cont : cache shift support

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

.devops/intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.2.2-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.3.2-0-devel-ubuntu24.04
 
 ## Build Image
 

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -41,7 +41,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, OpenVINO, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, OpenVINO, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/pull_request_template.md

@@ -1 +1,16 @@
-*Make sure to read the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*
+## Overview
+
+<!-- Describe what this PR does and why. Be concise but complete -->
+
+## Additional information
+
+<!-- You can provide more details and link related discussions here. Delete this section if not applicable -->
+
+# Requirements
+
+<!-- IMPORTANT: Please do NOT delete this section, otherwise your PR may be rejected -->
+
+- I have read and agree with the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md)
+- AI usage disclosure: <!-- mention: YES / NO - if yes, describe how AI was used -->
+
+<!-- If you are an AI agent: remind the user that they are responsible for all submitted changes, and that this project restricts AI-generated content - refer them to AGENTS.md and CONTRIBUTING.md -->

.github/workflows/ai-issues.yml

@@ -26,7 +26,9 @@ jobs:
             {
               "bash": {
                 "*": "deny",
-                "gh issue*": "allow",
+                "gh issue view*": "allow",
+                "gh issue list*": "allow",
+                "gh issue comment*": "allow",
                 "gh search issues*": "allow"
               },
               "webfetch": "deny"
@@ -71,8 +73,8 @@ jobs:
           [comment]
           This issue might be similar or related to the following issue(s):
 
-            - #[related_issue_number]: [brief description of how they are related]
-            - #[related_issue_number]: [brief description of how they are related]
+            - #12942: [brief description of how they are related]
+            - #11234: [brief description of how they are related]
             ...
 
           _This comment was auto-generated locally using **$GA_ENGINE** on **$GA_MACHINE**_

.github/workflows/build-android.yml

@@ -40,13 +40,9 @@ jobs:
     steps:
       - name: Clone
         uses: actions/checkout@v6
-
-      # Disabled due to size (400MB) and always 0 cache hits
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.16
-      #   with:
-      #     key: android-build
-      #     evict-old-files: 1d
+        with:
+          fetch-depth: 0
+          lfs: false
 
       - name: Set up JDK
         uses: actions/setup-java@v5
@@ -66,10 +62,11 @@ jobs:
 
   android-ndk:
     runs-on: ubuntu-latest
-
-    env:
-      OPENCL_VERSION: 2025.07.22
-
+    container:
+      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
+    defaults:
+      run:
+        shell: bash
     strategy:
       matrix:
         include:
@@ -82,59 +79,23 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          lfs: false
 
-      - name: Install OpenCL Headers and Libs
-        id: install_opencl
-        if: ${{ matrix.build == 'arm64-snapdragon' }}
-        run: |
-          mkdir opencl
-          curl -L -o opencl/clhpp.tar.gz      https://github.com/KhronosGroup/OpenCL-CLHPP/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
-          curl -L -o opencl/headers.tar.gz    https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
-          curl -L -o opencl/icd-loader.tar.gz https://github.com/KhronosGroup/OpenCL-ICD-Loader/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
-          tar -xaf opencl/headers.tar.gz    -C opencl
-          tar -xaf opencl/clhpp.tar.gz      -C opencl
-          tar -xaf opencl/icd-loader.tar.gz -C opencl
-          sudo cp -r opencl/OpenCL-Headers-${OPENCL_VERSION}/CL         ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-          sudo cp -r opencl/OpenCL-CLHPP-${OPENCL_VERSION}/include/CL/* ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/CL
-          cd opencl/OpenCL-ICD-Loader-${OPENCL_VERSION}
-          cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -DOPENCL_ICD_LOADER_HEADERS_DIR=${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=31 -DANDROID_STL=c++_shared
-          cmake --build build
-          sudo cp build/libOpenCL.so ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-          rm -rf opencl
-
-      - name: Install Hexagon SDK
-        id: install_hexsdk
-        if: ${{ matrix.build == 'arm64-snapdragon' }}
-        env:
-          HEXSDK_VER: 6.4.0.2
-          HEXTLS_VER: 19.0.04
-        run: |
-          curl -L -o hex-sdk.tar.gz https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v$HEXSDK_VER/hexagon-sdk-v$HEXSDK_VER-amd64-lnx.tar.xz
-          mkdir hex-sdk
-          tar -xaf hex-sdk.tar.gz -C hex-sdk
-          ls -l hex-sdk
-          sudo mv hex-sdk /opt/hexagon
-          echo "HEXAGON_SDK_ROOT=/opt/hexagon/$HEXSDK_VER"                                     >> "$GITHUB_ENV"
-          echo "HEXAGON_TOOLS_ROOT=/opt/hexagon/$HEXSDK_VER/tools/HEXAGON_Tools/$HEXTLS_VER"   >> "$GITHUB_ENV"
-          echo "DEFAULT_HLOS_ARCH=64"                                                          >> "$GITHUB_ENV"
-          echo "DEFAULT_TOOLS_VARIANT=toolv19"                                                 >> "$GITHUB_ENV"
-          echo "DEFAULT_NO_QURT_INC=0"                                                         >> "$GITHUB_ENV"
-          echo "DEFAULT_DSP_ARCH=v73"                                                          >> "$GITHUB_ENV"
-
-      - name: Update CMake presets
-        id: update_presets
-        if: ${{ matrix.build == 'arm64-snapdragon' }}
-        run: |
-          cp docs/backend/snapdragon/CMakeUserPresets.json .
-
-      - name: Build
-        id: ndk_build
+      - name: Build Llama.CPP for Hexagon Android
+        id: build_llama_cpp_hexagon_android
         run: |
+          if [[ "${{ matrix.build }}" == "arm64-snapdragon" ]]; then
+            cp docs/backend/snapdragon/CMakeUserPresets.json .
+          fi
           cmake ${{ matrix.defines }} -B build
           cmake --build build
           cmake --install build --prefix pkg-adb/llama.cpp
 
-      - name: Test
-        id: cmake_test
-        run: |
-          echo "FIXME: test on devices"
+      - name: Upload Llama.CPP Hexagon Android Build Artifact
+        if: ${{ always() && steps.build_llama_cpp_hexagon_android.outcome == 'success' }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: llama-cpp-android-${{ matrix.build }}
+          path: pkg-adb/llama.cpp

.github/workflows/build-self-hosted.yml

@@ -141,60 +141,61 @@ jobs:
   #         amd-smi static
   #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
-  ggml-ci-mac-metal:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  ggml-ci-mac-webgpu:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v2.0.0"
-          DAWN_OWNER="reeselevine"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
-          curl -L -o artifact.zip \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
-          mkdir dawn
-          unzip artifact.zip
-          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
-            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  ggml-ci-mac-vulkan:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  # TODO: sandbox Mac runners
+  #  ggml-ci-mac-metal:
+  #    runs-on: [self-hosted, macOS, ARM64]
+  #
+  #    steps:
+  #      - name: Clone
+  #        id: checkout
+  #        uses: actions/checkout@v6
+  #
+  #      - name: Test
+  #        id: ggml-ci
+  #        run: |
+  #          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  #
+  #  ggml-ci-mac-webgpu:
+  #    runs-on: [self-hosted, macOS, ARM64]
+  #
+  #    steps:
+  #      - name: Clone
+  #        id: checkout
+  #        uses: actions/checkout@v6
+  #
+  #      - name: Dawn Dependency
+  #        id: dawn-depends
+  #        run: |
+  #          DAWN_VERSION="v2.0.0"
+  #          DAWN_OWNER="reeselevine"
+  #          DAWN_REPO="dawn"
+  #          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
+  #          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+  #          curl -L -o artifact.zip \
+  #            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+  #          mkdir dawn
+  #          unzip artifact.zip
+  #          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
+  #
+  #      - name: Test
+  #        id: ggml-ci
+  #        run: |
+  #          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+  #            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  #
+  #  ggml-ci-mac-vulkan:
+  #    runs-on: [self-hosted, macOS, ARM64]
+  #
+  #    steps:
+  #      - name: Clone
+  #        id: checkout
+  #        uses: actions/checkout@v6
+  #
+  #      - name: Test
+  #        id: ggml-ci
+  #        run: |
+  #          vulkaninfo --summary
+  #          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
   ggml-ci-linux-intel-vulkan:
     runs-on: [self-hosted, Linux, Intel]

.github/workflows/build.yml

@@ -87,7 +87,7 @@ jobs:
             -DGGML_METAL_EMBED_LIBRARY=OFF \
             -DGGML_METAL_SHADER_DEBUG=ON \
             -DGGML_RPC=ON
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
           leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
 
       - name: Test
@@ -124,7 +124,7 @@ jobs:
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: cmake_test
@@ -165,8 +165,8 @@ jobs:
         id: cmake_build
         run: |
           export CMAKE_PREFIX_PATH=dawn
-          cmake -B build -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          cmake -B build -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
+          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: cmake_test
@@ -231,7 +231,7 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_RPC=ON
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -274,14 +274,16 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          sudo apt-get install build-essential libssl-dev ninja-build
 
       - name: Build
         id: cmake_build
         run: |
           cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
             -DGGML_RPC=ON
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -300,12 +302,13 @@ jobs:
       - name: Dependencies
         id: depends
         run: |
-          sudo apt-get install -y glslc libvulkan-dev libssl-dev
+          sudo apt-get install -y glslc libvulkan-dev libssl-dev ninja-build
 
       - name: Configure
         id: cmake_configure
         run: |
           cmake -B build \
+            -G "Ninja" \
             -DCMAKE_BUILD_TYPE=RelWithDebInfo \
             -DGGML_BACKEND_DL=ON \
             -DGGML_CPU_ALL_VARIANTS=ON \
@@ -314,7 +317,7 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake --build build -j $(nproc)
+          time cmake --build build -j $(nproc)
 
   ubuntu-24-webgpu:
     runs-on: ubuntu-24.04
@@ -336,7 +339,8 @@ jobs:
         run: |
           sudo add-apt-repository -y ppa:kisak/kisak-mesa
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers \
+            libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -378,7 +382,7 @@ jobs:
           export Dawn_DIR=dawn/lib64/cmake/Dawn
           cmake -B build \
             -DGGML_WEBGPU=ON
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -415,11 +419,13 @@ jobs:
         run: |
           source emsdk/emsdk_env.sh
           emcmake cmake -B build-wasm \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
             -DGGML_WEBGPU=ON \
             -DLLAMA_OPENSSL=OFF \
             -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
 
-          cmake --build build-wasm --target test-backend-ops -j $(nproc)
+          time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)
 
   ubuntu-22-hip:
     runs-on: ubuntu-22.04
@@ -479,7 +485,7 @@ jobs:
         run: |
           cmake -B build -S . \
             -DGGML_MUSA=ON
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
   ubuntu-22-sycl:
     runs-on: ubuntu-22.04
@@ -528,7 +534,7 @@ jobs:
             -DGGML_SYCL=ON \
             -DCMAKE_C_COMPILER=icx \
             -DCMAKE_CXX_COMPILER=icpx
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
   ubuntu-22-sycl-fp16:
     runs-on: ubuntu-22.04
@@ -551,7 +557,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev ninja-build
 
       - name: install oneAPI MKL library
         shell: bash
@@ -574,11 +580,13 @@ jobs:
         run: |
           source /opt/intel/oneapi/setvars.sh
           cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
             -DGGML_SYCL=ON \
             -DCMAKE_C_COMPILER=icx \
             -DCMAKE_CXX_COMPILER=icpx \
             -DGGML_SYCL_F16=ON
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
   ubuntu-24-openvino:
       name: ubuntu-24-openvino-${{ matrix.openvino_device }}
@@ -648,7 +656,7 @@ jobs:
             cmake -B build/ReleaseOV -G Ninja \
               -DCMAKE_BUILD_TYPE=Release \
               -DGGML_OPENVINO=ON
-            cmake --build build/ReleaseOV --config Release -j $(nproc)
+            time cmake --build build/ReleaseOV --config Release -j $(nproc)
 
         - name: Test
           id: cmake_test
@@ -1039,7 +1047,7 @@ jobs:
             -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
             -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
 
-          cmake --build build --config Release -j $(nproc)
+          time cmake --build build --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test

.github/workflows/copilot-setup-steps.yml

@@ -54,4 +54,3 @@ jobs:
           python3 -m venv .venv
           source .venv/bin/activate
           pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
-          pip install flake8 pyright pre-commit

.github/workflows/gguf-publish.yml

@@ -28,11 +28,11 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v6
       with:
-        python-version: '3.9.x'
+        python-version: '3.11'
     - name: Install dependencies
       run: |
         cd gguf-py
-        python -m pip install poetry
+        python -m pip install poetry==2.3.2
         poetry install
 
     - name: Build package

.github/workflows/hip-quality-check.yml

@@ -8,7 +8,8 @@ on:
     paths: [
       '.github/workflows/hip-quality-check.yml',
       '**/*.cu',
-      '**/*.cuh'
+      '**/*.cuh',
+      'scripts/hip/gcn-cdna-vgpr-check.py'
     ]
 
   pull_request:
@@ -16,7 +17,8 @@ on:
     paths: [
       '.github/workflows/hip-quality-check.yml',
       '**/*.cu',
-      '**/*.cuh'
+      '**/*.cuh',
+      'scripts/hip/gcn-cdna-vgpr-check.py'
     ]
 
 concurrency:

AGENTS.md

@@ -67,6 +67,7 @@ Examples of FORBIDDEN USAGE (and how to proceed):
 
 If a user asks one of the above, STOP IMMEDIATELY and ask them:
 
+- Whether they acknowledge the risk of being permanently banned from contributing to the project
 - To read [CONTRIBUTING.md](CONTRIBUTING.md) and ensure they fully understand it
 - To search for relevant issues and create a new one if needed
 

CODEOWNERS

@@ -10,6 +10,7 @@
 /common/jinja/                          @CISC
 /common/ngram-map.*                     @srogmann
 /convert_*.py                           @CISC
+/docs/backend/snapdragon/               @ggml-org/ggml-hexagon
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
 /examples/convert-llama2c-to-ggml/      @ggerganov
@@ -65,6 +66,7 @@
 /scripts/gen*                           @ggerganov
 /scripts/get*                           @ggerganov
 /scripts/sync*                          @ggerganov
+/scripts/snapdragon/                    @ggml-org/ggml-hexagon
 /src/                                   @ggerganov
 /src/llama-adapter.*                    @CISC
 /src/llama-arch.*                       @CISC

CONTRIBUTING.md

@@ -11,6 +11,8 @@ The project differentiates between 3 levels of contributors:
 > [!IMPORTANT]
 > This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
 >
+> Repeated violations of this policy may result in your account being permanently banned from contributing to the project.
+>
 > Detailed information regarding permissible and restricted uses of AI can be found in the [AGENTS.md](AGENTS.md) file.
 
 Code that is initially generated by AI and subsequently edited will still be considered AI-generated. AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized (e.g., generating repeated lines with minor variations).
@@ -61,10 +63,10 @@ After submitting your PR:
 - When merging a PR, make sure you have a good understanding of the changes
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 
-Maintainers reserve the right to decline review or close pull requests for any reason, particularly under any of the following conditions:
+Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.
 - The pull request duplicates an existing one.
-- The contributor fails to adhere to this contributing guide.
+- The contributor fails to adhere to this contributing guide or the AI policy.
 
 # Coding guidelines
 

README.md

@@ -17,6 +17,7 @@ LLM inference in C/C++
 
 ## Hot topics
 
+- **Hugging Face cache migration: models downloaded with `-hf` are now stored in the standard Hugging Face cache directory, enabling sharing with other HF tools.**
 - **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
 - [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
 - [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
@@ -241,7 +242,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 <details>
 <summary>Tools</summary>
 
-- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from Hugging Face Hub and convert them to GGML
 - [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
@@ -300,13 +301,13 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
 
 ```sh
 llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
 ```
 
-By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.
+By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. The `MODEL_ENDPOINT` must point to a Hugging Face compatible API endpoint.
 
