uint GATED = 0;

 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

@@ -16,12 +21,14 @@ layout(push_constant) uniform parameter
    uint n_tokens;

    uint nbx0; uint nbx1; uint nbx2;   // x
-    uint nbw0; uint nbw1;              // weights
+    uint nbw0; uint nbw1; uint nbw2;   // weights / gate
    uint nbd0; uint nbd1;              // dst

    uint x_offset;
    uint w_offset;
    uint d_offset;
+
+    float scale;
 };

 layout(binding = 0, std430) readonly buffer X { float data_x[]; };
@@ -36,10 +43,12 @@ void main() {
    const uint tid = gl_LocalInvocationID.x;
    const uint it  = gl_WorkGroupID.y;

-    if (tid < hc) {
-        w[tid] = data_w[w_offset + tid * nbw0 + it * nbw1];
+    if (GATED == 0) {
+        if (tid < hc) {
+            w[tid] = data_w[w_offset + tid * nbw0 + it * nbw1];
+        }
+        barrier();
    }
-    barrier();

    // After the barrier, so every invocation reaches it.
    const uint i0 = gl_WorkGroupID.x * BLOCK_SIZE + tid;
@@ -48,12 +57,19 @@ void main() {
    }

    const uint xb = x_offset + i0 * nbx0 + it * nbx2;
+    const uint wb = w_offset + i0 * nbw0 + it * nbw2;

    float result = 0.0f;
    [[unroll]]
    for (uint ih = 0; ih < hc; ++ih) {
-        result = fma(data_x[xb + ih * nbx1], w[ih], result);
+        float wv;
+        if (GATED == 1) {
+            wv = 1.0f / (1.0f + exp(-data_w[wb + ih * nbw1]));
+        } else {
+            wv = w[ih];
+        }
+        result = fma(data_x[xb + ih * nbx1], wv, result);
    }

-    data_d[d_offset + i0 * nbd0 + it * nbd1] = result;
+    data_d[d_offset + i0 * nbd0 + it * nbd1] = scale * result;
 }
@@ -31,7 +31,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};

 #if BDA
-layout (buffer_reference) buffer D_ptr {D_TYPE d;};
+layout (buffer_reference, buffer_reference_align = D_SIZE) buffer D_ptr {D_TYPE d;};
 #endif

 void im2col(const uint ow, const uint z_idx) {
@@ -50,7 +50,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};

 #if BDA
-layout (buffer_reference) buffer D_ptr {D_TYPE d;};
+layout (buffer_reference, buffer_reference_align = D_SIZE) buffer D_ptr {D_TYPE d;};
 #endif

 void main() {
@@ -313,6 +313,9 @@ void main() {

    // Workgroup has no work
    if (ic * BN >= _ne1) return;
+
+    uint required_work_items = (_ne1 - ic * BN) * BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B;
+    uint required_warp_c = (_ne1 - ic * BN + WN - 1) / WN;
 #endif

 #ifdef MUL_MAT_ID
@@ -363,6 +366,9 @@ void main() {
        [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
            load_a_to_shmem(pos_a, loadr_a, loadc_a + l, ir * BM + loadc_a + l, block, end_k);
        }
+        #ifdef MUL_MAT_ID
+        if (gl_LocalInvocationID.x < required_work_items) {
+        #endif
        [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
 #if !defined(MUL_MAT_ID)
            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic * BN + loadc_b + l, block, end_k);
@@ -370,6 +376,9 @@ void main() {
            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic, _ne1, block, end_k);
 #endif
        }
+        #ifdef MUL_MAT_ID
+        }
+        #endif

        barrier();

@@ -377,6 +386,9 @@ void main() {
        pos_b += BK / LOAD_VEC_B_EFF;

 #ifdef COOPMAT
+#ifdef MUL_MAT_ID
+        if (warp_c < required_warp_c) {
+#endif
        [[unroll]] for (uint i = 0; i < BK; i += TK) {
            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
                // Load from shared into cache
@@ -389,6 +401,9 @@ void main() {
                }
            }
        }
+#ifdef MUL_MAT_ID
+        }
+#endif
 #else
        [[unroll]] for (uint i = 0; i < BK / BK_STEP; i++) {
            // Load from shared into cache
@@ -454,6 +454,55 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
 }
 #endif

+#if defined(DATA_A_IQ3_S)
+// 2-byte loads for IQ3_S blocks (110 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint ib32 = ib % 8;
+
+    // grid indices for qs[2 * iqs] and qs[2 * iqs + 1]
+    const uint qs    = uint(data_a_packed16[ib_k].qs[ib32 * 4 + iqs]);
+    // their two high index bits
+    const uint qh    = uint(data_a_packed16[ib_k].qh[ib32 / 2]) >> ((ib32 & 1) * 8 + 2 * iqs);
+    // one sign bit per value, 8 values
+    const uint signs = uint(data_a_packed16[ib_k].signs[ib32 * 2 + iqs / 2]) >> ((iqs & 1) * 8);
+
+    // grid holds 4 values of 1..15, one per byte
+    const ivec4 vals0 = ivec4(unpack8(iq3s_grid[( qs       & 0xFF) | ((qh & 1) << 8)]));
+    const ivec4 vals1 = ivec4(unpack8(iq3s_grid[((qs >> 8) & 0xFF) | ((qh & 2) << 7)]));
+
+    // negate with (v ^ -s) - -s to avoid branches
+    const ivec4 m0 = -(ivec4(signs,      signs >> 1, signs >> 2, signs >> 3) & 1);
+    const ivec4 m1 = -(ivec4(signs >> 4, signs >> 5, signs >> 6, signs >> 7) & 1);
+
+    buf_a[buf_ib].qs[2 * iqs    ] = pack32(i8vec4((vals0 ^ m0) - m0));
+    buf_a[buf_ib].qs[2 * iqs + 1] = pack32(i8vec4((vals1 ^ m1) - m1));
+
+    if (iqs == 0) {
+        const uint scale = (uint(data_a_packed16[ib_k].scales[ib32 / 4]) >> ((ib32 & 3) * 4)) & 0xF;
+
+        buf_a[buf_ib].d = FLOAT_TYPE(float(data_a_packed16[ib_k].d) * float(1 + 2 * scale));
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].d = buf_a[buf_ib].d;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t q_sum = 0;
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        q_sum += dotPacked4x8EXT(cache_a[ib_a].qs[iqs], cache_b.qs[iqs]);
+    }
+
+    return ACC_TYPE(float(cache_a[ib_a].d) * float(cache_b.ds.x) * float(q_sum));
+}
+#endif
+
 void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
    if (is_in_bounds) {
        const uint ib_outer = ib / 4;
@@ -59,6 +59,12 @@ struct block_a_cache {
    int32_t qs[8];
    FLOAT_TYPE d;
 };
+#elif defined(DATA_A_IQ3_S)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE d;
+};
 #elif defined(DATA_A_Q2_K)
 #define QUANT_R_MMQ 4
 struct block_a_cache {
@@ -624,7 +624,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
        };