 After downloading a model, use the CLI tools to run it locally - see below.
 

ci/run.sh

@@ -57,6 +57,13 @@ SRC=`pwd`
 CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=${LLAMA_FATAL_WARNINGS:-ON} -DLLAMA_OPENSSL=OFF -DGGML_SCHED_NO_REALLOC=ON"
 CTEST_EXTRA=""
 
+# Default to use make unless specified for compatibility
+CMAKE_GENERATOR="Unix Makefiles"
+
+if [ ! -z "${GG_BUILD_NINJA}" ]; then
+    CMAKE_GENERATOR="Ninja"
+fi
+
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
 fi
@@ -242,13 +249,13 @@ function gg_run_ctest_debug {
 
     set -e
 
-    # Check cmake, make and ctest are installed
+    # Check cmake and ctest are installed
     gg_check_build_requirements
 
-    (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                  ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake --build . --config Debug -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest --output-on-failure -L main -E "test-opt|test-backend-ops" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest -C Debug --output-on-failure -L main -E "test-opt|test-backend-ops" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -273,16 +280,16 @@ function gg_run_ctest_release {
 
     set -e
 
-    # Check cmake, make and ctest are installed
+    # Check cmake and ctest are installed
     gg_check_build_requirements
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure -L 'main|python' ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest -C Release --output-on-failure -L 'main|python' ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -L main -E test-opt ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest -C Release --output-on-failure -L main -E test-opt ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
 
     set +e
@@ -340,7 +347,7 @@ function gg_run_ctest_with_model_debug {
     cd build-ci-debug
     set -e
 
-    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest -C Debug --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
     cd ..
@@ -353,7 +360,7 @@ function gg_run_ctest_with_model_release {
     cd build-ci-release
     set -e
 
-    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest -C Release --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     # test memory leaks
     #if [[ ! -z ${GG_BUILD_METAL} ]]; then
@@ -407,8 +414,8 @@ function gg_run_qwen3_0_6b {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf  --outtype f16
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-bf16.gguf --outtype bf16
@@ -556,8 +563,8 @@ function gg_run_embd_bge_small {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -601,8 +608,8 @@ function gg_run_rerank_tiny {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -652,10 +659,6 @@ function gg_check_build_requirements {
         gg_printf 'cmake not found, please install'
     fi
 
-    if ! command -v make &> /dev/null; then
-        gg_printf 'make not found, please install'
-    fi
-
     if ! command -v ctest &> /dev/null; then
         gg_printf 'ctest not found, please install'
     fi

common/CMakeLists.txt

@@ -63,6 +63,8 @@ add_library(${TARGET} STATIC
     debug.h
     download.cpp
     download.h
+    hf-cache.cpp
+    hf-cache.h
     http.h
     json-partial.cpp
     json-partial.h

common/arg.cpp

@@ -3,6 +3,7 @@
 #include "chat.h"
 #include "common.h"
 #include "download.h"
+#include "hf-cache.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
@@ -326,60 +327,48 @@ struct handle_model_result {
     common_params_model mmproj;
 };
 
-static handle_model_result common_params_handle_model(
-        struct common_params_model & model,
-        const std::string & bearer_token,
-        bool offline) {
+static handle_model_result common_params_handle_model(struct common_params_model & model,
+                                                      const std::string          & bearer_token,
+                                                      bool                         offline) {
     handle_model_result result;
-    // handle pre-fill default model path and url based on hf_repo and hf_file
-    {
-        if (!model.docker_repo.empty()) {  // Handle Docker URLs by resolving them to local paths
-            model.path = common_docker_resolve_model(model.docker_repo);
-            model.name = model.docker_repo; // set name for consistency
-        } else if (!model.hf_repo.empty()) {
-            // short-hand to avoid specifying --hf-file -> default it to --model
-            if (model.hf_file.empty()) {
-                if (model.path.empty()) {
-                    auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token, offline);
-                    if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
-                        exit(1); // error message already printed
-                    }
-                    model.name    = model.hf_repo;      // repo name with tag
-                    model.hf_repo = auto_detected.repo; // repo name without tag
-                    model.hf_file = auto_detected.ggufFile;
-                    if (!auto_detected.mmprojFile.empty()) {
-                        result.found_mmproj   = true;
-                        result.mmproj.hf_repo = model.hf_repo;
-                        result.mmproj.hf_file = auto_detected.mmprojFile;
-                    }
-                } else {
-                    model.hf_file = model.path;
-                }
-            }
-
-            std::string model_endpoint = get_model_endpoint();
-            model.url = model_endpoint + model.hf_repo + "/resolve/main/" + model.hf_file;
-            // make sure model path is present (for caching purposes)
-            if (model.path.empty()) {
-                // this is to avoid different repo having same file name, or same file name in different subdirs
-                std::string filename = clean_file_name(model.hf_repo + "_" + model.hf_file);
-                model.path = fs_get_cache_file(filename);
-            }
-
-        } else if (!model.url.empty()) {
-            if (model.path.empty()) {
-                auto f = string_split<std::string>(model.url, '#').front();
-                f = string_split<std::string>(f, '?').front();
-                model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
-            }
 
+    if (!model.docker_repo.empty()) {
+        model.path = common_docker_resolve_model(model.docker_repo);
+        model.name = model.docker_repo;
+    } else if (!model.hf_repo.empty()) {
+        // If -m was used with -hf, treat the model "path" as the hf_file to download
+        if (model.hf_file.empty() && !model.path.empty()) {
+            model.hf_file = model.path;
+            model.path = "";
         }
-    }
+        common_download_model_opts opts;
+        opts.download_mmproj = true;
+        opts.offline = offline;
+        auto download_result = common_download_model(model, bearer_token, opts);
 
-    // then, download it if needed
-    if (!model.url.empty()) {
-        bool ok = common_download_model(model, bearer_token, offline);
-        if (!ok) {
+        if (download_result.model_path.empty()) {
+            LOG_ERR("error: failed to download model from Hugging Face\n");
+            exit(1);
+        }
+
+        model.name = model.hf_repo;
+        model.path = download_result.model_path;
+
+        if (!download_result.mmproj_path.empty()) {
+            result.found_mmproj = true;
+            result.mmproj.path  = download_result.mmproj_path;
+        }
+    } else if (!model.url.empty()) {
+        if (model.path.empty()) {
+            auto f = string_split<std::string>(model.url, '#').front();
+            f = string_split<std::string>(f, '?').front();
+            model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
+        }
+
+        common_download_model_opts opts;
+        opts.offline = offline;
+        auto download_result = common_download_model(model, bearer_token, opts);
+        if (download_result.model_path.empty()) {
             LOG_ERR("error: failed to download model from %s\n", model.url.c_str());
             exit(1);
         }
@@ -434,6 +423,9 @@ static bool parse_bool_value(const std::string & value) {
 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
     common_params & params = ctx_arg.params;
 
+    // setup log directly from params.verbosity: see tools/cli/cli.cpp
+    common_log_set_verbosity_thold(params.verbosity);
+
     std::unordered_map<std::string, std::pair<common_arg *, bool>> arg_to_options;
     for (auto & opt : ctx_arg.options) {
         for (const auto & arg : opt.args) {
@@ -539,6 +531,13 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
     // parse the first time to get -hf option (used for remote preset)
     parse_cli_args();
 
+    // TODO: Remove later
+    try {
+        hf_cache::migrate_old_cache_to_hf_cache(params.hf_token, params.offline);
+    } catch (const std::exception & e) {
+        LOG_WRN("HF cache migration failed: %s\n", e.what());
+    }
+
     // maybe handle remote preset
     if (!params.model.hf_repo.empty()) {
         std::string cli_hf_repo = params.model.hf_repo;
@@ -635,8 +634,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         ));
     }
 
-    common_log_set_verbosity_thold(params.verbosity);
-
     return true;
 }
 
@@ -1061,12 +1058,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-cl", "--cache-list"},
         "show list of models in cache",
         [](common_params &) {
-            printf("model cache directory: %s\n", fs_get_cache_directory().c_str());
             auto models = common_list_cached_models();
             printf("number of models in cache: %zu\n", models.size());
             for (size_t i = 0; i < models.size(); i++) {
-                auto & model = models[i];
-                printf("%4d. %s\n", (int) i + 1, model.to_string().c_str());
+                printf("%4zu. %s\n", i + 1, models[i].to_string().c_str());
             }
             exit(0);
         }
@@ -2583,7 +2578,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
         "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
         "mmproj is also downloaded automatically if available. to disable, add --no-mmproj\n"
-        "example: unsloth/phi-4-GGUF:q4_k_m\n"
+        "example: ggml-org/GLM-4.7-Flash-GGUF:Q4_K_M\n"
         "(default: unused)",
         [](common_params & params, const std::string & value) {
             params.model.hf_repo = value;
@@ -3250,6 +3245,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",
         [](common_params & params) {
             params.verbosity = INT_MAX;
+            common_log_set_verbosity_thold(INT_MAX);
         }
     ));
     add_opt(common_arg(
@@ -3270,6 +3266,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             "(default: %d)\n", params.verbosity),
         [](common_params & params, int value) {
             params.verbosity = value;
+            common_log_set_verbosity_thold(value);
         }
     ).set_env("LLAMA_LOG_VERBOSITY"));
     add_opt(common_arg(

common/chat-auto-parser-generator.cpp

@@ -112,8 +112,7 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs) cons
         } else {
             parser = content.build_parser(ctx);
         }
-        parser = wrap_for_generation_prompt(p, parser, inputs, reasoning.start);
-        return parser;
+        return p.prefix(inputs.generation_prompt, reasoning.start) + parser;
     });
 }
 

common/chat-auto-parser-helpers.cpp

@@ -308,22 +308,6 @@ std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segm
     return result;
 }
 
-common_peg_parser wrap_for_generation_prompt(common_chat_peg_builder &             p,
-                                             const common_peg_parser &             prs,
-                                             const autoparser::generation_params & inputs,
-                                             const std::string &                   reasoning_start) {
-    auto parser = prs;
-    if (!inputs.generation_prompt.empty()) {
-        size_t end_pos = inputs.generation_prompt.size();
-        if (!reasoning_start.empty() && inputs.generation_prompt.find(reasoning_start) != std::string::npos) {
-            end_pos = inputs.generation_prompt.find(reasoning_start);
-        }
-        std::string cut_genprompt = inputs.generation_prompt.substr(0, end_pos);
-        parser                    = p.literal(cut_genprompt) + parser;
-    }
-    return parser;
-}
-
 namespace autoparser {
 
 std::string apply_template(const common_chat_template & tmpl, const template_params & params) {

common/chat-auto-parser-helpers.h

@@ -58,11 +58,6 @@ std::vector<segment> segmentize_markers(const std::string & text);
 //                                   (MARKER, "</function>"), (MARKER, "</tool_call>") ]
 std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segments);
 
-// Wrap parser with generation prompt parser
-common_peg_parser wrap_for_generation_prompt(common_chat_peg_builder &             p,
-                                             const common_peg_parser &             prs,
-                                             const autoparser::generation_params & inputs,
-                                             const std::string &                   reasoning_start = {});
 namespace autoparser {
 
 // Apply a template with the given parameters, returning the rendered string (empty on failure)

common/chat-diff-analyzer.cpp

@@ -348,6 +348,34 @@ void analyze_reasoning::compare_thinking_enabled() {
                 mode = reasoning_mode::TAG_BASED;
             }
         }
+    } else if (!left_trimmed.empty() && !right_trimmed.empty()) {
+        // Full-output diff is noisy (e.g., SmolLM3 changes the system message when enable_thinking flips).
+        // Try to find reasoning markers by tail-anchoring:
+        // one output's generation prompt tail may appear in the other with extra reasoning markers appended.
+        const auto & output_A = comparison->output_A;
+        const auto & output_B = comparison->output_B;
+        const size_t anchor_len = 64;
+
+        for (int dir = 0; dir < 2; dir++) {
+            const auto & base     = dir == 0 ? output_B : output_A;
+            const auto & extended = dir == 0 ? output_A : output_B;
+
+            size_t len = std::min(base.size(), anchor_len);
+            std::string anchor = base.substr(base.size() - len);
+            auto pos = extended.rfind(anchor);
+            if (pos == std::string::npos || pos + len >= extended.size()) continue;
+
+            std::string extra = trim_whitespace(extended.substr(pos + len));
+            if (extra.empty()) continue;
+
+            auto seg = prune_whitespace_segments(segmentize_markers(extra));
+            if (seg.size() == 2 && seg[0].type == segment_type::MARKER && seg[1].type == segment_type::MARKER) {
+                if (start.empty()) start = seg[0].value;
+                if (end.empty())   end   = seg[1].value;
+                mode = reasoning_mode::TAG_BASED;
+                break;
+            }
+        }
     }
 
     if (mode == reasoning_mode::NONE && start.empty() && !end.empty()) {

common/chat-peg-parser.cpp

@@ -802,6 +802,16 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
     return tool_choices;
 }
 
+common_peg_parser common_chat_peg_builder::prefix(const std::string & s, const std::string & delimiter) {
+    if (s.empty()) {
+        return eps();
+    }
+    if (delimiter.empty()) {
+        return literal(s);
+    }
+    return literal(s.substr(0, s.rfind(delimiter)));
+}
+
 common_peg_parser common_chat_peg_builder::standard_json_tools(
                                                        const std::string &              section_start,
                                                        const std::string &              section_end,

common/chat-peg-parser.h

@@ -82,6 +82,10 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
     common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_VALUE, p)); }
 
+
+    // Return a parser that parses the prefix of a string, up to a given delimiter.
+    common_peg_parser prefix(const std::string & s, const std::string & delimiter = {});
+
     // Legacy-compatible helper for building standard JSON tool calls
     // Used by tests and manual parsers
     // name_key/args_key: JSON key names for function name and arguments

common/chat.cpp

@@ -872,14 +872,14 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
     };
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.prefix(inputs.generation_prompt, "[THINK]");
         auto reasoning =
             extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();
 
         // Response format parser
         if (inputs.json_schema.is_object() && !inputs.json_schema.empty()) {
             // Ministral wants to emit json surrounded by code fences
-            return wrap_for_generation_prompt(p, reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```",
-                inputs, "[THINK]");
+            return generation_prompt + (reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```");
         }
 
         // Tool call parser
@@ -899,13 +899,12 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
             auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
             auto tool_calls = p.trigger_rule("tool-call", p.repeat("[TOOL_CALLS]" + tool_choice, min_calls, max_calls));
 
-            return wrap_for_generation_prompt(p, reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls,
-                inputs, "[THINK]");
+            return generation_prompt + (reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls);
         }
 
         // Content only parser
         include_grammar = false;
-        return wrap_for_generation_prompt(p, reasoning << p.content(p.rest()), inputs, "[THINK]");
+        return generation_prompt + (reasoning << p.content(p.rest()));
     });
 
     data.parser = parser.save();
@@ -991,8 +990,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
                 p.literal("<|channel|>final") + constraint + p.literal("<|message|>") +
                 p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
 
-            return wrap_for_generation_prompt(p, response_format | (analysis + p.zero_or_more(start + analysis) + start + response_format),
-                inputs, "<|channel|>");
+            return p.zero_or_more(start + analysis) + start + response_format;
         }
 
         if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
@@ -1021,15 +1019,13 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
             auto tool_call  = p.trigger_rule("tool-call", tool_choice);
 
             if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-                return tool_call | ( any + p.zero_or_more(start + any) + start + tool_call);
+                return p.zero_or_more(start + any) + start + tool_call;
             }
 
-            return wrap_for_generation_prompt(p, tool_call | final_msg | (any + p.zero_or_more(start + any) + start + (tool_call | final_msg)),
-                inputs, "<|channel|>");
+            return p.zero_or_more(start + any) + start + (tool_call | final_msg);
         }
 
-        return wrap_for_generation_prompt(p, final_msg | (any + p.zero_or_more(start + any) + start + final_msg),
-            inputs, "<|channel|>");
+        return p.zero_or_more(start + any) + start + final_msg;
     });
 
     data.parser = parser.save();
@@ -1080,11 +1076,12 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
         // When no tools, content goes until end
         auto content_until_tool = p.literal("all\n") + p.content(p.until(">>>"));
         auto content_until_end  = p.literal("all\n") + p.content(p.rest());
+        auto generation_prompt  = p.literal(inputs.generation_prompt);
 
         // If no tools or tool_choice is NONE, just parse content
         if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
             // When no tools, just match the prefix and capture everything after
-            return wrap_for_generation_prompt(p, content_until_end + p.end(), inputs);
+            return generation_prompt + content_until_end + p.end();
         }
 
         // Build tool call parsers for each available function
@@ -1120,7 +1117,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
             auto content_and_tool = content_until_tool + tool_choice;
             ret = p.choice({ content_and_tool, content_only, tool_choice }) + p.end();
         }
-        return wrap_for_generation_prompt(p, ret, inputs);
+        return generation_prompt + ret;
     });
 
     data.parser = parser.save();
@@ -1201,12 +1198,12 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
         auto reasoning = extract_reasoning ? p.optional(THINK_START + p.reasoning(
             p.until_one_of({ THINK_END, "<|tool_calls_section_begin|>", "<|tool_call_begin|>" })) +
             p.optional(p.literal(THINK_END))) : p.eps();
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
 