 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (!f16acc && !coopmat && !coopmat2 && !dot2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
+        if (!f16acc && !coopmat && !coopmat2 && !dot2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4" || tname == "iq3_s")) {
            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
        }
 #endif
@@ -106,6 +106,11 @@ struct ggml_webgpu_generic_shader_decisions {
    bool     inplace = false;
 };

+struct ggml_webgpu_get_rows_shader_decisions {
+    uint32_t wg_size    = 0;
+    bool     vectorized = false;
+};
+
 struct ggml_webgpu_binary_shader_decisions {
    uint32_t wg_size     = 0;
    bool     inplace     = false;
@@ -1551,8 +1556,8 @@ class ggml_webgpu_shader_lib {
        return argsort_merge_pipelines[order];
    }

-    webgpu_pipeline get_get_rows_pipeline(const ggml_webgpu_shader_lib_context & context) {
-        const bool vectorized                 = context.src0->type == GGML_TYPE_F32 && context.dst->ne[0] % 4 == 0;
+    webgpu_pipeline get_get_rows_pipeline(const ggml_webgpu_shader_lib_context & context, bool vec4_aligned) {
+        const bool vectorized = context.src0->type == GGML_TYPE_F32 && context.dst->ne[0] % 4 == 0 && vec4_aligned;
        ggml_webgpu_get_rows_pipeline_key key = {};
        key.src_type                          = context.src0->type;
        key.vectorized                        = (int) vectorized;
@@ -1669,8 +1674,9 @@ class ggml_webgpu_shader_lib {
        defines.push_back("WG_SIZE=" + std::to_string(context.max_wg_size));

        auto processed           = preprocessor.preprocess(wgsl_get_rows, defines);
-        auto decisions           = std::make_shared<ggml_webgpu_generic_shader_decisions>();
+        auto decisions           = std::make_shared<ggml_webgpu_get_rows_shader_decisions>();
        decisions->wg_size       = context.max_wg_size;
+        decisions->vectorized    = vectorized;
        webgpu_pipeline pipeline = ggml_webgpu_create_pipeline(device, processed, variant);
        pipeline.context         = decisions;
        get_rows_pipelines[key]  = pipeline;
@@ -1518,15 +1518,24 @@ static webgpu_encoded_op ggml_webgpu_get_rows(webgpu_context & ctx,
    shader_lib_ctx.dst                            = dst;
    shader_lib_ctx.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup;

-    webgpu_pipeline pipeline  = ctx->shader_lib->get_get_rows_pipeline(shader_lib_ctx);
-    auto *          decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
+    const uint32_t offset_src  = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type));
+    const uint32_t offset_dst  = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type));
+    const uint32_t stride_src1 = (uint32_t) (src->nb[1] / ggml_type_size(src->type));
+    const uint32_t stride_src2 = (uint32_t) (src->nb[2] / ggml_type_size(src->type));
+    const uint32_t stride_src3 = (uint32_t) (src->nb[3] / ggml_type_size(src->type));

-    std::vector<uint32_t> params = { (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+    const bool vec4_aligned = offset_src % 4 == 0 && offset_dst % 4 == 0 && stride_src1 % 4 == 0 &&
+                              stride_src2 % 4 == 0 && stride_src3 % 4 == 0;
+
+    webgpu_pipeline pipeline  = ctx->shader_lib->get_get_rows_pipeline(shader_lib_ctx, vec4_aligned);
+    auto *          decisions = static_cast<ggml_webgpu_get_rows_shader_decisions *>(pipeline.context.get());
+
+    std::vector<uint32_t> params = { offset_src,
                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, idx) / ggml_type_size(idx->type)),
-                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
-                                     (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
-                                     (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
-                                     (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
+                                     offset_dst,
+                                     stride_src1,
+                                     stride_src2,
+                                     stride_src3,
                                     (uint32_t) (idx->nb[0] / ggml_type_size(idx->type)),
                                     (uint32_t) (idx->nb[1] / ggml_type_size(idx->type)),
                                     (uint32_t) (idx->nb[2] / ggml_type_size(idx->type)),
@@ -1544,7 +1553,7 @@ static webgpu_encoded_op ggml_webgpu_get_rows(webgpu_context & ctx,
                                                  ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, idx),
                                                  ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst) };

-    uint32_t blocks_per_row = (uint32_t) (dst->ne[0] / (src->type == GGML_TYPE_F32 && dst->ne[0] % 4 == 0 ? 4 : 1));
+    uint32_t blocks_per_row = (uint32_t) (dst->ne[0] / (decisions->vectorized ? 4 : 1));
    uint32_t total_rows     = (uint32_t) (dst->ne[1] * dst->ne[2] * dst->ne[3]);
    uint32_t total_threads  = float_parallel ? blocks_per_row * total_rows : total_rows;
    uint32_t wg_x           = CEIL_DIV(total_threads, decisions->wg_size);
@@ -4333,22 +4342,12 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                           src0->type == GGML_TYPE_F32 && (src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32));
            break;
        case GGML_OP_GET_ROWS:
-            {
-                const size_t storage_alignment =
-                    ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment;
-                const size_t src_address_unit =
-                    src0->type == GGML_TYPE_F32 && op->ne[0] % 4 == 0 ? 4 * sizeof(float) : ggml_type_size(src0->type);
-                if (ggml_webgpu_tensor_misalignment(src0, storage_alignment) % src_address_unit != 0) {
-                    break;
-                }
-                if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 ||
-                    ggml_webgpu_supported_qtype(src0->type)) {
-                    supports_op = (op->type == GGML_TYPE_F32);
-                } else if (src0->type == GGML_TYPE_I32) {
-                    supports_op = op->type == GGML_TYPE_I32;
-                }
-                break;
+            if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_webgpu_supported_qtype(src0->type)) {
+                supports_op = (op->type == GGML_TYPE_F32);
+            } else if (src0->type == GGML_TYPE_I32) {
+                supports_op = op->type == GGML_TYPE_I32;
            }
+            break;
        case GGML_OP_MUL_MAT:
            {
                switch (src1->type) {
@@ -238,6 +238,7 @@ struct gguf_reader {
        : callback(callback),
          userdata(userdata),
          max_chunk_read(max_chunk_read),
+          start_offset(data_offset),
          data_offset(data_offset),
          nbytes_remain(nbytes_remain) {
        GGML_ASSERT(max_chunk_read > 0);
@@ -366,6 +367,11 @@ struct gguf_reader {
        return data_offset;
    }

+    // position in the file where the GGUF data starts, alignment is relative to it, not to the file
+    uint64_t start() const {
+        return start_offset;
+    }
+
    bool seek(uint64_t absolute_offset) const {
        const uint64_t end_offset = uint64_t(data_offset) + nbytes_remain;
        if (absolute_offset > end_offset) {
@@ -415,6 +421,7 @@ private:
    gguf_reader_callback_t callback = nullptr;
    void * userdata = nullptr;
    size_t max_chunk_read = 0;
+    uint64_t start_offset = 0;
    mutable uint64_t data_offset = 0;
    mutable uint64_t nbytes_remain = 0;
 };
@@ -763,7 +770,7 @@ static struct gguf_context * gguf_init_from_reader(const struct gguf_reader & gr
    GGML_ASSERT(int64_t(ctx->info.size()) == n_tensors);

    // we require the data section to be aligned, so take into account any padding
-    if (n_tensors > 0 && !gr.seek(GGML_PAD(gr.tell(), ctx->alignment))) {
+    if (n_tensors > 0 && !gr.seek(gr.start() + GGML_PAD(gr.tell() - gr.start(), ctx->alignment))) {
        GGML_LOG_ERROR("%s: failed to seek to beginning of data section\n", __func__);
        gguf_free(ctx);
        return nullptr;
@@ -5876,7 +5876,7 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
    GGMLQuantizationType.Q5_0:    (32, 2 + 4 + 16),
    GGMLQuantizationType.Q5_1:    (32, 2 + 2 + 4 + 16),
    GGMLQuantizationType.Q8_0:    (32, 2 + 32),
-    GGMLQuantizationType.Q8_1:    (32, 4 + 4 + 32),
+    GGMLQuantizationType.Q8_1:    (32, 2 + 2 + 32),
    GGMLQuantizationType.Q2_K:    (256, 2 + 2 + QK_K // 16 + QK_K // 4),
    GGMLQuantizationType.Q3_K:    (256, 2 + QK_K // 4 + QK_K // 8 + 12),
    GGMLQuantizationType.Q4_K:    (256, 2 + 2 + QK_K // 2 + 12),
@@ -518,6 +518,8 @@ extern "C" {
              struct llama_model_params   params);

    // Load a model from an open FILE pointer
+    // The GGUF is read from the current position, so it can be embedded in a larger file
+    // mmap needs the GGUF data section at a file offset to be aligned to the CPU tensor alignment (32 bytes)
    LLAMA_API struct llama_model * llama_model_load_from_file_ptr(
                                   FILE * file,
              struct llama_model_params   params);
@@ -681,6 +683,11 @@ extern "C" {
            struct llama_model * model,
            const char * path_lora);

+    // Load a LoRA adapter from an open FILE pointer, reading from its current position
+    LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init_from_file_ptr(
+            struct llama_model * model,
+            FILE * file);
+
    // Functions to access the adapter's GGUF metadata scalar values
    // - The functions return the length of the string on success, or -1 on failure
    // - The output string is always null-terminated and cleared on failure
@@ -8,7 +8,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})

 if (EMSCRIPTEN)
 else()
-    if (NOT GGML_BACKEND_DL)
+    if (NOT GGML_BACKEND_DL AND GGML_CPU)
        add_subdirectory(vdot)
    endif()
 endif()
@@ -55,7 +55,9 @@ fi
 # Some generic functions
 usage() {
    echo "Usage: $0 [ --include-path <dir> ] --generate <build-dir> libXXX [ libYYY ... ]" >&2
-    echo "       $0 --check <old-build-dir> <new-build-dir>" >&2
+    echo "       $0 [ --strict ] --check <old-build-dir> <new-build-dir>" >&2
+    echo "" >&2
+    echo "  --strict: fail on any API/ABI change, including backwards-compatible additions" >&2
 }

 get_cmake_project_name() {
@@ -72,6 +74,7 @@ get_cmake_version() {
 # Option parsing and validation
 DO_GEN=0
 DO_CHECK=0
+STRICT=0
 BUILD_DIR=
 BUILD_DIR_NEW=
 INCLUDE_PATHS=
@@ -119,6 +122,10 @@ while [ "$#" -gt 0 ]; do
        INCLUDE_PATHS="$INCLUDE_PATHS $2"
        shift 2
        ;;
+    --strict)
+        STRICT=1
+        shift
+        ;;
    -h | --help)
        usage
        exit 1
@@ -258,13 +265,24 @@ elif [ "$DO_CHECK" -eq 1 ]; then
        abidiff "$xml_file" "$xml_file_new"
        res=$?
        [ "$((res & 8))" -ne 0 ] && ABI_RESULT=1
+        # check bit 2 for compatible ABI changes (like new symbols) and if
+        # STRICT is set then handle this as an error.
+        [ "$STRICT" -eq 1 ] && [ "$((res & 4))" -ne 0 ] && ABI_RESULT=1
    done