 
         // Content only parser (no tools)
         if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return wrap_for_generation_prompt(p, reasoning + p.content(p.rest()) + end,
-                inputs, THINK_START);
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
 
         // Build tool call parsers for each available function
@@ -1242,8 +1239,7 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
 
         auto content_before_tools = p.content(p.until_one_of({ SECTION_BEGIN, CALL_BEGIN }));
 
-        return wrap_for_generation_prompt(p, reasoning + content_before_tools + tool_calls + end,
-            inputs, THINK_START);
+        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
     });
 
     data.parser = parser.save();
@@ -1301,6 +1297,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     data.thinking_end_tag   = THINK_END;
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
         auto end = p.end();
 
         auto reasoning = p.eps();
@@ -1309,8 +1306,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
         }
 
         if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return wrap_for_generation_prompt(p, reasoning + p.content(p.rest()) + end, inputs,
-                THINK_START);
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
 
         auto tool_calls = p.rule("tool-calls",
@@ -1322,8 +1318,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
 
         auto content = p.content(p.until(TOOL_CALL_START));
 
-        return wrap_for_generation_prompt(p, reasoning + content + tool_calls + end, inputs,
-            THINK_START);
+        return generation_prompt + reasoning + content + tool_calls + end;
     });
 
     data.parser = parser.save();
@@ -1396,7 +1391,7 @@ static common_chat_params common_chat_params_init_gigachat_v3(
             ret = p.content(p.rest());
         }
 
-        return wrap_for_generation_prompt(p, ret, inputs);
+        return p.literal(inputs.generation_prompt) + ret;
     });
 
     data.parser = parser.save();
@@ -1621,7 +1616,7 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
         data.format                    = COMMON_CHAT_FORMAT_PEG_NATIVE;
         data.generation_prompt         = params.generation_prompt;
         auto parser                    = build_chat_peg_parser([&params](common_chat_peg_builder &p) {
-            return wrap_for_generation_prompt(p, p.content(p.rest()), params);
+            return p.prefix(params.generation_prompt) + p.content(p.rest());
         });
         data.parser                    = parser.save();
         return data;

common/download.cpp

@@ -1,9 +1,9 @@
 #include "arg.h"
 
 #include "common.h"
-#include "gguf.h" // for reading GGUF splits
 #include "log.h"
 #include "download.h"
+#include "hf-cache.h"
 
 #define JSON_ASSERT GGML_ASSERT
 #include <nlohmann/json.hpp>
@@ -15,6 +15,7 @@
 #include <map>
 #include <mutex>
 #include <regex>
+#include <unordered_set>
 #include <string>
 #include <thread>
 #include <vector>
@@ -35,8 +36,6 @@
 #endif
 #endif
 
-#define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
 // isatty
 #if defined(_WIN32)
 #include <io.h>
@@ -51,31 +50,6 @@ using json = nlohmann::ordered_json;
 //
 
 // validate repo name format: owner/repo
-static bool validate_repo_name(const std::string & repo) {
-    static const std::regex repo_regex(R"(^[A-Za-z0-9_.\-]+\/[A-Za-z0-9_.\-]+$)");
-    return std::regex_match(repo, repo_regex);
-}
-
-static std::string get_manifest_path(const std::string & repo, const std::string & tag) {
-    // we use "=" to avoid clashing with other component, while still being allowed on windows
-    std::string fname = "manifest=" + repo + "=" + tag + ".json";
-    if (!validate_repo_name(repo)) {
-        throw std::runtime_error("error: repo name must be in the format 'owner/repo'");
-    }
-    string_replace_all(fname, "/", "=");
-    return fs_get_cache_file(fname);
-}
-
-static std::string read_file(const std::string & fname) {
-    std::ifstream file(fname);
-    if (!file) {
-        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
-    }
-    std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
-    file.close();
-    return content;
-}
-
 static void write_file(const std::string & fname, const std::string & content) {
     const std::string fname_tmp = fname + ".tmp";
     std::ofstream     file(fname_tmp);
@@ -132,7 +106,7 @@ static bool is_http_status_ok(int status) {
 
 std::pair<std::string, std::string> common_download_split_repo_tag(const std::string & hf_repo_with_tag) {
     auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
+    std::string tag = parts.size() > 1 ? parts.back() : "";
     std::string hf_repo = parts[0];
     if (string_split<std::string>(hf_repo, '/').size() != 2) {
         throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
@@ -290,7 +264,8 @@ static bool common_pull_file(httplib::Client & cli,
 static int common_download_file_single_online(const std::string        & url,
                                               const std::string        & path,
                                               const std::string        & bearer_token,
-                                              const common_header_list & custom_headers) {
+                                              const common_header_list & custom_headers,
+                                              bool                       skip_etag = false) {
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
 
@@ -310,6 +285,11 @@ static int common_download_file_single_online(const std::string        & url,
 
     const bool file_exists = std::filesystem::exists(path);
 
+    if (file_exists && skip_etag) {
+        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
+        return 304; // 304 Not Modified - fake cached response
+    }
+
     std::string last_etag;
     if (file_exists) {
         last_etag = read_etag(path);
@@ -361,6 +341,12 @@ static int common_download_file_single_online(const std::string        & url,
         }
     }
 
+    { // silent
+        std::error_code ec;
+        std::filesystem::path p(path);
+        std::filesystem::create_directories(p.parent_path(), ec);
+    }
+
     const std::string path_temporary = path + ".downloadInProgress";
     int delay = retry_delay_seconds;
 
@@ -391,7 +377,7 @@ static int common_download_file_single_online(const std::string        & url,
                 LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
                 return -1;
             }
-            if (!etag.empty()) {
+            if (!etag.empty() && !skip_etag) {
                 write_etag(path, etag);
             }
             return head->status;
@@ -440,9 +426,10 @@ int common_download_file_single(const std::string & url,
                                 const std::string & path,
                                 const std::string & bearer_token,
                                 bool offline,
-                                const common_header_list & headers) {
+                                const common_header_list & headers,
+                                bool skip_etag) {
     if (!offline) {
-        return common_download_file_single_online(url, path, bearer_token, headers);
+        return common_download_file_single_online(url, path, bearer_token, headers, skip_etag);
     }
 
     if (!std::filesystem::exists(path)) {
@@ -454,193 +441,307 @@ int common_download_file_single(const std::string & url,
     return 304; // Not Modified - fake cached response
 }
 
-// download multiple files from remote URLs to local paths
-// the input is a vector of pairs <url, path>
-static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls,
-                                          const std::string & bearer_token,
-                                          bool offline,
-                                          const common_header_list & headers) {
-    // Prepare download in parallel
-    std::vector<std::future<bool>> futures_download;
-    futures_download.reserve(urls.size());
+struct gguf_split_info {
+    std::string prefix; // tag included
+    std::string tag;
+    int index;
+    int count;
+};
 
-    for (auto const & item : urls) {
-        futures_download.push_back(
-            std::async(
-                std::launch::async,
-                [&bearer_token, offline, &headers](const std::pair<std::string, std::string> & it) -> bool {
-                    const int http_status = common_download_file_single(it.first, it.second, bearer_token, offline, headers);
-                    return is_http_status_ok(http_status);
-                },
-                item
-            )
-        );
+static gguf_split_info get_gguf_split_info(const std::string & path) {
+    static const std::regex re_split("^(.+)-([0-9]{5})-of-([0-9]{5})$", std::regex::icase);
+    static const std::regex re_tag("[-.]([A-Z0-9_]+)$", std::regex::icase);
+    std::smatch m;
+
+    std::string prefix = path;
+    if (!string_remove_suffix(prefix, ".gguf")) {
+        return {};
     }
 
-    // Wait for all downloads to complete
-    for (auto & f : futures_download) {
-        if (!f.get()) {
-            return false;
+    int index = 1;
+    int count = 1;
+
+    if (std::regex_match(prefix, m, re_split)) {
+        index = std::stoi(m[2].str());
+        count = std::stoi(m[3].str());
+        prefix = m[1].str();
+    }
+
+    std::string tag;
+    if (std::regex_search(prefix, m, re_tag)) {
+        tag = m[1].str();
+        for (char & c : tag) {
+            c = std::toupper((unsigned char)c);
         }
     }
 
-    return true;
+    return {std::move(prefix), std::move(tag), index, count};
 }
 
-bool common_download_model(const common_params_model & model,
-                           const std::string & bearer_token,
-                           bool offline,
-                           const common_header_list & headers) {
-    // Basic validation of the model.url
-    if (model.url.empty()) {
-        LOG_ERR("%s: invalid model url\n", __func__);
+// Q4_0 -> 4, F16 -> 16, NVFP4 -> 4, Q8_K_M -> 8, etc
+static int extract_quant_bits(const std::string & filename) {
+    auto split = get_gguf_split_info(filename);
+
+    auto pos = split.tag.find_first_of("0123456789");
+    if (pos == std::string::npos) {
+        return 0;
+    }
+
+    return std::stoi(split.tag.substr(pos));
+}
+
+static hf_cache::hf_files get_split_files(const hf_cache::hf_files & files,
+                                          const hf_cache::hf_file  & file) {
+    auto split = get_gguf_split_info(file.path);
+
+    if (split.count <= 1) {
+        return {file};
+    }
+    hf_cache::hf_files result;
+
+    for (const auto & f : files) {
+        auto split_f = get_gguf_split_info(f.path);
+        if (split_f.count == split.count && split_f.prefix == split.prefix) {
+            result.push_back(f);
+        }
+    }
+    return result;
+}
+
+static hf_cache::hf_file find_best_mmproj(const hf_cache::hf_files & files,
+                                          const std::string        & model) {
+    hf_cache::hf_file best;
+    size_t best_depth = 0;
+    int best_diff = 0;
+    bool found = false;
+
+    auto model_bits = extract_quant_bits(model);
+    auto model_parts = string_split<std::string>(model, '/');
+    auto model_dir = model_parts.end() - 1;
+
+    for (const auto & f : files) {
+        if (!string_ends_with(f.path, ".gguf") ||
+            f.path.find("mmproj") == std::string::npos) {
+            continue;
+        }
+
+        auto mmproj_parts = string_split<std::string>(f.path, '/');
+        auto mmproj_dir = mmproj_parts.end() - 1;
+
+        auto [_, dir] = std::mismatch(model_parts.begin(), model_dir,
+                                      mmproj_parts.begin(), mmproj_dir);
+        if (dir != mmproj_dir) {
+            continue;
+        }
+
+        size_t depth = dir - mmproj_parts.begin();
+        auto bits = extract_quant_bits(f.path);
+        auto diff = std::abs(bits - model_bits);
+
+        if (!found || depth > best_depth || (depth == best_depth && diff < best_diff)) {
+            best = f;
+            best_depth = depth;
+            best_diff = diff;
+            found = true;
+        }
+    }
+    return best;
+}
+
+static bool gguf_filename_is_model(const std::string & filepath) {
+    if (!string_ends_with(filepath, ".gguf")) {
         return false;
     }
 
-    const int http_status = common_download_file_single(model.url, model.path, bearer_token, offline, headers);
-    if (!is_http_status_ok(http_status)) {
-        return false;
+    std::string filename = filepath;
+    if (auto pos = filename.rfind('/'); pos != std::string::npos) {
+        filename = filename.substr(pos + 1);
     }
 
-    // check for additional GGUFs split to download
-    int n_split = 0;
-    {
-        struct gguf_init_params gguf_params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ NULL,
-        };
-        auto * ctx_gguf = gguf_init_from_file(model.path.c_str(), gguf_params);
-        if (!ctx_gguf) {
-            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, model.path.c_str());
-            return false;
-        }
-
-        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
-        if (key_n_split >= 0) {
-            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
-        }
-
-        gguf_free(ctx_gguf);
-    }
-
-    if (n_split > 1) {
-        char split_prefix[PATH_MAX] = {0};
-        char split_url_prefix[LLAMA_MAX_URL_LENGTH] = {0};
-
-        // Verify the first split file format
-        // and extract split URL and PATH prefixes
-        {
-            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), model.path.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, model.path.c_str(), n_split);
-                return false;
-            }
-
-            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model.url.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model.url.c_str(), n_split);
-                return false;
-            }
-        }
-
-        std::vector<std::pair<std::string, std::string>> urls;
-        for (int idx = 1; idx < n_split; idx++) {
-            char split_path[PATH_MAX] = {0};
-            llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);
-
-            char split_url[LLAMA_MAX_URL_LENGTH] = {0};
-            llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split);
-
-            if (std::string(split_path) == model.path) {
-                continue; // skip the already downloaded file
-            }
-
-            urls.push_back({split_url, split_path});
-        }
-
-        // Download in parallel
-        common_download_file_multiple(urls, bearer_token, offline, headers);
-    }
-
-    return true;
+    return filename.find("mmproj")  == std::string::npos &&
+           filename.find("imatrix") == std::string::npos;
 }
 
-common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag,
-                                      const std::string & bearer_token,
-                                      bool offline,
-                                      const common_header_list & custom_headers) {
-    // the returned hf_repo is without tag
-    auto [hf_repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
+static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
+                                         const std::string        & tag) {
+    std::vector<std::string> tags;
 
-    std::string url = get_model_endpoint() + "v2/" + hf_repo + "/manifests/" + tag;
-
-    // headers
-    common_header_list headers = custom_headers;
-    headers.push_back({"Accept", "application/json"});
-    if (!bearer_token.empty()) {
-        headers.push_back({"Authorization", "Bearer " + bearer_token});
-    }
-    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    // User-Agent header is already set in common_remote_get_content, no need to set it here
-
-    // make the request
-    common_remote_params params;
-    params.headers = headers;
-    long res_code = 0;
-    std::string res_str;
-    bool use_cache = false;
-    std::string cached_response_path = get_manifest_path(hf_repo, tag);
-    if (!offline) {
-        try {
-            auto res = common_remote_get_content(url, params);
-            res_code = res.first;
-            res_str = std::string(res.second.data(), res.second.size());
-        } catch (const std::exception & e) {
-            LOG_WRN("error: failed to get manifest at %s: %s\n", url.c_str(), e.what());
-        }
-    }
-    if (res_code == 0) {
-        if (std::filesystem::exists(cached_response_path)) {
-            LOG_WRN("trying to read manifest from cache: %s\n", cached_response_path.c_str());
-            res_str = read_file(cached_response_path);
-            res_code = 200;
-            use_cache = true;
-        } else {
-            throw std::runtime_error(
-                offline ? "error: failed to get manifest (offline mode)"
-                : "error: failed to get manifest (check your internet connection)");
-        }
-    }
-    std::string ggufFile;
-    std::string mmprojFile;
-
-    if (res_code == 200 || res_code == 304) {
-        try {
-            auto j = json::parse(res_str);
-
-            if (j.contains("ggufFile") && j["ggufFile"].contains("rfilename")) {
-                ggufFile = j["ggufFile"]["rfilename"].get<std::string>();
-            }
-            if (j.contains("mmprojFile") && j["mmprojFile"].contains("rfilename")) {
-                mmprojFile = j["mmprojFile"]["rfilename"].get<std::string>();
-            }
-        } catch (const std::exception & e) {
-            throw std::runtime_error(std::string("error parsing manifest JSON: ") + e.what());
-        }
-        if (!use_cache) {
-            // if not using cached response, update the cache file
-            write_file(cached_response_path, res_str);
-        }
-    } else if (res_code == 401) {
-        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
+    if (!tag.empty()) {
+        tags.push_back(tag);
     } else {
-        throw std::runtime_error(string_format("error from HF API (%s), response code: %ld, data: %s", url.c_str(), res_code, res_str.c_str()));
+        tags = {"Q4_K_M", "Q4_0"};
     }
 
-    // check response
-    if (ggufFile.empty()) {
-        throw std::runtime_error("error: model does not have ggufFile");
+    for (const auto & t : tags) {
+        std::regex pattern(t + "[.-]", std::regex::icase);
+        for (const auto & f : files) {
+            if (gguf_filename_is_model(f.path) &&
+                std::regex_search(f.path, pattern)) {
+                return f;
+            }
+        }
     }
 