    if [ "$API_RESULT" -gt 0 ]; then
-        echo "ERROR: API changed with possible backwards-compatibility problems." >&2
+        if [ "$STRICT" -eq 1 ]; then
+            echo "ERROR: API changed — a minor version bump is required." >&2
+        else
+            echo "ERROR: API changed with backwards-incompatible problems — a major version bump is required." >&2
+        fi
    fi
    if [ "$ABI_RESULT" -gt 0 ]; then
-        echo "ERROR: ABI changed with possible backwards-compatibility problems." >&2
+        if [ "$STRICT" -eq 1 ]; then
+            echo "ERROR: ABI changed — a minor version bump is required." >&2
+        else
+            echo "ERROR: ABI changed with backwards-incompatible problems — a major version bump is required." >&2
+        fi
    fi
    if [ "$((API_RESULT + ABI_RESULT))" -gt 0 ]; then
        exit 1
@@ -0,0 +1,131 @@
+#!/bin/bash
+# Check API/ABI compatibility between the previous release tag and current HEAD.
+#
+# Finds the most recent vX.Y.Z tag, checks it out in a temporary git worktree,
+# builds both versions with shared libs enabled, and uses check-apiabi-compat.sh
+# to compare the results.
+#
+# Exit codes:
+#   0: compatible, or check was skipped
+#   1: backwards-incompatible changes found, or build failed
+#
+# Options:
+#   --tag <version>: compare against this tag instead of the latest release
+#
+# Environment:
+#   SKIP_APIABI_CHECK:    set to 1 or true to skip
+#   APIABI_COMPARE_TAG:   equivalent to --tag (used by CI)
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+usage() {
+    echo "Usage: $0 [--tag <version>]" >&2
+    echo "  --tag <version>  Compare against this release tag (default: latest)" >&2
+}
+
+COMPARE_TAG="${APIABI_COMPARE_TAG:-}"
+while [[ "$#" -gt 0 ]]; do
+    case "$1" in
+    --tag)
+        if [[ -z "${2:-}" ]]; then usage; exit 1; fi
+        COMPARE_TAG="$2"
+        shift 2
+        ;;
+    --tag=*)
+        COMPARE_TAG="${1#*=}"
+        shift
+        ;;
+    -h | --help)
+        usage; exit 0
+        ;;
+    *)
+        usage; exit 1
+        ;;
+    esac
+done
+
+if [[ "${SKIP_APIABI_CHECK:-}" == "1" || "${SKIP_APIABI_CHECK:-}" == "true" ]]; then
+    echo "SKIP_APIABI_CHECK is set - skipping API/ABI compatibility check"
+    exit 0
+fi
+
+if ! command -v abi-compliance-checker >/dev/null 2>&1 || ! command -v abidw >/dev/null 2>&1; then
+    echo "Warning: abi-compliance-checker or abigail-tools not installed - skipping API/ABI check"
+    exit 0
+fi
+
+discover_libs() {
+    local build_dir="$1"
+    local libs=()
+    for dir in "$build_dir/src" "$build_dir/bin"; do
+        [[ -d "$dir" ]] || continue
+        for f in "$dir"/lib*.so; do
+            [[ -f "$f" ]] && libs+=("$(basename "$f" .so)")
+        done
+    done
+    echo "${libs[@]}"
+}
+
+if [[ -n "${COMPARE_TAG}" ]]; then
+    PREV_TAG="${COMPARE_TAG}"
+    if ! git -C "$REPO_ROOT" rev-parse --verify "${PREV_TAG}^{}" >/dev/null 2>&1; then
+        echo "Error: tag '${PREV_TAG}' not found in repository." >&2
+        exit 1
+    fi
+else
+    PREV_TAG=$(git -C "$REPO_ROOT" tag --sort=-v:refname | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' | head -n 1 || true)
+    if [[ -z "${PREV_TAG}" ]]; then
+        echo "Warning: no previous release tag found - skipping API/ABI check"
+        exit 0
+    fi
+fi
+OLD_VERSION="${PREV_TAG#v}"
+OLD_MAJOR="${OLD_VERSION%%.*}"
+OLD_MINOR="${OLD_VERSION#*.}"; OLD_MINOR="${OLD_MINOR%%.*}"
+
+NEW_MAJOR=$(grep "set(LLAMA_VERSION_MAJOR" "$REPO_ROOT/CMakeLists.txt" | sed 's/.*MAJOR \([0-9]*\).*/\1/')
+NEW_MINOR=$(grep "set(LLAMA_VERSION_MINOR" "$REPO_ROOT/CMakeLists.txt" | sed 's/.*MINOR \([0-9]*\).*/\1/')
+
+if [[ "$NEW_MAJOR" -gt "$OLD_MAJOR" ]]; then
+    echo "Major version increment ($OLD_MAJOR -> $NEW_MAJOR): API/ABI breaking changes are expected, skipping compatibility check."
+    exit 0
+fi
+
+CHECK_FLAGS=()
+if [[ "$NEW_MINOR" -eq "$OLD_MINOR" ]]; then
+    echo "Patch version bump detected: checking for any API/ABI changes (a minor bump is required if any are found)..."
+    CHECK_FLAGS+=(--strict)
+else
+    echo "Minor version bump detected: checking for backwards-incompatible API/ABI changes..."
+fi
+
+echo "Checking API/ABI compatibility against ${PREV_TAG}..."
+
+WORKTREE_DIR=$(mktemp -d)
+BUILD_OLD=$(mktemp -d)
+BUILD_NEW=$(mktemp -d)
+
+cleanup() {
+    git -C "$REPO_ROOT" worktree remove --force "$WORKTREE_DIR" 2>/dev/null || true
+    rm -rf "$WORKTREE_DIR" "$BUILD_OLD" "$BUILD_NEW"
+}
+trap cleanup EXIT
+
+git -C "$REPO_ROOT" worktree add "$WORKTREE_DIR" "$PREV_TAG"
+
+cmake -S "$WORKTREE_DIR" -B "$BUILD_OLD" -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=RelWithDebInfo
+cmake --build "$BUILD_OLD" --parallel "$(nproc)"
+OLD_LIBS=($(discover_libs "$BUILD_OLD"))
+echo "Libraries found in old build: ${OLD_LIBS[*]}"
+
+cmake -S "$REPO_ROOT"    -B "$BUILD_NEW" -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=RelWithDebInfo
+cmake --build "$BUILD_NEW" --parallel "$(nproc)"
+NEW_LIBS=($(discover_libs "$BUILD_NEW"))
+echo "Libraries found in new build: ${NEW_LIBS[*]}"
+
+(cd "$WORKTREE_DIR" && "$SCRIPT_DIR/check-apiabi-compat.sh" --include-path ggml/include --generate "$BUILD_OLD" "${OLD_LIBS[@]}")
+(cd "$REPO_ROOT"    && "$SCRIPT_DIR/check-apiabi-compat.sh" --include-path ggml/include --generate "$BUILD_NEW" "${NEW_LIBS[@]}")
+(cd "$REPO_ROOT"    && "$SCRIPT_DIR/check-apiabi-compat.sh" "${CHECK_FLAGS[@]}" --check "$BUILD_OLD" "$BUILD_NEW")
@@ -165,6 +165,23 @@ else
    fi
 fi