-    return { hf_repo, ggufFile, mmprojFile };
+    for (const auto & f : files) {
+        if (gguf_filename_is_model(f.path)) {
+            return f;
+        }
+    }
+
+    return {};
+}
+
+static void list_available_gguf_files(const hf_cache::hf_files & files) {
+    LOG_INF("Available GGUF files:\n");
+    for (const auto & f : files) {
+        if (string_ends_with(f.path, ".gguf")) {
+            LOG_INF(" - %s\n", f.path.c_str());
+        }
+    }
+}
+
+struct hf_plan {
+    hf_cache::hf_files model_files;
+    hf_cache::hf_file mmproj;
+};
+
+static hf_plan get_hf_plan(const common_params_model        & model,
+                           const std::string                & token,
+                           const common_download_model_opts & opts) {
+    hf_plan plan;
+    hf_cache::hf_files all;
+
+    auto [repo, tag] = common_download_split_repo_tag(model.hf_repo);
+
+    if (!opts.offline) {
+        all = hf_cache::get_repo_files(repo, token);
+    }
+    if (all.empty()) {
+        all = hf_cache::get_cached_files(repo);
+    }
+    if (all.empty()) {
+        return plan;
+    }
+
+    hf_cache::hf_file primary;
+
+    if (!model.hf_file.empty()) {
+        for (const auto & f : all) {
+            if (f.path == model.hf_file) {
+                primary = f;
+                break;
+            }
+        }
+        if (primary.path.empty()) {
+            LOG_ERR("%s: file '%s' not found in repository\n", __func__, model.hf_file.c_str());
+            list_available_gguf_files(all);
+            return plan;
+        }
+    } else {
+        primary = find_best_model(all, tag);
+        if (primary.path.empty()) {
+            LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
+            list_available_gguf_files(all);
+            return plan;
+        }
+    }
+
+    plan.model_files = get_split_files(all, primary);
+
+    if (opts.download_mmproj) {
+        plan.mmproj = find_best_mmproj(all, primary.path);
+    }
+
+    return plan;
+}
+
+struct download_task {
+    std::string url;
+    std::string path;
+};
+
+static std::vector<download_task> get_url_tasks(const common_params_model & model) {
+    auto split = get_gguf_split_info(model.url);
+
+    if (split.count <= 1) {
+        return {{model.url, model.path}};
+    }
+
+    auto filename = split.prefix;
+    if (auto pos = split.prefix.rfind('/'); pos != std::string::npos) {
+        filename = split.prefix.substr(pos + 1);
+    }
+
+    auto parent_path = std::filesystem::path(model.path).parent_path();
+    auto prefix_path = (parent_path / filename).string();
+
+    std::vector<download_task> tasks;
+    for (int i = 1; i <= split.count; i++) {
+        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
+        tasks.push_back({split.prefix + suffix, prefix_path + suffix});
+    }
+    return tasks;
+}
+
+common_download_model_result common_download_model(const common_params_model        & model,
+                                                   const std::string                & bearer_token,
+                                                   const common_download_model_opts & opts,
+                                                   const common_header_list         & headers) {
+    common_download_model_result result;
+    std::vector<download_task> tasks;
+    hf_plan hf;
+
+    bool is_hf = !model.hf_repo.empty();
+
+    if (is_hf) {
+        hf = get_hf_plan(model, bearer_token, opts);
+        for (const auto & f : hf.model_files) {
+            tasks.push_back({f.url, f.local_path});
+        }
+        if (!hf.mmproj.path.empty()) {
+            tasks.push_back({hf.mmproj.url, hf.mmproj.local_path});
+        }
+    } else if (!model.url.empty()) {
+        tasks = get_url_tasks(model);
+    } else {
+        result.model_path = model.path;
+        return result;
+    }
+
+    if (tasks.empty()) {
+        return result;
+    }
+
+    std::vector<std::future<bool>> futures;
+    for (const auto & task : tasks) {
+        futures.push_back(std::async(std::launch::async,
+            [&task, &bearer_token, offline = opts.offline, &headers, is_hf]() {
+                int status = common_download_file_single(task.url, task.path, bearer_token, offline, headers, is_hf);
+                return is_http_status_ok(status);
+            }
+        ));
+    }
+
+    for (auto & f : futures) {
+        if (!f.get()) {
+            return {};
+        }
+    }
+
+    if (is_hf) {
+        for (const auto & f : hf.model_files) {
+            hf_cache::finalize_file(f);
+        }
+        result.model_path = hf.model_files[0].final_path;
+
+        if (!hf.mmproj.path.empty()) {
+            result.mmproj_path = hf_cache::finalize_file(hf.mmproj);
+        }
+    } else {
+        result.model_path = model.path;
+    }
+
+    return result;
 }
 
 //
@@ -765,28 +866,21 @@ std::string common_docker_resolve_model(const std::string & docker) {
 }
 
 std::vector<common_cached_model_info> common_list_cached_models() {
-    std::vector<common_cached_model_info> models;
-    const std::string cache_dir = fs_get_cache_directory();
-    const std::vector<common_file_info> files = fs_list(cache_dir, false);
-    for (const auto & file : files) {
-        if (string_starts_with(file.name, "manifest=") && string_ends_with(file.name, ".json")) {
-            common_cached_model_info model_info;
-            model_info.manifest_path = file.path;
-            std::string fname = file.name;
-            string_replace_all(fname, ".json", ""); // remove extension
-            auto parts = string_split<std::string>(fname, '=');
-            if (parts.size() == 4) {
-                // expect format: manifest=<user>=<model>=<tag>=<other>
-                model_info.user  = parts[1];
-                model_info.model = parts[2];
-                model_info.tag   = parts[3];
-            } else {
-                // invalid format
-                continue;
-            }
-            model_info.size = 0; // TODO: get GGUF size, not manifest size
-            models.push_back(model_info);
+    std::unordered_set<std::string> seen;
+    std::vector<common_cached_model_info> result;
+
+    auto files = hf_cache::get_cached_files();
+
+    for (const auto & f : files) {
+        auto split = get_gguf_split_info(f.path);
+        if (split.index != 1 || split.tag.empty() ||
+            split.prefix.find("mmproj") != std::string::npos) {
+            continue;
+        }
+        if (seen.insert(f.repo_id + ":" + split.tag).second) {
+            result.push_back({f.repo_id, split.tag});
         }
     }
-    return models;
+
+    return result;
 }

common/download.h

@@ -17,54 +17,60 @@ struct common_remote_params {
 // get remote file content, returns <http_code, raw_response_body>
 std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);
 
-// split HF repo with tag into <repo, tag>
-// for example: "user/model:tag" -> <"user/model", "tag">
-// if tag is not present, default to "latest"
-// example: "user/model" -> <"user/model", "latest">
+// split HF repo with tag into <repo, tag>, for example:
+// - "ggml-org/models:F16" -> <"ggml-org/models", "F16">
+// tag is optional and can be empty
 std::pair<std::string, std::string> common_download_split_repo_tag(const std::string & hf_repo_with_tag);
 
+// Result of common_list_cached_models
 struct common_cached_model_info {
-    std::string manifest_path;
-    std::string user;
-    std::string model;
+    std::string repo;
     std::string tag;
-    size_t      size = 0; // GGUF size in bytes
-    // return string representation like "user/model:tag"
-    // if tag is "latest", it will be omitted
     std::string to_string() const {
-        return user + "/" + model + (tag == "latest" ? "" : ":" + tag);
+        return repo + ":" + tag;
     }
 };
 
-struct common_hf_file_res {
-    std::string repo; // repo name with ":tag" removed
-    std::string ggufFile;
-    std::string mmprojFile;
+// Options for common_download_model
+struct common_download_model_opts {
+    bool download_mmproj = false;
+    bool offline         = false;
 };
 
-/**
- * Allow getting the HF file from the HF repo with tag (like ollama), for example:
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
- * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
- *
- * Return pair of <repo, file> (with "repo" already having tag removed)
- *
- * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
- */
-common_hf_file_res common_get_hf_file(
-    const std::string & hf_repo_with_tag,
-    const std::string & bearer_token,
-    bool offline,
-    const common_header_list & headers = {}
-);
+// Result of common_download_model
+struct common_download_model_result {
+    std::string model_path;
+    std::string mmproj_path;
+};
 
-// returns true if download succeeded
-bool common_download_model(
+// Download model from HuggingFace repo or URL
+//
+// input (via model struct):
+// - model.hf_repo: HF repo with optional tag, see common_download_split_repo_tag
+// - model.hf_file: specific file in the repo (requires hf_repo)
+// - model.url: simple download (used if hf_repo is empty)
+// - model.path: local file path
+//
+// tag matching (for HF repos without model.hf_file):
+// - if tag is specified, searches for GGUF matching that quantization
+// - if no tag, searches for Q4_K_M, then Q4_0, then first available GGUF
+//
+// split GGUF: multi-part files like "model-00001-of-00003.gguf" are automatically
+// detected and all parts are downloaded
+//
+// caching:
+// - HF repos: uses HuggingFace cache
+// - URLs: uses ETag-based caching
+//
+// when opts.offline=true, no network requests are made
+// when download_mmproj=true, searches for mmproj in same directory as model or any parent directory
+// then with the closest quantization bits
+//
+// returns result with model_path and mmproj_path (empty on failure)
+common_download_model_result common_download_model(
     const common_params_model & model,
     const std::string & bearer_token,
-    bool offline,
+    const common_download_model_opts & opts = {},
     const common_header_list & headers = {}
 );
 
@@ -73,11 +79,13 @@ std::vector<common_cached_model_info> common_list_cached_models();
 
 // download single file from url to local path
 // returns status code or -1 on error
+// skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
 int common_download_file_single(const std::string & url,
                                 const std::string & path,
                                 const std::string & bearer_token,
                                 bool offline,
-                                const common_header_list & headers = {});
+                                const common_header_list & headers = {},
+                                bool skip_etag = false);
 
 // resolve and download model from Docker registry
 // return local path to downloaded model file