+echo "Checking API/ABI compatibility..."
+set +e
+bash "$SCRIPT_DIR/check-release-apiabi.sh"
+APIABI_RESULT=$?
+set -e
+if [[ $APIABI_RESULT -ne 0 ]]; then
+    if [[ "$DRY_RUN" == "true" ]]; then
+        echo "Warning: API/ABI check found backwards-incompatible changes (dry run, continuing)."
+        CHECKS_PASSED=false
+    else
+        echo "Error: API/ABI check found backwards-incompatible changes."
+        exit 1
+    fi
+else
+    echo "API/ABI compatibility check passed - OK"
+fi
+
 if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
    echo "checks_passed=${CHECKS_PASSED}" >> "$GITHUB_OUTPUT"
 fi
@@ -6,6 +6,8 @@

 #include <map>
 #include <cassert>
+#include <cerrno>
+#include <cstring>
 #include <sstream>
 #include <stdexcept>

@@ -146,22 +148,23 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
    return nullptr;
 }

-static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
-    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
-
+static void llama_adapter_lora_init_impl(llama_model & model, FILE * file, llama_adapter_lora & adapter) {
    ggml_context * ctx_init;
    gguf_init_params meta_gguf_params = {
        /* .no_alloc = */ true,
        /* .ctx      = */ &ctx_init,
    };

-    gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) };
+    gguf_context_ptr ctx_gguf { gguf_init_from_file_ptr(file, meta_gguf_params) };
    if (!ctx_gguf) {
-        throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
+        throw std::runtime_error("failed to load lora adapter from file");
    }

    ggml_context_ptr ctx { ctx_init };

+    // must come after gguf_init_from_file_ptr, the llama_file constructor moves the file position
+    llama_file gguf_file(file);
+
    // check metadata
    {
        const gguf_context * gguf_ctx = ctx_gguf.get();
@@ -393,7 +396,6 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_

    // set tensor data
    {
-        llama_file gguf_file(path_lora, "rb");
        std::vector<uint8_t> read_buf;
        auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) {
            const size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
@@ -421,10 +423,30 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
 }

 llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
+    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
+
+    FILE * file = ggml_fopen(path_lora, "rb");
+    if (!file) {
+        LLAMA_LOG_ERROR("%s: failed to open '%s': %s\n", __func__, path_lora, strerror(errno));
+        return nullptr;
+    }
+
+    llama_adapter_lora * adapter = llama_adapter_lora_init_from_file_ptr(model, file);
+    fclose(file);
+
+    return adapter;
+}
+
+llama_adapter_lora * llama_adapter_lora_init_from_file_ptr(llama_model * model, FILE * file) {
+    if (!file) {
+        LLAMA_LOG_ERROR("%s: file is NULL\n", __func__);
+        return nullptr;
+    }
+
    llama_adapter_lora * adapter = new llama_adapter_lora(model);

    try {
-        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
+        llama_adapter_lora_init_impl(*model, file, *adapter);
        return adapter;
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@@ -599,8 +599,11 @@ void llama_context::sched_reserve() {

    LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);

-    gf_res_prev.reset(new llm_graph_result(max_nodes));
+    for (auto & res : gf_res_prev) {
+        res.reset();
+    }
    gf_res_reserve.reset(new llm_graph_result(max_nodes));
+    gf_res_prev_active = nullptr;

    sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, cparams.pipeline_parallel, cparams.op_offload));

@@ -816,10 +819,14 @@ bool llama_context::memory_update(bool optimize) {
                }
        }

-        // reset the previous graph result to make sure that it won't be reused
-        // TODO: change the mctx->apply() to return information if a graph reserve is needed
-        //       reset the graph result only if the memory module did reset the scheduler
-        gf_res_prev->reset();
+        // reset the previous graph results to make sure that they won't be reused
+        // TODO: make mctx->apply() report if a graph reserve is needed, then reset graph results only if the memory module reset the scheduler
+        for (auto & res : gf_res_prev) {
+            if (res) {
+                res->reset();
+            }
+        }
+        gf_res_prev_active = nullptr;

        if (!mctx->apply()) {
            LLAMA_LOG_ERROR("%s: failed to apply memory update\n", __func__);
@@ -1340,14 +1347,14 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
        return nullptr;
    }

-    auto * res = gf_res_prev.get();
+    auto * res = get_gf_res_prev();
    auto * gf  = res->get_gf();

    // the new graph parameters
    // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
    const auto gparams = graph_params(res, ubatch, mctx, gtype);

-    if (!graph_reuse_disable && res->can_reuse(gparams)) {
+    if (!graph_reuse_disable && gf_res_prev_active == res && res->can_reuse(gparams)) {
        //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);

        // with pipeline parallelism, the previous graph_compute_async may still be running
@@ -1359,6 +1366,7 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll

        n_reused++;
    } else {
+        gf_res_prev_active = nullptr;
        res->reset();

        ggml_backend_sched_reset(sched.get());
@@ -1381,6 +1389,8 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
            ret = GGML_STATUS_ALLOC_FAILED;
            return nullptr;
        }
+
+        gf_res_prev_active = res;
    }

    // set the input data for the input tensors
@@ -2361,6 +2371,14 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
    return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }

+llm_graph_result * llama_context::get_gf_res_prev() {
+    auto & res = gf_res_prev[n_outputs > 0];
+    if (!res) {
+        res.reset(new llm_graph_result(gf_res_reserve->get_max_nodes()));
+    }
+    return res.get();
+}
+
 // pack sampler outputs into as few sequences as possible before using sequences without samplers
 static void ubatch_prepare_reserve(
              llama_ubatch                            & ubatch,
@@ -2430,8 +2448,13 @@ ggml_cgraph * llama_context::graph_reserve(

    ggml_backend_sched_reset(sched.get());

-    // when the scheduler is reset, we cannot reuse the old graph, so we reset the previous graph result to prevent that
-    gf_res_prev->reset();
+    // when the scheduler is reset, we cannot reuse old graphs, so we reset the previous graph results
+    for (auto & res : gf_res_prev) {
+        if (res) {
+            res->reset();
+        }
+    }
+    gf_res_prev_active = nullptr;

    // store the n_outputs as it is, and restore it afterwards
    // TODO: not sure if needed, might simplify in the future by removing this
@@ -3521,10 +3544,12 @@ void llama_context::opt_epoch_iter(
                break;
            }

-            auto * res = gf_res_prev.get();
+            auto * res = get_gf_res_prev();

            const auto gparams = graph_params(res, ubatch, mctx.get(), ctx_type_to_graph_type(cparams.ctx_type));

+            // the optimizer graph is allocated outside sched, so the next decode must rebuild
+            gf_res_prev_active = nullptr;
            res->reset();

            auto * gf = model.build_graph(gparams);
@@ -11,6 +11,7 @@
 #include "ggml-cpp.h"
 #include "ggml-opt.h"

+#include <array>
 #include <map>
 #include <vector>

@@ -254,6 +255,8 @@ public:
    bool set_sampler(llama_seq_id seq_id, llama_sampler * sampler);

 private:
+    llm_graph_result * get_gf_res_prev();
+
    llm_graph_params graph_params(
                        llm_graph_result * res,
                      const llama_ubatch & ubatch,
@@ -364,9 +367,12 @@ private:
    std::vector<ggml_backend_buffer_type_t> backend_buft;
    std::vector<size_t>                     backend_buf_exp_size; // expected buffer sizes

-    llm_graph_result_ptr gf_res_prev;
+    // Separate arenas give batches with and without outputs distinct CUDA graph cache keys.
+    std::array<llm_graph_result_ptr, 2> gf_res_prev;
    llm_graph_result_ptr gf_res_reserve;

+    llm_graph_result * gf_res_prev_active = nullptr;
+
    // host buffer for the model output (logits and embeddings)
    ggml_backend_buffer_ptr buf_output;

@@ -685,6 +685,13 @@ llama_model_loader::llama_model_loader(
            throw std::runtime_error(format("%s: failed to load model from file pointer", __func__));
        }

+        // mmap places tensors at their file offsets, so an embedded GGUF must be aligned in the file too
+        const size_t tensor_align = ggml_backend_buft_get_alignment(ggml_backend_cpu_buffer_type());
+        if (use_mmap && gguf_get_data_offset(metadata) % tensor_align != 0) {
+            throw std::runtime_error(format("%s: GGUF data section at file offset %zu is not %zu byte aligned, cannot mmap",
+                __func__, gguf_get_data_offset(metadata), tensor_align));
+        }
+
        get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
        llm_kv = LLM_KV(llm_arch_from_string(arch_name));