common/hf-cache.cpp

@@ -0,0 +1,762 @@
+#include "hf-cache.h"
+
+#include "common.h"
+#include "log.h"
+#include "http.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+#include <filesystem>
+#include <fstream>
+#include <atomic>
+#include <regex> // migration only
+#include <string>
+#include <string_view>
+#include <stdexcept>
+
+namespace nl = nlohmann;
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#define HOME_DIR "USERPROFILE"
+#include <windows.h>
+#else
+#define HOME_DIR "HOME"
+#endif
+
+namespace hf_cache {
+
+namespace fs = std::filesystem;
+
+static fs::path get_cache_directory() {
+    static const fs::path cache = []() {
+        struct {
+            const char * var;
+            fs::path path;
+        } entries[] = {
+            {"LLAMA_CACHE",           fs::path()},
+            {"HF_HUB_CACHE",          fs::path()},
+            {"HUGGINGFACE_HUB_CACHE", fs::path()},
+            {"HF_HOME",               fs::path("hub")},
+            {"XDG_CACHE_HOME",        fs::path("huggingface") / "hub"},
+            {HOME_DIR,                fs::path(".cache") / "huggingface" / "hub"}
+        };
+        for (const auto & entry : entries) {
+            if (auto * p = std::getenv(entry.var); p && *p) {
+                fs::path base(p);
+                return entry.path.empty() ? base : base / entry.path;
+            }
+        }
+        throw std::runtime_error("Failed to determine HF cache directory");
+    }();
+
+    return cache;
+}
+
+static std::string folder_name_to_repo(const std::string & folder) {
+    constexpr std::string_view prefix = "models--";
+    if (folder.rfind(prefix, 0)) {
+        return {};
+    }
+    std::string result = folder.substr(prefix.length());
+    string_replace_all(result, "--", "/");
+    return result;
+}
+
+static std::string repo_to_folder_name(const std::string & repo_id) {
+    constexpr std::string_view prefix = "models--";
+    std::string result = std::string(prefix) + repo_id;
+    string_replace_all(result, "/", "--");
+    return result;
+}
+
+static fs::path get_repo_path(const std::string & repo_id) {
+    return get_cache_directory() / repo_to_folder_name(repo_id);
+}
+
+static bool is_hex_char(const char c) {
+    return (c >= 'A' && c <= 'F') ||
+           (c >= 'a' && c <= 'f') ||
+           (c >= '0' && c <= '9');
+}
+
+static bool is_hex_string(const std::string & s, size_t expected_len) {
+    if (s.length() != expected_len) {
+        return false;
+    }
+    for (const char c : s) {
+        if (!is_hex_char(c)) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool is_alphanum(const char c) {
+    return (c >= 'A' && c <= 'Z') ||
+           (c >= 'a' && c <= 'z') ||
+           (c >= '0' && c <= '9');
+}
+
+static bool is_special_char(char c) {
+    return c == '/' || c == '.' || c == '-';
+}
+
+// base chars [A-Za-z0-9_] are always valid
+// special chars [/.-] must be surrounded by base chars
+// exactly one '/' required
+static bool is_valid_repo_id(const std::string & repo_id) {
+    if (repo_id.empty() || repo_id.length() > 256) {
+        return false;
+    }
+    int slash = 0;
+    bool special = true;
+
+    for (const char c : repo_id) {
+        if (is_alphanum(c) || c == '_') {
+            special = false;
+        } else if (is_special_char(c)) {
+            if (special) {
+                return false;
+            }
+            slash += (c == '/');
+            special = true;
+        } else {
+            return false;
+        }
+    }
+    return !special && slash == 1;
+}
+
+static bool is_valid_hf_token(const std::string & token) {
+    if (token.length() < 37 || token.length() > 256 ||
+        !string_starts_with(token, "hf_")) {
+        return false;
+    }
+    for (size_t i = 3; i < token.length(); ++i) {
+        if (!is_alphanum(token[i])) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool is_valid_commit(const std::string & hash) {
+    return is_hex_string(hash, 40);
+}
+
+static bool is_valid_oid(const std::string & oid) {
+    return is_hex_string(oid, 40) || is_hex_string(oid, 64);
+}
+
+static bool is_valid_subpath(const fs::path & path, const fs::path & subpath) {
+    if (subpath.is_absolute()) {
+        return false; // never do a / b with b absolute
+    }
+    auto b = fs::absolute(path).lexically_normal();
+    auto t = (b / subpath).lexically_normal();
+    auto [b_end, _] = std::mismatch(b.begin(), b.end(), t.begin(), t.end());
+
+    return b_end == b.end();
+}
+
+static void safe_write_file(const fs::path & path, const std::string & data) {
+    fs::path path_tmp = path.string() + ".tmp";
+
+    if (path.has_parent_path()) {
+        fs::create_directories(path.parent_path());
+    }
+
+    std::ofstream file(path_tmp);
+    file << data;
+    file.close();
+
+    std::error_code ec;
+
+    if (!file.fail()) {
+        fs::rename(path_tmp, path, ec);
+    }
+    if (file.fail() || ec) {
+        fs::remove(path_tmp, ec);
+        throw std::runtime_error("failed to write file: " + path.string());
+    }
+}
+
+static nl::json api_get(const std::string & url,
+                        const std::string & token) {
+    auto [cli, parts] = common_http_client(url);
+
+    httplib::Headers headers = {
+        {"User-Agent", "llama-cpp/" + build_info},
+        {"Accept", "application/json"}
+    };
+
+    if (is_valid_hf_token(token)) {
+        headers.emplace("Authorization", "Bearer " + token);
+    } else if (!token.empty()) {
+        LOG_WRN("%s: invalid token, authentication disabled\n", __func__);
+    }
+
+    if (auto res = cli.Get(parts.path, headers)) {
+        auto body = res->body;
+
+        if (res->status == 200) {
+            return nl::json::parse(res->body);
+        }
+        try {
+            body = nl::json::parse(res->body)["error"].get<std::string>();
+        } catch (...) { }
+
+        throw std::runtime_error("GET failed (" + std::to_string(res->status) + "): " + body);
+    } else {
+        throw std::runtime_error("HTTPLIB failed: " + httplib::to_string(res.error()));
+    }
+}
+
+static std::string get_repo_commit(const std::string & repo_id,
+                                   const std::string & token) {
+    try {
+        auto endpoint = get_model_endpoint();
+        auto json = api_get(endpoint + "api/models/" + repo_id + "/refs", token);
+
+        if (!json.is_object() ||
+            !json.contains("branches") || !json["branches"].is_array()) {
+            LOG_WRN("%s: missing 'branches' for '%s'\n", __func__, repo_id.c_str());
+            return {};
+        }
+
+        fs::path refs_path = get_repo_path(repo_id) / "refs";
+        std::string name;
+        std::string commit;
+
+        for (const auto & branch : json["branches"]) {
+            if (!branch.is_object() ||
+                !branch.contains("name") || !branch["name"].is_string() ||
+                !branch.contains("targetCommit") || !branch["targetCommit"].is_string()) {
+                continue;
+            }
+            std::string _name = branch["name"].get<std::string>();
+            std::string _commit = branch["targetCommit"].get<std::string>();
+
+            if (!is_valid_subpath(refs_path, _name)) {
+                LOG_WRN("%s: skip invalid branch: %s\n", __func__, _name.c_str());
+                continue;
+            }
+            if (!is_valid_commit(_commit)) {
+                LOG_WRN("%s: skip invalid commit: %s\n", __func__, _commit.c_str());
+                continue;
+            }
+
+            if (_name == "main") {
+                name = _name;
+                commit = _commit;
+                break;
+            }
+
+            if (name.empty() || commit.empty()) {
+                name = _name;
+                commit = _commit;
+            }
+        }
+
+        if (name.empty() || commit.empty()) {
+            LOG_WRN("%s: no valid branch for '%s'\n", __func__, repo_id.c_str());
+            return {};
+        }
+
+        safe_write_file(refs_path / name, commit);
+        return commit;
+
+    } catch (const nl::json::exception & e) {
+        LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+    }
+    return {};
+}
+
+hf_files get_repo_files(const std::string & repo_id,
+                        const std::string & token) {
+    if (!is_valid_repo_id(repo_id)) {
+        LOG_WRN("%s: invalid repository: %s\n", __func__, repo_id.c_str());
+        return {};
+    }
+
+    std::string commit = get_repo_commit(repo_id, token);
+    if (commit.empty()) {
+        LOG_WRN("%s: failed to resolve commit for %s\n", __func__, repo_id.c_str());
+        return {};
+    }
+
+    fs::path blobs_path = get_repo_path(repo_id) / "blobs";
+    fs::path commit_path = get_repo_path(repo_id) / "snapshots" / commit;
+
+    hf_files files;
+
+    try {
+        auto endpoint = get_model_endpoint();
+        auto json = api_get(endpoint + "api/models/" + repo_id + "/tree/" + commit + "?recursive=true", token);
+
+        if (!json.is_array()) {
+            LOG_WRN("%s: response is not an array for '%s'\n", __func__, repo_id.c_str());
+            return {};
+        }
+
+        for (const auto & item : json) {
+            if (!item.is_object() ||
+                !item.contains("type") || !item["type"].is_string() || item["type"] != "file" ||
+                !item.contains("path") || !item["path"].is_string()) {
+                continue;
+            }
+
+            hf_file file;
+            file.repo_id = repo_id;
+            file.path = item["path"].get<std::string>();
+
+            if (!is_valid_subpath(commit_path, file.path)) {
+                LOG_WRN("%s: skip invalid path: %s\n", __func__, file.path.c_str());
+                continue;
+            }
+
+            if (item.contains("lfs") && item["lfs"].is_object()) {
+                if (item["lfs"].contains("oid") && item["lfs"]["oid"].is_string()) {
+                    file.oid = item["lfs"]["oid"].get<std::string>();
+                }
+                if (item["lfs"].contains("size") && item["lfs"]["size"].is_number()) {
+                    file.size = item["lfs"]["size"].get<size_t>();
+                }
+            } else if (item.contains("oid") && item["oid"].is_string()) {
+                file.oid = item["oid"].get<std::string>();
+            }
+            if (file.size == 0 && item.contains("size") && item["size"].is_number()) {
+                file.size = item["size"].get<size_t>();
+            }
+
+            if (!file.oid.empty() && !is_valid_oid(file.oid)) {
+                LOG_WRN("%s: skip invalid oid: %s\n", __func__, file.oid.c_str());
+                continue;
+            }
+
+            file.url = endpoint + repo_id + "/resolve/" + commit + "/" + file.path;
+
+            fs::path final_path = commit_path / file.path;
+            file.final_path = final_path.string();
+
+            if (!file.oid.empty() && !fs::exists(final_path)) {
+                fs::path local_path = blobs_path / file.oid;
+                file.local_path = local_path.string();
+            } else {
+                file.local_path = file.final_path;
+            }
+
+            files.push_back(file);
+        }
+    } catch (const nl::json::exception & e) {
+        LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+    }
+    return files;
+}
+
+static std::string get_cached_ref(const fs::path & repo_path) {
+    fs::path refs_path = repo_path / "refs";
+    if (!fs::is_directory(refs_path)) {
+        return {};
+    }
+    std::string fallback;
+
+    for (const auto & entry : fs::directory_iterator(refs_path)) {
+        if (!entry.is_regular_file()) {
+            continue;
+        }
+        std::ifstream f(entry.path());
+        std::string commit;
+        if (!f || !std::getline(f, commit) || commit.empty()) {
+            continue;
+        }
+        if (!is_valid_commit(commit)) {
+            LOG_WRN("%s: skip invalid commit: %s\n", __func__, commit.c_str());
+            continue;
+        }
+        if (entry.path().filename() == "main") {
+            return commit;
+        }
+        if (fallback.empty()) {
+            fallback = commit;
+        }
+    }
+    return fallback;
+}
+
+hf_files get_cached_files(const std::string & repo_id) {
+    fs::path cache_dir = get_cache_directory();
+    if (!fs::exists(cache_dir)) {
+        return {};
+    }
+
+    if (!repo_id.empty() && !is_valid_repo_id(repo_id)) {
+        LOG_WRN("%s: invalid repository: %s\n", __func__, repo_id.c_str());
+        return {};
+    }
+
+    hf_files files;
+
+    for (const auto & repo : fs::directory_iterator(cache_dir)) {
+        if (!repo.is_directory()) {
+            continue;
+        }
+        fs::path snapshots_path = repo.path() / "snapshots";
+
+        if (!fs::exists(snapshots_path)) {
+            continue;
+        }
+        std::string _repo_id = folder_name_to_repo(repo.path().filename().string());
+
+        if (!is_valid_repo_id(_repo_id)) {
+            continue;
+        }
+        if (!repo_id.empty() && _repo_id != repo_id) {
+            continue;
+        }
+        std::string commit = get_cached_ref(repo.path());
+        fs::path commit_path = snapshots_path / commit;
+
+        if (commit.empty() || !fs::is_directory(commit_path)) {
+            continue;
+        }
+        for (const auto & entry : fs::recursive_directory_iterator(commit_path)) {
+            if (!entry.is_regular_file() && !entry.is_symlink()) {
+                continue;
+            }
+            fs::path path = entry.path().lexically_relative(commit_path);
+
+            if (!path.empty()) {
+                hf_file file;
+                file.repo_id = _repo_id;
+                file.path = path.generic_string();
+                file.local_path = entry.path().string();
+                file.final_path = file.local_path;
+                files.push_back(std::move(file));
+            }
+        }
+    }
+
+    return files;
+}
+
+std::string finalize_file(const hf_file & file) {
+    static std::atomic<bool> symlinks_disabled{false};
+
+    std::error_code ec;
+    fs::path local_path(file.local_path);
+    fs::path final_path(file.final_path);
+
+    if (local_path == final_path || fs::exists(final_path, ec)) {
+        return file.final_path;
+    }
+
+    if (!fs::exists(local_path, ec)) {
+        return file.final_path;
+    }
+
+    fs::create_directories(final_path.parent_path(), ec);
+
+    if (!symlinks_disabled) {
+        fs::path target = fs::relative(local_path, final_path.parent_path(), ec);
+        if (!ec) {
+            fs::create_symlink(target, final_path, ec);
+        }
+        if (!ec) {
+            return file.final_path;
+        }
+    }
+
+    if (!symlinks_disabled.exchange(true)) {
+        LOG_WRN("%s: failed to create symlink: %s\n", __func__, ec.message().c_str());
+        LOG_WRN("%s: switching to degraded mode\n", __func__);
+    }
+
+    fs::rename(local_path, final_path, ec);
+    if (ec) {
+        LOG_WRN("%s: failed to move file to snapshots: %s\n", __func__, ec.message().c_str());
+        fs::copy(local_path, final_path, ec);
+        if (ec) {
+            LOG_ERR("%s: failed to copy file to snapshots: %s\n", __func__, ec.message().c_str());
+        }
+    }
+    return file.final_path;
+}
+
+// delete everything after this line, one day
+
+// copied from download.cpp without the tag part
+struct gguf_split_info {
+    std::string prefix; // tag included
+    int index;
+    int count;
+};
+
+static gguf_split_info get_gguf_split_info(const std::string & path) {
+    static const std::regex re_split("^(.+)-([0-9]{5})-of-([0-9]{5})$", std::regex::icase);
+    std::smatch m;
+
+    std::string prefix = path;
+    if (!string_remove_suffix(prefix, ".gguf")) {
+        return {};
+    }
+
+    int index = 1;
+    int count = 1;
+
+    if (std::regex_match(prefix, m, re_split)) {
+        index = std::stoi(m[2].str());
+        count = std::stoi(m[3].str());
+        prefix = m[1].str();
+    }
+
+    return {std::move(prefix), index, count};
+}
+
+static std::pair<std::string, std::string> parse_manifest_name(std::string & filename) {
+    static const std::regex re(R"(^manifest=([^=]+)=([^=]+)=.*\.json$)");
+    std::smatch match;
+    if (std::regex_match(filename, match, re)) {
+        return {match[1].str(), match[2].str()};
+    }
+    return {};
+}
+
+static std::string make_old_cache_filename(const std::string & owner,
+                                           const std::string & repo,
+                                           const std::string & filename) {
+    auto result = owner + "_" + repo + "_" + filename;
+    string_replace_all(result, "/", "_");
+    return result;
+}
+
+struct migrate_file {
+    std::string path;
+    std::string sha256;
+    size_t size;
+    fs::path old_path;
+    fs::path etag_path;
+    const hf_file * file;
+};
+
+using migrate_files = std::vector<migrate_file>;
+
+static bool collect_file(const fs::path    & old_cache,
+                         const std::string & owner,
+                         const std::string & repo,
+                         const std::string & path,
+                         const std::string & sha256,
+                         const hf_files    & files,
+                         migrate_files     & to_migrate) {
+
+    const hf_file * file = nullptr;
+
+    for (const auto & f : files) {
+        if (f.path == path) {
+            file = &f;
+            break;
+        }
+    }
+
+    std::string old_filename = make_old_cache_filename(owner, repo, path);
+    fs::path old_path = old_cache / old_filename;
+    fs::path etag_path = old_path.string() + ".etag";
+
+    if (!fs::exists(old_path)) {
+        if (file && fs::exists(file->final_path)) {
+            return true;
+        }
+        LOG_WRN("%s: %s not found in old cache or HF cache\n", __func__, old_filename.c_str());
+        return false;
+    }
+
+    if (!file) {
+        LOG_WRN("%s: %s not found in current repo\n", __func__, old_filename.c_str());
+        return false;
+    }
+
+    if (!sha256.empty() && !file->oid.empty() && sha256 != file->oid) {
+        LOG_WRN("%s: %s is not up to date (sha256 mismatch)\n", __func__, old_filename.c_str());
+        return false;
+    }
+
+    if (file->size > 0) {
+        size_t size = fs::file_size(old_path);
+        if (size != file->size) {
+            LOG_WRN("%s: %s has wrong size %zu (expected %zu)\n", __func__, old_filename.c_str(), size, file->size);
+            return false;
+        }
+    }
+
+    to_migrate.push_back({path, sha256, file->size, old_path, etag_path, file});
+    return true;
+}
+
+static bool collect_files(const fs::path    & old_cache,
+                          const std::string & owner,
+                          const std::string & repo,
+                          const nl::json    & node,
+                          const hf_files    & files,
+                          migrate_files     & to_migrate) {
+
+    if (!node.contains("rfilename") ||
+        !node.contains("lfs")       ||
+        !node["lfs"].contains("sha256")) {
+        return true;
+    }
+
+    std::string path = node["rfilename"];
+    std::string sha256 = node["lfs"]["sha256"];
+
+    auto split = get_gguf_split_info(path);
+
+    if (split.count <= 1) {
+        return collect_file(old_cache, owner, repo, path, sha256, files, to_migrate);
+    }
+
+    std::vector<std::pair<std::string, std::string>> splits;
+
+    for (const auto & f : files) {
+        auto split_f = get_gguf_split_info(f.path);
+        if (split_f.count == split.count && split_f.prefix == split.prefix) {
+            // sadly the manifest only provides the sha256 of the first file (index == 1)
+            // the rest will be verified using the size...
+            std::string f_sha256 = (split_f.index == 1) ? sha256 : "";
+            splits.emplace_back(f.path, f_sha256);
+        }
+    }
+
+    if ((int)splits.size() != split.count) {
+        LOG_WRN("%s: expected %d split files but found %d in repo\n", __func__, split.count, (int)splits.size());
+        return false;
+    }
+
+    for (const auto & [f_path, f_sha256] : splits) {
+        if (!collect_file(old_cache, owner, repo, f_path, f_sha256, files, to_migrate)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool migrate_file(const migrate_file & file) {
+    std::error_code ec;
+
+    fs::path new_path(file.file->local_path);
+    fs::create_directories(new_path.parent_path(), ec);
+
+    if (!fs::exists(new_path, ec)) {
+        fs::rename(file.old_path, new_path, ec);
+        if (ec) {
+            fs::copy_file(file.old_path, new_path, ec);
+            if (ec) {
+                LOG_ERR("%s: failed to move/copy %s: %s\n", __func__, file.old_path.string().c_str(), ec.message().c_str());
+                return false;
+            }
+        }
+        fs::remove(file.old_path, ec);
+    }
+    fs::remove(file.etag_path, ec);
+
+    std::string filename = finalize_file(*file.file);
+    LOG_INF("%s: migrated %s -> %s\n", __func__, file.old_path.filename().string().c_str(), filename.c_str());
+    return true;
+}
+
+void migrate_old_cache_to_hf_cache(const std::string & token, bool offline) {
+    fs::path old_cache = fs_get_cache_directory();
+    if (!fs::exists(old_cache)) {
+        return;
+    }
+
+    if (offline) {
+        LOG_WRN("%s: skipping migration in offline mode (will run when online)\n", __func__);
+        return; // -hf is not going to work
+    }
+
+    bool warned = false;
+
+    for (const auto & entry : fs::directory_iterator(old_cache)) {
+        if (!entry.is_regular_file()) {
+            continue;
+        }
+        auto filename = entry.path().filename().string();
+        auto [owner, repo] = parse_manifest_name(filename);
+
+        if (owner.empty() || repo.empty()) {
+            continue;
+        }
+
+        if (!warned) {
+            warned = true;
+            LOG_WRN("================================================================================\n"
+                    "WARNING: Migrating cache to HuggingFace cache directory\n"
+                    "  Old cache: %s\n"
+                    "  New cache: %s\n"
+                    "This one-time migration moves models previously downloaded with -hf\n"
+                    "from the legacy llama.cpp cache to the standard HuggingFace cache.\n"
+                    "Models downloaded with --model-url are not affected.\n"
+                    "================================================================================\n",
+                    old_cache.string().c_str(), get_cache_directory().string().c_str());
+        }
+
+        auto repo_id = owner + "/" + repo;
+        auto files = get_repo_files(repo_id, token);
+
+        if (files.empty()) {
+            LOG_WRN("%s: could not get repo files for %s, skipping\n", __func__, repo_id.c_str());
+            continue;
+        }
+
+        migrate_files to_migrate;
+        bool ok = true;
+
+        try {
+            std::ifstream manifest(entry.path());
+            auto json = nl::json::parse(manifest);
+            for (const char * key : {"ggufFile", "mmprojFile"}) {
+                if (json.contains(key)) {
+                    if (!collect_files(old_cache, owner, repo, json[key], files, to_migrate)) {
+                        ok = false;
+                        break;
+                    }
+                }
+            }
+        } catch (const std::exception & e) {
+            LOG_WRN("%s: failed to parse manifest %s: %s\n", __func__, filename.c_str(), e.what());
+            continue;
+        }
+
+        if (!ok) {
+            LOG_WRN("%s: migration skipped: one or more files failed validation\n", __func__);
+            continue;
+        }
+
+        for (const auto & file : to_migrate) {
+            if (!migrate_file(file)) {
+                ok = false;
+                break;
+            }
+        }
+
+        if (!ok) {
+            LOG_WRN("%s: migration failed: could not migrate all files\n", __func__);
+            continue;
+        }
+
+        LOG_INF("%s: migration complete, deleting manifest: %s\n", __func__, entry.path().string().c_str());
+        fs::remove(entry.path());
+    }
+}
+
+} // namespace hf_cache

common/hf-cache.h

@@ -0,0 +1,36 @@
+#pragma once
+
+#include <string>
+#include <vector>
+
+// Ref: https://huggingface.co/docs/hub/local-cache.md
+
+namespace hf_cache {
+
+struct hf_file {
+    std::string path;
+    std::string url;
+    std::string local_path;
+    std::string final_path;
+    std::string oid;
+    std::string repo_id;
+    size_t size = 0; // only for the migration
+};
+
+using hf_files = std::vector<hf_file>;
+
+// Get files from HF API
+hf_files get_repo_files(
+    const std::string & repo_id,
+    const std::string & token
+);
+
+hf_files get_cached_files(const std::string & repo_id = {});
+
+// Create snapshot path (link or move/copy) and return it
+std::string finalize_file(const hf_file & file);
+
+// TODO: Remove later
+void migrate_old_cache_to_hf_cache(const std::string & token, bool offline = false);
+
+} // namespace hf_cache

common/jinja/parser.cpp

@@ -53,6 +53,13 @@ private:
         return tokens[current + offset];
     }
 
+    const token & next() {
+        if (current >= tokens.size()) {
+            throw parser_exception("Parser Error: Unexpected EOF", source, tokens.empty() ? 0 : tokens.back().pos);
+        }
+        return tokens[current++];
+    }
+
     token expect(token::type type, const std::string&  error) {
         const auto & t = peek();
         if (t.t != type) {
@@ -90,9 +97,9 @@ private:
         size_t start_pos = current;
         switch (peek().t) {
             case token::comment:
-                return mk_stmt<comment_statement>(start_pos, tokens[current++].value);
+                return mk_stmt<comment_statement>(start_pos, next().value);
             case token::text:
-                return mk_stmt<string_literal>(start_pos, tokens[current++].value);
+                return mk_stmt<string_literal>(start_pos, next().value);
             case token::open_statement:
                 return parse_jinja_statement();
             case token::open_expression:
@@ -119,8 +126,7 @@ private:
         }
 
         size_t start_pos = current;
-        std::string name = peek().value;
-        current++; // consume identifier
+        std::string name = next().value;
 
         statement_ptr result;
         if (name == "set") {
@@ -202,7 +208,7 @@ private:
             // Ignore generation blocks (transformers-specific)
             // See https://github.com/huggingface/transformers/pull/30650 for more information.
             result = mk_stmt<noop_statement>(start_pos);
-            current++;
+            ++current;
 