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
TheArchitectit Claude Opus 4.6 Oliver Simons	b23701f77d	cuda : fix CUB argsort corruption caused by in-place keys (#28389 ) argsort_f32_i32_cuda_cub called the one-shot DeviceRadixSort::SortPairs API with d_keys_in == d_keys_out (temp_keys, temp_keys). CUB's internal double-buffer ping-pong requires distinct key buffers: with aliased buffers the sort partially overwrites its own input mid-pass and emits a corrupted permutation, surfacing as intermittent garbage indices (e.g. backend top_k over a 248k-column vocab on Maxwell/CUDA 12.5/CCCL 2.x, which then triggered out-of-bounds gathers in downstream get_rows). Use a distinct keys-out buffer for all six call sites (plain and segmented, ascending and descending, size-query and execute). --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Oliver Simons <osimons@nvidia.com>	2026-09-19 07:32:52 +02:00
dsproule	60081bb2b5	opencl: add support for bin kernel `flash_attn_f32_f16_bin` (#29046 ) * opencl: add `flash_attn_f32_f16_bin` * opencl: guarded prefill fa	2026-09-18 16:32:31 -07:00
Todor Boinovski	2b1847030c	hexagon: add ROLL op support (#29105 )	2026-09-18 15:05:10 -07:00
Todor Boinovski	50631b3d2c	hexagon: im2col update (#29103 ) * ggml-hexagon: accept 1D and padded IM2COL ops * ggml-hexagon: make pure-DDR IM2COL kernel is_2D-aware * ggml-hexagon: extend IM2COL DMA patch-embed fast path to 1D * ggml-hexagon: add blocked-staging general IM2COL DMA kernel	2026-09-18 14:20:48 -07:00
Todor Boinovski	18a04f09c2	hexagon: HMX flash-attention head_dim padding (support DK=DV=72) (#26539 ) Allow HMX flash-attention to run with head_dim not a multiple of 64 (e.g. SigLIP head_dim=72), by operating on DK/DV rounded up to 64 with zero-filled tail lanes.	2026-09-18 13:15:08 -07:00
shaofeiqi	ec92815050	opencl: add bin kernel `kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin` (#28678 ) * opencl: add A8 Q6_K non-MoE binary kernel * opencl: fix layout compatibility	2026-09-18 10:50:15 -07:00
bri-prism	4fea119de3	ggml-cpu: add F16 input to the FWHT (#27779 ) * ggml-cpu: add F16 input to the FWHT The CPU FWHT accepts F32 input only. This change makes the source type a template parameter. The CPU path now accepts F16 input and F32 input. The CPU MUL_MAT reference now converts an F16 src1 to F32. It does this when the caller sets the Hadamard hint. No backend has an F16 FWHT kernel yet. The test cases come with the backend changes that add one. * ggml-cpu: assert the F16 FWHT input path, and use the bulk converter Address review feedback. The F16 branch writes plain floats into wdata, which is only correct when vec_dot_type is F32. That invariant held because supports_op only accepts an F16 src1 for the Hadamard hint with F32 src0 and dst, but nothing enforced it. Assert it next to the existing src1 type check so widening supports_op cannot silently break the write. Replace the hand-rolled conversion loop with ggml_cpu_fp16_to_fp32.	2026-09-18 17:17:38 +03:00
Alexey Kopytko	5b335f413e	ggml : check for allocation failures to prevent crashes (#28149 ) * ggml : check for allocation failures to prevent crashes * wording	2026-09-18 16:58:25 +03:00
Pascal	542348a35c	Model-Saver: Write the SWA pattern, 15 more architectures roundtrip (#29042 ) * llama: read the SWA pattern as a period or a per-layer array Add llama_model_base::load_swa_pattern(), which reads sliding_window_pattern either as one flag per layer or as a period expanded by set_swa_pattern(), and use it in every loader that reads the key as a period. These loaders silently ignored an array and applied their default period, although the converters of olmo2, gemma3n and exaone4 write arrays. The published GGUFs match the defaults, so their outputs do not change. The loaders that already accepted both forms lose their duplicated scalar-then-array block, and use their declared default period when the key is absent. * model-saver: write the SWA pattern and the MLA SWA geometry Write sliding_window_pattern as one flag per layer, nextn layers included, for every model using SWA. The array is never collapsed to a scalar, since the loaders read a scalar as a period. Also write the MLA key/value lengths and KV LoRA rank of the SWA layers, required by dots3note. This enables the saver for plamo3, gemma3, cohere2, cohere2moe, olmo2, exaone-moe, afmoe, mimo2, spark2_5, muse-glimmer, mellum, laguna, granite_swa, dots3note and maple, all passing the bit-exact roundtrip of test-llama-archs.	2026-09-18 15:20:03 +02:00
Aaron Teo	d663dd3f3a	ci: change ubuntu-latest to ubuntu-24.04 (#29079 )	2026-09-18 21:17:19 +08:00
Masashi Yoshimura	44be98f057	ggml-webgpu: fix supports_op condition for GET_ROWS (#28978 ) * fix get_rows vec4 handling * Add src strides checking to vec4_aligned of get_rows and the new test case.	2026-09-18 20:47:07 +09:00
z	911f6cdc8a	ggml : handle graph buffer reservation failure (#26070 )	2026-09-18 12:31:19 +03:00
Daniel Varga	bbd488c42a	vulkan: add IQ3_S MMQ matmul kernels (#28822 ) * vulkan: add IQ3_S MMQ matmul kernels * Make block_a_to_shmem do 2-byte loads (110 bytes is divisible by 2) * Align the check, IQ3_S is also using K tile size	2026-09-18 11:46:02 +03:00
Sait Furkan Teke	dc85f89c7e	vocab : add ufakzeka pre-tokenizer (#29033 ) * vocab : add ufakzeka pre-tokenizer * vocab : move ufakzeka to the models list and regenerate the hash mapping	2026-09-18 11:45:11 +03:00
Nikita Gordeev	8ed1a55efc	cmake : fix build when GGML_CPU=OFF and GGML_CUDA=ON (#29026 ) * fix: build fails when GGML_CPU=OFF and GGML_CUDA=ON * fix: eol in examples/convert-llama2c-to-ggml/CMakeLists.txt file	2026-09-18 11:44:03 +03:00
yanghongandSigbjørn Skjæret	bb11ebb682	gguf-py: fix Q8_1 block size in GGML_QUANT_SIZES (2+2+32) (#29036 ) * gguf-py: fix Q8_1 block size in GGML_QUANT_SIZES * --whitespace --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>	2026-09-18 11:43:06 +03:00
Sigbjørn Skjæret	f03cf3e9b8	ci : disable GHA cache for copilot (#29068 )	2026-09-18 10:20:53 +02:00
Sigbjørn Skjæret	bdcbaaf6e7	ci : bump android-actions/setup-android to 4.0.4 (#29065 )	2026-09-18 09:04:14 +02:00
drluoto	5c53396b89	vulkan: raise the hoisted row-id limit for mul_mat_id from 256 to 512 experts (#28501 ) * vulkan: raise the hoisted row-id limit for mul_mat_id to 512 experts The expert-count shader (count_experts.comp) sizes its shared arrays with BLOCK_SIZE, which is 256. Because of that, row-id hoisting is switched off for any model with more than 256 experts, and every mul_mat_id workgroup has to rescan the whole ids tensor on its own. Qwen3.8-Flash-Next has 512 experts and was quietly running on that slow path. This change sizes the arrays with a separate MAX_EXPERTS constant (512), clears them in a loop instead of one entry per thread, and raises the matching limit on the host side. On Strix Halo at batch 2048 the expert matmuls drop from 12.5 to 9.5 ms (iq3_s) and from 14.0 to 7.5 ms (iq4_nl) per op, and prompt processing gets about 19 % faster at 8k tokens. test-backend-ops MUL_MAT_ID passes (891/891) with new 512-expert test cases. Assisted-by: Claude Fable 5.1 * vulkan: raise the hoisted row-id limit for mul_mat_id to 1024 experts Follow-up to review feedback: 1024 matches LLAMA_MAX_EXPERTS instead of stopping at 512. The three shared arrays in count_experts.comp grow to 3 * 1024 * 4 = 12 KiB, which fits the 16 KiB that Vulkan guarantees for maxComputeSharedMemorySize. Adds mul_mat_id test cases at 1024 experts alongside the existing 512 ones. test-backend-ops MUL_MAT_ID passes on Vulkan (RADV, Strix Halo, Radeon 8060S): 889/889.	2026-09-18 09:00:15 +02:00