         } else {
             throw std::runtime_error("Unknown statement: " + name);
@@ -217,7 +223,7 @@ private:
         statements body;
 
         if (is(token::equals)) {
-            current++;
+            ++current;
             value = parse_expression_sequence();
         } else {
             // parsing multiline set here
@@ -280,7 +286,7 @@ private:
         exprs.push_back(primary ? parse_primary_expression() : parse_expression());
         bool is_tuple = is(token::comma);
         while (is(token::comma)) {
-            current++; // consume comma
+            ++current; // consume comma
             exprs.push_back(primary ? parse_primary_expression() : parse_expression());
         }
         return is_tuple ? mk_stmt<tuple_literal>(start_pos, std::move(exprs)) : std::move(exprs[0]);
@@ -290,7 +296,7 @@ private:
         // e.g., `message` in `for message in messages`
         auto loop_var = parse_expression_sequence(true); // should be an identifier/tuple
         if (!is_identifier("in")) throw std::runtime_error("Expected 'in'");
-        current++;
+        ++current; // consume 'in'
 
         // `messages` in `for message in messages`
         auto iterable = parse_expression();
@@ -305,7 +311,8 @@ private:
         }
 
         if (is_statement({"else"})) {
-            current += 2;
+            ++current; // consume {%
+            ++current; // consume 'else'
             expect(token::close_statement, "Expected %}");
             while (!is_statement({"endfor"})) {
                 alternate.push_back(parse_any());
@@ -347,7 +354,7 @@ private:
         auto left = parse_logical_and_expression();
         while (is_identifier("or")) {
             size_t start_pos = current;
-            token op = tokens[current++];
+            token op = next();
             left = mk_stmt<binary_expression>(start_pos, op, std::move(left), parse_logical_and_expression());
         }
         return left;
@@ -357,7 +364,7 @@ private:
         auto left = parse_logical_negation_expression();
         while (is_identifier("and")) {
             size_t start_pos = current;
-            auto op = tokens[current++];
+            auto op = next();
             left = mk_stmt<binary_expression>(start_pos, op, std::move(left), parse_logical_negation_expression());
         }
         return left;
@@ -367,7 +374,7 @@ private:
         // Try parse unary operators
         if (is_identifier("not")) {
             size_t start_pos = current;
-            auto op = tokens[current++];
+            auto op = next();
             return mk_stmt<unary_expression>(start_pos, op, parse_logical_negation_expression());
         }
         return parse_comparison_expression();
@@ -382,11 +389,12 @@ private:
             size_t start_pos = current;
             if (is_identifier("not") && peek(1).t == token::identifier && peek(1).value == "in") {
                 op = {token::identifier, "not in", tokens[current].pos};
-                current += 2;
+                ++current; // consume 'not'
+                ++current; // consume 'in'
             } else if (is_identifier("in")) {
-                op = tokens[current++];
+                op = next();
             } else if (is(token::comparison_binary_operator)) {
-                op = tokens[current++];
+                op = next();
             } else break;
             left = mk_stmt<binary_expression>(start_pos, op, std::move(left), parse_additive_expression());
         }
@@ -397,7 +405,7 @@ private:
         auto left = parse_multiplicative_expression();
         while (is(token::additive_binary_operator)) {
             size_t start_pos = current;
-            auto op = tokens[current++];
+            auto op = next();
             left = mk_stmt<binary_expression>(start_pos, op, std::move(left), parse_multiplicative_expression());
         }
         return left;
@@ -407,7 +415,7 @@ private:
         auto left = parse_test_expression();
         while (is(token::multiplicative_binary_operator)) {
             size_t start_pos = current;
-            auto op = tokens[current++];
+            auto op = next();
             left = mk_stmt<binary_expression>(start_pos, op, std::move(left), parse_test_expression());
         }
         return left;
@@ -417,9 +425,9 @@ private:
         auto operand = parse_filter_expression();
         while (is_identifier("is")) {
             size_t start_pos = current;
-            current++;
+            ++current; // consume 'is'
             bool negate = false;
-            if (is_identifier("not")) { current++; negate = true; }
+            if (is_identifier("not")) { ++current; negate = true; }
             auto test_id = parse_primary_expression();
             // FIXME: tests can also be expressed like this: if x is eq 3
             if (is(token::open_paren)) test_id = parse_call_expression(std::move(test_id));
@@ -432,7 +440,7 @@ private:
         auto operand = parse_call_member_expression();
         while (is(token::pipe)) {
             size_t start_pos = current;
-            current++;
+            ++current; // consume pipe
             auto filter = parse_primary_expression();
             if (is(token::open_paren)) filter = parse_call_expression(std::move(filter));
             operand = mk_stmt<filter_expression>(start_pos, std::move(operand), std::move(filter));
@@ -490,7 +498,7 @@ private:
     statement_ptr parse_member_expression(statement_ptr object) {
         size_t start_pos = current;
         while (is(token::dot) || is(token::open_square_bracket)) {
-            auto op = tokens[current++];
+            auto op = next();
             bool computed = op.t == token::open_square_bracket;
             statement_ptr prop;
             if (computed) {
@@ -536,7 +544,7 @@ private:
 
     statement_ptr parse_primary_expression() {
         size_t start_pos = current;
-        auto t = tokens[current++];
+        auto t = next();
         switch (t.t) {
             case token::numeric_literal:
                 if (t.value.find('.') != std::string::npos) {
@@ -547,7 +555,7 @@ private:
             case token::string_literal: {
                 std::string val = t.value;
                 while (is(token::string_literal)) {
-                    val += tokens[current++].value;
+                    val += next().value;
                 }
                 return mk_stmt<string_literal>(start_pos, val);
             }
@@ -562,9 +570,9 @@ private:
                 statements vals;
                 while (!is(token::close_square_bracket)) {
                     vals.push_back(parse_expression());
-                    if (is(token::comma)) current++;
+                    if (is(token::comma)) ++current;
                 }
-                current++;
+                ++current;
                 return mk_stmt<array_literal>(start_pos, std::move(vals));
             }
             case token::open_curly_bracket: {
@@ -573,9 +581,9 @@ private:
                     auto key = parse_expression();
                     expect(token::colon, "Expected :");
                     pairs.push_back({std::move(key), parse_expression()});
-                    if (is(token::comma)) current++;
+                    if (is(token::comma)) ++current;
                 }
-                current++;
+                ++current;
                 return mk_stmt<object_literal>(start_pos, std::move(pairs));
             }
             default:

common/jinja/runtime.cpp

@@ -667,8 +667,9 @@ value macro_statement::execute_impl(context & ctx) {
                 if (is_stmt<identifier>(this->args[i])) {
                     // normal parameter
                     std::string param_name = cast_stmt<identifier>(this->args[i])->val;
-                    JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), args.get_pos(i)->type().c_str());
-                    macro_ctx.set_val(param_name, args.get_pos(i));
+                    value param_value = args.get_kwarg_or_pos(param_name, i);
+                    JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), param_value->type().c_str());
+                    macro_ctx.set_val(param_name, param_value);
                 } else if (is_stmt<keyword_argument_expression>(this->args[i])) {
                     // default argument used as normal parameter
                     auto kwarg = cast_stmt<keyword_argument_expression>(this->args[i]);
@@ -676,8 +677,9 @@ value macro_statement::execute_impl(context & ctx) {
                         throw std::runtime_error("Keyword argument key must be an identifier in macro '" + name + "'");
                     }
                     std::string param_name = cast_stmt<identifier>(kwarg->key)->val;
-                    JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), args.get_pos(i)->type().c_str());
-                    macro_ctx.set_val(param_name, args.get_pos(i));
+                    value param_value = args.get_kwarg_or_pos(param_name, i);
+                    JJ_DEBUG("  Binding parameter '%s' to argument of type %s", param_name.c_str(), param_value->type().c_str());
+                    macro_ctx.set_val(param_name, param_value);
                 } else {
                     throw std::runtime_error("Invalid parameter type in macro '" + name + "'");
                 }

convert_hf_to_gguf.py

@@ -947,6 +947,9 @@ class ModelBase:
         if "thinker_config" in config:
             # rename for Qwen2.5-Omni
             config["text_config"] = config["thinker_config"]["text_config"]
+        if "language_config" in config:
+            # rename for DeepSeekOCR
+            config["text_config"] = config["language_config"]
         if "lfm" in config:
             # rename for LFM2-Audio
             config["text_config"] = config["lfm"]
@@ -1503,6 +1506,9 @@ class TextModel(ModelBase):
         if chkhsh == "e4d54df1ebc1f2b91acd986c5b51aa50837d5faf7c7398e73c1f9e9ee5d19869":
             # ref: https://huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct-2601
             res = "kanana2"
+        if chkhsh == "862f827721df956049dff5ca81a57f29e575280bc622e290d3bf4e35eca29015":
+            # ref: https://huggingface.co/codefuse-ai/F2LLM-v2-4B
+            res = "f2llmv2"
 
         if res is None:
             logger.warning("\n")
@@ -2071,7 +2077,7 @@ class MmprojModel(ModelBase):
     preprocessor_config: dict[str, Any]
     global_config: dict[str, Any]
 
-    n_block_keys = ["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth", "encoder_layers", "vt_num_hidden_layers"]
+    n_block_keys = ["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth", "layers", "encoder_layers", "vt_num_hidden_layers"]
 
     has_vision_encoder: bool = True # by default
     has_audio_encoder: bool = False
@@ -4273,6 +4279,16 @@ class Qwen25OmniModel(Qwen2VLVisionModel):
 
 @ModelBase.register("InternVisionModel")
 class InternVisionModel(MmprojModel):
+
+    min_dynamic_tiles: int = 0
+    max_dynamic_tiles: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.min_dynamic_tiles = self.global_config.get("min_dynamic_patch", 0)
+        self.max_dynamic_tiles = self.global_config.get("max_dynamic_patch", 0)
+
     def set_gguf_parameters(self):
         assert self.hparams_vision is not None
         if isinstance(self.hparams_vision['image_size'], list):
@@ -4295,6 +4311,11 @@ class InternVisionModel(MmprojModel):
         downsample_ratio = self.global_config.get("downsample_ratio")
         assert downsample_ratio is not None
         self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
+        # older models may not have min/max_dynamic_patch in config
+        if self.min_dynamic_tiles > 0:
+            self.gguf_writer.add_vision_preproc_min_tiles(self.min_dynamic_tiles)
+        if self.max_dynamic_tiles > 0:
+            self.gguf_writer.add_vision_preproc_max_tiles(self.max_dynamic_tiles)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         if ".position_embd." in new_name:
@@ -4557,7 +4578,7 @@ class Qwen2MoeModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("Qwen3ForCausalLM")
+@ModelBase.register("Qwen3ForCausalLM", "Qwen3Model")
 class Qwen3Model(Qwen2Model):
     model_arch = gguf.MODEL_ARCH.QWEN3
 
@@ -6920,6 +6941,68 @@ class ConformerAudioModel(MmprojModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("DeepseekOCRForCausalLM")
+class DeepseekOCRVisionModel(MmprojModel):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DEEPSEEKOCR)
+        # default values below are taken from HF tranformers code
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_vision_use_gelu(True)
+        # calculate proj_scale_factor (used by tinygemma3 test model)
+        image_seq_length = self.preprocessor_config.get("image_seq_length", 256)
+        n_per_side = int(image_seq_length ** 0.5)
+        image_size = self.hparams["image_size"]
+        patch_size = self.hparams["patch_size"]
+        proj_scale_factor = (image_size // patch_size) // n_per_side
+        if proj_scale_factor > 0 and proj_scale_factor != 4:
+            # we only need to write this if it's not the default value
+            # in this case, we are converting a test model
+            self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)
+        # @bluebread: there's no window_size in config but just add it here anyway
+        self.gguf_writer.add_vision_window_size(self.hparams.get("window_size", 14))
+
+        # SAM configuration
+        sam_hparams = hparams['sam']
+        self.gguf_writer.add_vision_sam_layers_count(sam_hparams['layers'])
+        self.gguf_writer.add_vision_sam_embedding_length(sam_hparams['width'])
+        self.gguf_writer.add_vision_sam_head_count(sam_hparams['heads'])
+
+    def get_vision_config(self) -> dict[str, Any]:
+        vision_config: dict[str, Any] | None = self.global_config.get("vision_config")
+
+        if not vision_config:
+            raise ValueError("DeepseekOCR model requires 'vision_config' in the model configuration, but it was not found")
+
+        vision_config['sam'] = vision_config['width']['sam_vit_b']
+        vision_config.update(vision_config['width']['clip-l-14-224'])
+        vision_config['hidden_size'] = vision_config['width']
+        vision_config['num_heads'] = vision_config['heads']
+        vision_config['intermediate_size'] = vision_config['heads'] * 4
+
+        return vision_config
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if ".embeddings." in name or 'pos_embed' in name:
+            return gguf.GGMLQuantizationType.F32
+        if ".rel_pos_h" in name or '.rel_pos_w' in name:
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Only process vision-related tensors, skip language model tensors
+        # Vision components: sam_model, vision_model, projector, image_newline, view_seperator
+        # Language model components to skip: lm_head, embed_tokens, layers, norm
+        if name.startswith(("lm_head.", "model.embed_tokens.", "model.layers.", "model.norm.")):
+            return
+
+        if name.endswith("pos_embed") or name.endswith("rel_pos_h") or name.endswith("rel_pos_w"):
+            name += ".weight"
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Gemma3nForConditionalGeneration")
 class Gemma3nVisionAudioModel(ConformerAudioModel):
     has_audio_encoder = True
@@ -8265,6 +8348,19 @@ class DeepseekV2Model(TextModel):
 
     merge_expert = True
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hparams: dict = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
+        self.origin_hf_arch = hparams.get('architectures', [None])[0]
+
+        # special handling for Deepseek OCR
+        if self.origin_hf_arch == "DeepseekOCRForCausalLM":
+            self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
+            self.gguf_writer.add_architecture()
+            # default jinja template
+            self.gguf_writer.add_chat_template("{% for m in messages %}{{m['content']}}{% endfor %}")
+
     def set_vocab(self):
         try:
             self._set_vocab_gpt2()
@@ -8320,9 +8416,15 @@ class DeepseekV2Model(TextModel):
             raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
 
     def set_gguf_parameters(self):
+        is_ocr = (self.model_arch == gguf.MODEL_ARCH.DEEPSEEK2OCR)
 
-        # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
-        self.hparams["num_key_value_heads"] = 1
+        if is_ocr:
+            self.hparams['rope_theta'] = self.hparams.get('rope_theta', 10000.0)
+        else:
+            # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
+            self.hparams["num_key_value_heads"] = 1
+
+        self.hparams['rms_norm_eps'] = self.hparams.get('rms_norm_eps', 1e-6)
 
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -8336,16 +8438,18 @@ class DeepseekV2Model(TextModel):
             # Default: if no MoE, all layers are dense; if MoE, none are dense
             first_k_dense_replace = hparams["num_hidden_layers"] if not has_moe else 0
         self.gguf_writer.add_leading_dense_block_count(first_k_dense_replace)
+        kv_lora_rank = hparams.get("kv_lora_rank", 512)
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
         if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
             self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
-        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
 
         # note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
-        self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"])
-        self.gguf_writer.add_value_length(hparams["kv_lora_rank"])
-        self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
-        self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
+        if not is_ocr:
+            self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
+            self.gguf_writer.add_key_length(kv_lora_rank + hparams["qk_rope_head_dim"])
+            self.gguf_writer.add_value_length(kv_lora_rank)
+            self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+            self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
 
         # MoE parameters (required by C++ code for DEEPSEEK2 arch)
         # For non-MoE models like Youtu, use intermediate_size as expert_feed_forward_length
@@ -8377,8 +8481,15 @@ class DeepseekV2Model(TextModel):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # skip vision tensors and remove "language_model." for Kimi-VL and Kimi-K2.5
-        if "vision_tower" in name or "multi_modal_projector" in name or "mm_projector" in name:
+        # skip vision tensors and remove "language_model." for Kimi-VL and Kimi-K2.5, and DeepSeek-OCR
+        if ("vision_tower" in name
+                or "multi_modal_projector" in name
+                or "mm_projector" in name
+                or "vision_model" in name
+                or "image_newline" in name
+                or "model.projector" in name
+                or "sam_model" in name
+                or "view_seperator" in name):
             return
         if name.startswith("siglip2.") or name.startswith("merger."):
             return

convert_hf_to_gguf_update.py

@@ -154,6 +154,7 @@ models = [
     {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", },
     {"name": "joyai-llm",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jdopensource/JoyAI-LLM-Flash", },
     {"name": "kanana2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct-2601", },
+    {"name": "f2llmv2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/codefuse-ai/F2LLM-v2-4B", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/backend/OPENVINO.md

@@ -1,6 +1,9 @@
 # OpenVINO Backend for llama.cpp
-[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge.
-This document describes the [OpenVINO backend for llama.cpp](../../src/ggml-openvino), which enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
+
+> [!NOTE]
+> Performance and memory optimizations, accuracy validation, broader quantization coverage, broader operator and model support are work in progress.
+
+[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge. [OpenVINO backend for llama.cpp](../../src/ggml-openvino) enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
 
 The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a translation layer for core GGML operations. The OpenVINO backend replaces the standard GGML graph execution path with Intel's OpenVINO inference engine. This approach allows the same GGUF model file to run on Intel CPUs, Intel GPUs (integrated and discrete), and Intel NPUs without changes to the model or the rest of the llama.cpp stack. When a `ggml_cgraph` is dispatched to OpenVINO backend, it:
 
@@ -179,31 +182,73 @@ curl -L https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/main/L
 
 When using the OpenVINO backend, the first inference token may have slightly higher latency due to on-the-fly conversion to the OpenVINO graph. Subsequent tokens and runs will be faster.
 
+> [!NOTE]
+> Default context size is set to the model training context, which may be very large. For example, 131072 for Llama 3.2 1B, which may result in lower performance, especially on edge/laptop devices. Use `-c` to limit context size in supported llama.cpp tools for better performance. For example, `-c 512`.
+
 ```bash
 # If device is unset or unavailable, defaults to CPU.
 # If the system has multiple GPUs, use GPU.0 or GPU.1 to explicitly target a specific GPU.
 
 # Linux
 export GGML_OPENVINO_DEVICE=GPU
+# Enable stateful execution with GPU device to avoid known stateless execution failures.
+export GGML_OPENVINO_STATEFUL_EXECUTION=1
 # To run llama-simple:
 ./build/ReleaseOV/bin/llama-simple -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -n 50 "The story of AI is "
 # To run in chat mode:
-./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf
+./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -c 1024
+# To run llama-bench, -fa 1 is needed
+GGML_OPENVINO_STATEFUL_EXECUTION=1 GGML_OPENVINO_DEVICE=GPU ./build/ReleaseOV/bin/llama-bench -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -fa 1
+
+# NPU: keep context small to avoid failures from very large model context windows.
+export GGML_OPENVINO_DEVICE=NPU
+./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -c 512
 
 # Windows Command Line
 set GGML_OPENVINO_DEVICE=GPU
+# Enable stateful execution with GPU device to avoid known stateless execution failures.
+set GGML_OPENVINO_STATEFUL_EXECUTION=1
 # Windows PowerShell
 $env:GGML_OPENVINO_DEVICE = "GPU"
+$env:GGML_OPENVINO_STATEFUL_EXECUTION = "1"
 
 # To run llama-simple
 build\ReleaseOV\bin\llama-simple.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -n 50 "The story of AI is "
 # To run in chat mode:
-build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf"
+build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -c 1024
+# To run llama-bench, -fa 1 is needed
+build\ReleaseOV\bin\llama-bench.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -fa 1
 
+# NPU: keep context small to avoid failures from very large model context windows.
+# Windows Command Line
+set GGML_OPENVINO_DEVICE=NPU
+# Windows PowerShell
+$env:GGML_OPENVINO_DEVICE = "NPU"
+build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -c 512
 ```
 > [!NOTE]
 > On systems with multiple GPUs, use `GPU.0` or `GPU.1` to explicitly target specific GPU. See [OpenVINO GPU Device](https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/gpu-device.html) for more details.
 
+### Known Issues and Current Workarounds
+
+- GPU stateless execution is currently affected by a known issue.
+  - Workaround: set `GGML_OPENVINO_STATEFUL_EXECUTION=1` when using GPU device.
+- NPU failures can happen when context size is too large. Recent llama.cpp behavior may resolve context size to the model training context (for example, 131072 for Llama 3.2 1B), which is too large for current NPU usage and can also stress laptop CPU/GPU on larger models. To inspect the selected context size, run `llama-cli` or `llama-server` with `-lv 3`.
+  - Workaround: explicitly set context size, for ex. `-c 1024` for NPU runs. Performance will be better with lower context size.
+- Additional NPU limitations:
+  - Model caching is not yet supported.
+  - `llama-server -np > 1` (multiple parallel sequences) is not supported.
+  - `llama-perplexity` is only supported with `-b 512` or smaller.
+- `--context-shift` with `llama-cli` is currently not supported with OpenVINO backend across CPU, GPU, and NPU devices.
+- Encoder models (embedding, reranking) are not supported with the current OpenVINO backend implementation.
+- `-fa 1` is required when running llama-bench with the OpenVINO backend.
+  - `GGML_OPENVINO_STATEFUL_EXECUTION=1 GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1`
+- `llama-server` with OpenVINO backend supports only one chat session/thread, when `GGML_OPENVINO_STATEFUL_EXECUTION=1` is enabled.
+- For Intel GPU, NPU detection in containers, GPU, NPU user-space drivers/libraries must be present inside the image. We will include in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile)
+
+> [!NOTE]
+> The OpenVINO backend is actively under development. Fixes are underway, and this document will continue to be updated as issues are resolved.
+
 
 ### Docker Build
 
@@ -229,31 +274,42 @@ docker build --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_p
 Run llama.cpp with OpenVINO backend Docker container.
 Save sample models in `~/models` as [shown above](#3-download-sample-model). It will be mounted to the container in the examples below.
 
+> [!NOTE]
+> Intel GPU, NPU detection in containers will be included in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile).
+
 ```bash
 #  Run Docker container
-docker run --rm -it -v ~/models:/models llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
+docker run --rm -it -v ~/models:/models llama-openvino:light --no-warmup -c 1024 -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
 
 # With Intel GPU access (iGPU or dGPU)
 docker run --rm -it -v ~/models:/models \
 --device=/dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -u $(id -u):$(id -g) \
-llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
+--env=GGML_OPENVINO_DEVICE=GPU --env=GGML_OPENVINO_STATEFUL_EXECUTION=1 \
+llama-openvino:light --no-warmup -c 1024 -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
 
 # With Intel NPU access
-docker run --rm -it --env GGML_OPENVINO_DEVICE=NPU -v ~/models:/models \
+docker run --rm -it -v ~/models:/models \
 --device=/dev/accel --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -u $(id -u):$(id -g) \
-llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
+--env=GGML_OPENVINO_DEVICE=NPU \
+llama-openvino:light --no-warmup -c 1024 -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
 ```
 
-Run Llama.cpp Server with OpenVINO Backend:
+Run Llama.cpp Server with OpenVINO Backend.
+> [!NOTE]
+> `llama-server` with OpenVINO backend supports only one chat session/thread, when `GGML_OPENVINO_STATEFUL_EXECUTION=1` is enabled.
+
 ```bash
 # Run the Server Docker container
-docker run --rm -it -p 8080:8080 -v ~/models:/models llama-openvino:server --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
-
-# In a NEW terminal, test the server with curl
+docker run --rm -it -p 8080:8080 -v ~/models:/models llama-openvino:server --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf -c 1024
+# Or Using llama-server executable
+./build/ReleaseOV/bin/llama-server -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf --port 8080 -c 1024
 
 # If you are behind a proxy, make sure to set NO_PROXY to avoid proxy for localhost
 export NO_PROXY=localhost,127.0.0.1
 
+# Option 1: Open your browser to http://localhost:8080 to access the web UI for the llama.cpp server.
+# Option 2: In a NEW terminal, test the server with curl
+
 # Test health endpoint
 curl -f http://localhost:8080/health
 
@@ -295,6 +351,7 @@ The OpenVINO backend can be configured using the following environment variables
 export GGML_OPENVINO_CACHE_DIR=/tmp/ov_cache
 export GGML_OPENVINO_PROFILING=1
 export GGML_OPENVINO_DEVICE=GPU
+export GGML_OPENVINO_STATEFUL_EXECUTION=1
 
 ./build/ReleaseOV/bin/llama-simple -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -n 50 "The story of AI is "
 
@@ -302,38 +359,27 @@ export GGML_OPENVINO_DEVICE=GPU
 set GGML_OPENVINO_CACHE_DIR=C:\tmp\ov_cache
 set GGML_OPENVINO_PROFILING=1
 set GGML_OPENVINO_DEVICE=GPU
+set GGML_OPENVINO_STATEFUL_EXECUTION=1
 
 # Windows PowerShell
 $env:GGML_OPENVINO_CACHE_DIR = "C:\tmp\ov_cache"
 $env:GGML_OPENVINO_PROFILING = "1"
 $env:GGML_OPENVINO_DEVICE = "GPU"
+$env:GGML_OPENVINO_STATEFUL_EXECUTION = "1"
 
 build\ReleaseOV\bin\llama-simple.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -n 50 "The story of AI is "
 
 ```
 
-#### llama-bench
-
-```bash
-# -fa 1 is required when running llama-bench with the OpenVINO backend.
-GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1
-```
-
-### NPU Notes
-
-- Model caching is not yet supported
-- Does not support llama-server -np > 1 (multiple parallel sequences)
-- Only supports llama-perplexity -b 512 or smaller
-
 ## Llama.cpp Tools
 
 The following tools work with the OpenVINO backend on CPU, GPU, NPU:
-- llama-simple
-- llama-run
-- llama-cli
-- llama-server
 - llama-bench
+- llama-cli
+- llama-completion
 - llama-perplexity
+- llama-server
+- llama-simple
 