Sigbjørn Skjæret	972d2313bc	ci : add missing evict-old-files (#29041 )	2026-09-17 19:05:40 +02:00
Pedro Cuenca	c77ae695c9	rpc : skip ACCEL devices (#29020 )	2026-09-17 18:36:53 +03:00
David Friehs	b49650adb3	model : skip gate_up_exps if TENSOR_SKIP is set (#29014 ) required for qwen35moe if MTP tensors are fused but not loaded	2026-09-17 13:53:54 +02:00
Kartik Gulia	7076180486	model : extend Nemotron MTP support (#29018 ) * first fix * removed unnecessary declarations	2026-09-17 13:52:36 +02:00
Ravi PanchumarthyandMostafa Faheem	ebbb185227	openvino : Update OpenVINO to 2026.4;fix clangd,MSVC warnings; (#29009 ) * Update to openvino-2026.4 * Update OV docs * ggml-openvino : fix clangd and MSVC warnings * fix int to ptr cast, more internal linkage enforcement, and avoiding duplicate switch case --------- Co-authored-by: Mostafa Faheem <mostafaaafaheem@gmail.com>	2026-09-17 12:46:14 +02:00
Benjamin Babik Ben Babik Pascal	4ff829ec2e	ui: fix removed reasoning menu in single model mode on desktop (#27985 ) * ui: fix accidentally removed reasoning menu in single model mode on desktop * ui: formatting task run to fix storybook test * ui: mount the add menu reasoning submenu outside router mode only The models selector already owns the reasoning submenu in router mode, so the add menu only mounts it in single model mode. The first enabled item of the add menu is now the reasoning submenu, the accessibility story expects it. --------- Co-authored-by: Ben Babik <work@benjaminbabik.com> Co-authored-by: Pascal <admin@serveurperso.com>	2026-09-17 11:18:13 +02:00
Ruben Ortlam	f172be756a	vulkan: split buffers and debug code into separate files, add shared headers (#28732 )	2026-09-17 11:16:35 +03:00
Daniel Bevenius	87f9c82f2d	ci : add API/ABI check to make-release workflow [no ci] (#28947 ) * ci : add API/ABI check to make-release workflow [no ci] This commit adds an API/ABI compatibility check to the make-release workflow. The motivation for this to allow us to detect any potential breaking changes in API/ABI compatibility between releases and fail the the release if there are any. The workflow can be triggered manually as before and this check can be skipped if needed as it does take some time which might be useful when doing a dry-run and not specifically interested in the API/ABI check. By default this will check the current release against the latest release, but this can also be configured in the workflow, or in the script run on the command line, to check a different tag. * add check for minor version bumps [no ci] This commit also changes the build type to be RelWithDebInfo so that the reported information is more useful.	2026-09-17 09:55:16 +02:00
midagedevandSigbjørn Skjæret	7f6f0c2a9d	chat : add message delimiters to the DeepSeek V3.2/V4 parser (#29008 ) * chat : add message delimiters to the DeepSeek V3.2/V4 parser Assisted-by: Claude Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>	2026-09-17 09:41:08 +02:00
Yuri KhrustalevandJohannes Gäßler	81aeaeb74b	gguf : align the data section relative to the GGUF start, not the file (#28993 ) * gguf : align the data section relative to the GGUF start, not the file gguf_init_from_file_ptr reads a GGUF from the current file position, but padded the data section from file offset 0, so a GGUF embedded at an offset that is not a multiple of the alignment loaded without error and returned wrong tensor data. Also adds llama_adapter_lora_init_from_file_ptr, and disables mmap with a warning when an embedded data section is not aligned, instead of asserting in ggml. Assisted-by: Claude Opus 5 * llama : load lora from path through the FILE* variant The test now checks that mmap is disabled only for an unaligned offset. Assisted-by: Claude Fable 5.1 * Update ggml/src/gguf.cpp Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * Update include/llama.h Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * llama : error on unaligned mmap of an embedded GGUF, drop test-load-file-ptr --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2026-09-17 09:19:44 +02:00
Neo Zhang	c9a5eeeb34	sycl : fix the B70 mem allocate error when >19.3GB (#28953 )	2026-09-17 09:56:03 +03:00
Jiang, Fish	7490357f22	vulkan: skip unneeded MoE work in mul_mm coopmat1 path (#25483 )	2026-09-17 09:53:07 +03:00
Titaniumtown	817e5f83eb	sycl: ssm_conv: fuse the SiLU epilogue into the ssm_conv kernel (#28929 )	2026-09-17 09:51:50 +03:00
lhez	c57da6fd81	opencl: fix various warnings (#28984 ) * opencl: fix warnings * opencl: fix warnings for non adreno	2026-09-17 09:48:05 +03:00
Johannes Gäßler	79bfc1d43a	docs: remove JG as CODEOWNER for test-llama-archs (#29003 )	2026-09-17 08:14:58 +02:00
Abir Deol	05f2dcfdba	vulkan: fix buffer_reference alignment in im2col shaders (#28996 ) Both im2col.comp and im2col_3d.comp declare D_ptr without an explicit buffer_reference_align, so glslang emits writes through it as Aligned 16. The shaders advance the pointer by D_SIZE, a per-variant define set to 4 for float and 2 for float16_t, so most write addresses are not 16-byte aligned. This triggers VUID-RuntimeSpirv-PhysicalStorageBuffer64-06315 under GPU-AV. Declaring buffer_reference_align = D_SIZE matches the alignment to the actual write stride and takes validation hits from 20 to 0 for both IM2COL and IM2COL_3D. Fixes #28960	2026-09-17 07:14:33 +02:00
Ruben Ortlam	35822afe58	vulkan: support qwen4exp hc ops (#28988 ) * vulkan: support qwen4exp hc ops * fix stale comment [no-ci]	2026-09-17 06:34:23 +02:00
Michael Taylor	aa39d7a3e1	[SYCL] Fix function signature for `ggml_backend_sycl_split_buffer_type` (#28981 )	2026-09-16 22:15:47 -04:00
Jeff Bolz	4bc272fd72	vulkan: work around NV bug with argsort_large.comp (#28975 )	2026-09-16 18:40:07 -05:00
David Friehs	fb27a525d2	TP: fix split state and granularity for fused QKV gemma4, qwen35 (#28965 ) * model: calculate split states for attn_qkv from n_head * n_embd_head_k required for gemma4 with --fuse-qkv, where n_embd is 5376 but Q is 8192. * model: handle fused full attention layers for qwen35/qwen35moe * model: add TODO: [TAG_SPLIT_QGATE_QWEN]	2026-09-16 22:02:12 +03:00
Eve	c6824a9e42	ci: switch fast jobs back to github (#28959 ) * switch jobs to ubuntu-slim * ubuntu slim almost takes 15 minutes for check requirements so use something faster	2026-09-16 16:50:26 +00:00
Gaurav Garg	2f3fd02526	Enable CUDA graph for MTP draft (#28549 ) * Improve CUDA graph usage for MTP * Rename field * Address review feedback	2026-09-16 21:46:54 +05:30