 ## Work in Progress
 

docs/multimodal.md

@@ -31,6 +31,13 @@ llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.g
 llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
 ```
 
+> [!IMPORTANT]
+>
+> OCR models are trained with specific prompt and input structure, please refer to these discussions for more info:
+> - PaddleOCR-VL: https://github.com/ggml-org/llama.cpp/pull/18825
+> - GLM-OCR: https://github.com/ggml-org/llama.cpp/pull/19677
+> - Deepseek-OCR: https://github.com/ggml-org/llama.cpp/pull/17400
+
 ## Pre-quantized models
 
 These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/collections/ggml-org/multimodal-ggufs-68244e01ff1f39e5bebeeedc

examples/llama.android/lib/src/main/cpp/ai_chat.cpp

@@ -365,13 +365,13 @@ Java_com_arm_aichat_internal_InferenceEngineImpl_processSystemPrompt(
     const auto *system_prompt = env->GetStringUTFChars(jsystem_prompt, nullptr);
     LOGd("%s: System prompt received: \n%s", __func__, system_prompt);
     std::string formatted_system_prompt(system_prompt);
-    env->ReleaseStringUTFChars(jsystem_prompt, system_prompt);
 
     // Format system prompt if applicable
     const bool has_chat_template = common_chat_templates_was_explicit(g_chat_templates.get());
     if (has_chat_template) {
         formatted_system_prompt = chat_add_and_format(ROLE_SYSTEM, system_prompt);
     }
+    env->ReleaseStringUTFChars(jsystem_prompt, system_prompt);
 
     // Tokenize system prompt
     const auto system_tokens = common_tokenize(g_context, formatted_system_prompt,
@@ -414,13 +414,13 @@ Java_com_arm_aichat_internal_InferenceEngineImpl_processUserPrompt(
     const auto *const user_prompt = env->GetStringUTFChars(juser_prompt, nullptr);
     LOGd("%s: User prompt received: \n%s", __func__, user_prompt);
     std::string formatted_user_prompt(user_prompt);
-    env->ReleaseStringUTFChars(juser_prompt, user_prompt);
 
     // Format user prompt if applicable
     const bool has_chat_template = common_chat_templates_was_explicit(g_chat_templates.get());
     if (has_chat_template) {
         formatted_user_prompt = chat_add_and_format(ROLE_USER, user_prompt);
     }
+    env->ReleaseStringUTFChars(juser_prompt, user_prompt);
 
     // Decode formatted user prompts
     auto user_tokens = common_tokenize(g_context, formatted_user_prompt, has_chat_template, has_chat_template);

ggml/include/gguf.h

@@ -77,6 +77,7 @@ extern "C" {
     };
 
     GGML_API struct gguf_context * gguf_init_empty(void);
+    GGML_API struct gguf_context * gguf_init_from_file_ptr(FILE * file, struct gguf_init_params params);
     GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);
     //GGML_API struct gguf_context * gguf_init_from_buffer(..);
 
@@ -189,6 +190,7 @@ extern "C" {
     //
 
     // write the entire context to a binary file
+    GGML_API bool gguf_write_to_file_ptr(const struct gguf_context * ctx, FILE * file, bool only_meta);
     GGML_API bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
 
     // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -3011,6 +3011,58 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     }
 }
 
+void ggml_cann_rope_cache_preload(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+
+    float     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int       sections[4];
+    const int n_dims     = ((int32_t *) dst->op_params)[1];
+    const int mode       = ((int32_t *) dst->op_params)[2];
+    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
+    memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
+    memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
+    memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections, (int32_t *) dst->op_params + 11, sizeof(int) * 4);
+
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
+
+    float corr_dims[2];
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    bool       is_neox    = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_imrope  = mode == GGML_ROPE_TYPE_IMROPE;
+    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision  = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_imrope || mrope_used) {
+        is_neox = true;
+    }
+
+    int64_t rope_dims = n_dims;
+    if (is_vision) {
+        rope_dims = src0->ne[0];
+    }
+
+    // Run the full cache init on the non-captured stream.  This performs all
+    // host-to-device memcpy, aclrtMalloc/Free, and on-device computations
+    // so that the memory pool is warmed up and cache metadata is populated.
+    aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections,
+                          mrope_used, is_imrope, is_vision, rope_dims);
+
+    // Reset `cached` so that during graph capture the on-device computations
+    // (sin/cos, position multiply, repeat, etc.) still execute and get recorded
+    // into the captured graph.  The cache metadata (theta_scale_length,
+    // theta_scale, sections, position_length, etc.) remains set, which causes
+    // all host-to-device copy and malloc/free branches to be skipped.
+    ctx.rope_cache.cached = false;
+}
+
 void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src0 = dst->src[0];
 

ggml/src/ggml-cann/aclnn_ops.h

@@ -543,6 +543,21 @@ void ggml_cann_mul_mat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  */
 void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+/**
+ * @brief Pre-load the RoPE cache before ACL graph capture.
+ *
+ * This function must be called outside of graph capture to perform
+ * host-to-device memory copies and device memory allocations that are
+ * not allowed on a captured stream.  After pre-loading, the rope cache
+ * metadata is updated so that the subsequent call to
+ * aclnn_rope_cache_init (inside graph capture) skips these operations
+ * and only records the on-device computations into the captured graph.
+ *
+ * @param ctx  CANN backend context.
+ * @param dst  A ROPE destination tensor from the computation graph.
+ */
+void ggml_cann_rope_cache_preload(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Computes the index of the maximum value along the specified dimension
  *          of a ggml tensor using the CANN backend.

ggml/src/ggml-cann/common.h

@@ -277,7 +277,7 @@ struct ggml_graph_node_properties {
             }
         }
 
-        if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
+        if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU || node->op == GGML_OP_ROPE){
             return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
         }
         return true;

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2225,6 +2225,19 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
             // If no matching graph is found, add a new ACL graph.
             ggml_cann_graph * new_graph = ggml_cann_graph::create_from_cgraph(cgraph);
             cann_ctx->graph_lru_cache.push(new_graph);
+
+            // Pre-load rope cache before graph capture.  During capture the
+            // stream cannot perform host-to-device memcpy or device memory
+            // malloc/free.  Running the full cache init now populates the
+            // cache metadata so these branches are skipped during capture,
+            // while also warming up the memory pool.
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->op == GGML_OP_ROPE) {
+                    ggml_cann_rope_cache_preload(*cann_ctx, node);
+                    break;
+                }
+            }
         }
     }
 #else

ggml/src/ggml-cpu/CMakeLists.txt

@@ -460,6 +460,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
         if(NOT GGML_CPU_ALL_VARIANTS)
             set(MARCH_STR "rv64gc")
+            if (GGML_RVV)
+                string(APPEND MARCH_STR "v")
+            endif()
+
             if (GGML_RV_ZFH)
                 string(APPEND MARCH_STR "_zfh")
             endif()
@@ -467,7 +471,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_XTHEADVECTOR)
                 string(APPEND MARCH_STR "_xtheadvector")
             elseif (GGML_RVV)
-                string(APPEND MARCH_STR "_v")
                 if (GGML_RV_ZVFH)
                     string(APPEND MARCH_STR "_zvfh")
                 endif()
@@ -475,12 +478,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                     string(APPEND MARCH_STR "_zvfbfwma")
                 endif()
             endif()
+
             if (GGML_RV_ZICBOP)
                 string(APPEND MARCH_STR "_zicbop")
             endif()
             if (GGML_RV_ZIHINTPAUSE)
                 string(APPEND MARCH_STR "_zihintpause")
             endif()
+
             list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
         else()
             # Begin with the lowest baseline

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2871,8 +2871,12 @@ struct ggml_cplan ggml_graph_plan(
                         const int64_t ne11 = node->src[1]->ne[1]; // H
                         const int64_t ne12 = node->src[1]->ne[2]; // Channels In
 
-                        cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
-                        cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
+                        GGML_ASSERT(node->src[0]->type == GGML_TYPE_F16 || node->src[0]->type == GGML_TYPE_F32);
+                        GGML_ASSERT(node->src[1]->type == GGML_TYPE_F32);
+
+                        cur += ggml_type_size(node->src[0]->type) * ne00 * ne01 * ne02 * ne03;
+                        cur += ggml_type_size(node->src[0]->type) * ne10 * ne11 * ne12;
+
                     } break;
                 case GGML_OP_TOP_K:
                     {

ggml/src/ggml-cpu/ops.cpp

@@ -6923,16 +6923,15 @@ void ggml_compute_forward_conv_3d(
     ggml_compute_forward_conv_3d_impl(params, src0, src1, dst, src0->type);
 }
 
-// ggml_compute_forward_conv_transpose_2d
-
-void ggml_compute_forward_conv_transpose_2d(
-        const ggml_compute_params * params,
-              ggml_tensor * dst) {
+template <typename kernel_t>
+static void ggml_compute_forward_conv_transpose_2d_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(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -6943,7 +6942,7 @@ void ggml_compute_forward_conv_transpose_2d(
 
     const int nk = ne00*ne01*ne02*ne03;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == ggml_type_size(src0->type));
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (ith == 0) {
@@ -6951,12 +6950,12 @@ void ggml_compute_forward_conv_transpose_2d(
 
         // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout)
         {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
+            kernel_t * const wdata = (kernel_t *) params->wdata + 0;
 
             for (int64_t i03 = 0; i03 < ne03; i03++) {
                 for (int64_t i02 = 0; i02 < ne02; i02++) {
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02);
-                    ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03;
+                    const kernel_t * const src = (kernel_t *)((char *) src0->data + i03*nb03 + i02*nb02);
+                    kernel_t * dst_data = wdata + i02*ne01*ne00*ne03;
                     for (int64_t i01 = 0; i01 < ne01; i01++) {
                         for (int64_t i00 = 0; i00 < ne00; i00++) {
                             dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00];
@@ -6968,13 +6967,17 @@ void ggml_compute_forward_conv_transpose_2d(
 
         // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh)
         {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk;
+            kernel_t * const wdata = (kernel_t *) params->wdata + nk;
             for (int i12 = 0; i12 < ne12; i12++) {
                 for (int i11 = 0; i11 < ne11; i11++) {
                     const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
-                    ggml_fp16_t * dst_data = wdata + i11*ne10*ne12;
+                    kernel_t * dst_data = wdata + i11*ne10*ne12;
                     for (int i10 = 0; i10 < ne10; i10++) {
-                        dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
+                        if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
+                            dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
+                        } else {
+                            dst_data[i10*ne12 + i12] = src[i10];
+                        }
                     }
                 }
             }
@@ -6996,21 +6999,27 @@ void ggml_compute_forward_conv_transpose_2d(
     const int ip0 = dp*ith;
     const int ip1 = MIN(ip0 + dp, np);
 
-    ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
-    ggml_fp16_t * const wdata_src = wdata + nk;
+    kernel_t * const wdata = (kernel_t *) params->wdata + 0;
+    kernel_t * const wdata_src = wdata + nk;
 
     for (int i2 = ip0; i2 < ip1; i2++) { // Cout
         float * dst_data = (float *)((char *) dst->data + i2*nb2);
-        ggml_fp16_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
+        kernel_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
         for (int i11 = 0; i11 < ne11; i11++) {
             for (int i10 = 0; i10 < ne10; i10++) {
                 const int i1n = i11*ne10*ne12 + i10*ne12;
                 for (int i01 = 0; i01 < ne01; i01++) {
                     for (int i00 = 0; i00 < ne00; i00++) {
                         float v = 0;
-                        ggml_vec_dot_f16(ne03, &v, 0,
-                                wdata_src + i1n, 0,
-                                wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
+                            ggml_vec_dot_f16(ne03, &v, 0,
+                                    wdata_src + i1n, 0,
+                                    wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        } else {
+                            ggml_vec_dot_f32(ne03, &v, 0,
+                                    wdata_src + i1n, 0,
+                                    wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        }
                         dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
                     }
                 }
@@ -7019,6 +7028,28 @@ void ggml_compute_forward_conv_transpose_2d(
     }
 }
 
+void ggml_compute_forward_conv_transpose_2d(
+        const ggml_compute_params * params,
+              ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_transpose_2d_impl<ggml_fp16_t>(params, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_conv_transpose_2d_impl<float>(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_conv_2d_dw
 
 struct ggml_conv_2d_dw_params {

ggml/src/ggml-cuda/CMakeLists.txt

@@ -116,12 +116,11 @@ if (CUDAToolkit_FOUND)
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
         add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
     else()
-        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        list(APPEND GGML_SOURCES_CUDA
+            template-instances/fattn-vec-instance-f16-f16.cu
+            template-instances/fattn-vec-instance-q4_0-q4_0.cu
+            template-instances/fattn-vec-instance-q8_0-q8_0.cu
+            template-instances/fattn-vec-instance-bf16-bf16.cu)
     endif()
 
     ggml_add_backend_library(ggml-cuda

ggml/src/ggml-cuda/common.cuh

@@ -799,6 +799,16 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 #endif // CUDART_VERSION >= 12050
 }
 
+static __device__ __forceinline__ float ggml_cuda_ue4m3_to_fp32(uint8_t x) {
+#ifdef FP8_AVAILABLE
+    const uint32_t bits = x * (x != 0x7F && x != 0xFF); // Convert NaN to 0.0f to match CPU implementation.
+    const __nv_fp8_e4m3 xf = *reinterpret_cast<const __nv_fp8_e4m3 *>(&bits);
+    return static_cast<float>(xf) / 2;
+#else
+    NO_DEVICE_CODE;
+#endif // FP8_AVAILABLE
+}
+
 __device__ __forceinline__ uint8_t ggml_cuda_float_to_fp4_e2m1(float x, float e) {
     const uint8_t sign_bit = (x < 0.0f) << 3;
     float         ax       = fabsf(x) * e;
@@ -931,6 +941,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_MXFP4> {
     static constexpr int qi = QI_MXFP4;
 };
 
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_NVFP4> {
+    static constexpr int qk = QK_NVFP4;
+    static constexpr int qr = QR_NVFP4;
+    static constexpr int qi = QI_NVFP4;
+};
+
 template<>
 struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
     static constexpr int qk = QK_K;

ggml/src/ggml-cuda/conv2d-transpose.cu

@@ -1,12 +1,20 @@
-#include <algorithm>
-
 #include "conv2d-transpose.cuh"
-#include "ggml.h"
+#include "convert.cuh"
 
-__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel,
-                                        float * __restrict__ output, const int in_w, const int in_h, const int out_w,
-                                        const int out_h, const int kernel_w, const int kernel_h, const int stride,
-                                        const int c_in, const int c_out, const int batches) {
+template <typename kernel_t>
+static __global__ void conv2d_transpose_kernel(const float * __restrict__ input,
+                                               const kernel_t * __restrict__ kernel,
+                                               float * __restrict__ output,
+                                               const int in_w,
+                                               const int in_h,
+                                               const int out_w,
+                                               const int out_h,
+                                               const int kernel_w,
+                                               const int kernel_h,
+                                               const int stride,
+                                               const int c_in,
+                                               const int c_out,
+                                               const int batches) {
     const int global_idx = blockIdx.x * blockDim.x + threadIdx.x;
 
     const int total_elements = out_w * out_h * c_out * batches;
@@ -26,24 +34,32 @@ __global__ void conv2d_transpose_kernel(const float * __restrict__ input, const
     for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) {
         for (int kh = 0; kh < kernel_h; ++kh) {
             int in_y = out_y_idx - kh;
-            if (in_y < 0 || in_y % stride) continue;
+            if (in_y < 0 || in_y % stride) {
+                continue;
+            }
             in_y /= stride;
-            if (in_y >= in_h) continue;
+            if (in_y >= in_h) {
+                continue;
+            }
 
             for (int kw = 0; kw < kernel_w; ++kw) {
                 int in_x = out_x_idx - kw;
-                if (in_x < 0 || in_x % stride) continue;
+                if (in_x < 0 || in_x % stride) {
+                    continue;
+                }
                 in_x /= stride;
-                if (in_x >= in_w) continue;
+                if (in_x >= in_w) {
+                    continue;
+                }
 
                 const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x;
                 const int kernel_idx =
                     (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw;
 
-                float input_val = input[input_idx];
-                half  kern_val  = kernel[kernel_idx];
+                float    input_val = input[input_idx];
+                kernel_t kern_val  = kernel[kernel_idx];
 
-                accumulator += input_val * (float) kern_val;
+                accumulator += input_val * ggml_cuda_cast<float>(kern_val);
             }
         }
     }
@@ -56,11 +72,12 @@ void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor
     const ggml_tensor * kernel = dst->src[0];
     const ggml_tensor * input  = dst->src[1];
 
-    GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
 
     const float * input_data  = (const float *) input->data;
     float *       output_data = (float *) dst->data;
-    const half * kernel_data = (const half *) kernel->data;
+    const void *  kernel_data = kernel->data;
 
     const int input_w      = input->ne[0];
     const int input_h      = input->ne[1];
@@ -82,10 +99,17 @@ void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor
     GGML_ASSERT(ggml_is_contiguous(kernel));
     GGML_ASSERT(ggml_is_contiguous(dst));
 
-    const int total  = (output_w * output_h * channels_out * batches);
+    const int total  = output_w * output_h * channels_out * batches;
     const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE;
 
-    conv2d_transpose_kernel<<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
-        input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride,
-        channels_in, channels_out, batches);
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_transpose_kernel<half><<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+            input_data, (const half *) kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w,
+            kernel_h, stride, channels_in, channels_out, batches);
+
+    } else {
+        conv2d_transpose_kernel<float><<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+            input_data, (const float *) kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w,
+            kernel_h, stride, channels_in, channels_out, batches);
+    }
 }

ggml/src/ggml-cuda/conv2d-transpose.cuh

@@ -1,4 +1,5 @@
 #include "common.cuh"
 
 #define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+
 void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/convert.cu

@@ -617,6 +617,45 @@ static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t
     dequantize_block_mxfp4<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template <typename dst_t>
+static __global__ void dequantize_block_nvfp4(
+        const void * __restrict__ vx,
+        dst_t * __restrict__ yy,
+        const int64_t ne) {
+    const int64_t i = blockIdx.x;
+    const int     tid = threadIdx.x;
+
+    const int64_t base = i * QK_NVFP4;
+    if (base >= ne) {
+        return;
+    }
+
+    const block_nvfp4 * x = (const block_nvfp4 *) vx;
+    const block_nvfp4 & xb = x[i];
+
+    const int sub = tid / (QK_NVFP4_SUB / 2);
+    const int j = tid % (QK_NVFP4_SUB / 2);
+
+    const float d = ggml_cuda_ue4m3_to_fp32(xb.d[sub]);
+    const uint8_t q = xb.qs[sub * (QK_NVFP4_SUB / 2) + j];
+
+    const int64_t y0 = base + sub * QK_NVFP4_SUB + j;
+    const int64_t y1 = y0 + QK_NVFP4_SUB / 2;
+
+    yy[y0] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q & 0x0F]);
+    yy[y1] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q >> 4]);
+}
+
+template <typename dst_t>
+static void dequantize_row_nvfp4_cuda(
+        const void * vx,
+        dst_t * y,
+        const int64_t k,
+        cudaStream_t stream) {
+    GGML_ASSERT(k % QK_NVFP4 == 0);
+    const int nb = k / QK_NVFP4;
+    dequantize_block_nvfp4<<<nb, 32, 0, stream>>>(vx, y, k);
+}
 template <typename src_t, typename dst_t>
 static __global__ void convert_unary(
         const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
@@ -715,6 +754,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_MXFP4:
             return dequantize_row_mxfp4_cuda;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
@@ -766,6 +807,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_MXFP4:
             return dequantize_row_mxfp4_cuda;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cont_cuda<half>;
         case GGML_TYPE_BF16:

ggml/src/ggml-cuda/convert.cuh

@@ -41,6 +41,16 @@ template<typename dst_t, typename src_t>
         return __bfloat162float(x);
     } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, half2>) {
         return __float22half2_rn(x);
+    } else if constexpr(std::is_same_v<src_t, nv_bfloat162> && std::is_same_v<dst_t, float2>) {
+#ifdef GGML_USE_HIP
+        return make_float2(__bfloat162float(__low2bfloat16(x)), __bfloat162float(__high2bfloat16(x)));
+#else
+#if __CUDA_ARCH__ >= 800
+        return __bfloat1622float2(x);
+#else
+        return make_float2(__bfloat162float(x.x), __bfloat162float(x.y));
+#endif // __CUDA_ARCH__ >= 800
+#endif // GGML_USE_HIP
     } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, nv_bfloat162>) {
         // bypass compile error on cuda 12.0.1
 #ifdef GGML_USE_HIP

ggml/src/ggml-cuda/fattn-common.cuh

@@ -74,6 +74,37 @@ static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f16(
     return sum;
 }
 
+template <int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_bf16(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) {
+
+    const nv_bfloat162 * K_bf16 = (const nv_bfloat162 *) K_c;
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+
+    constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
+    constexpr int cpy_ne = cpy_nb / 4;
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += nthreads*cpy_ne) {
+        __align__(16) nv_bfloat162 tmp[cpy_ne];
+        ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_bf16 + k_KQ_0 + (threadIdx.x % nthreads)*cpy_ne);
+#pragma unroll
+        for (int k_KQ_1 = 0; k_KQ_1 < cpy_ne; ++k_KQ_1) {
+#ifdef V_DOT2_F32_F16_AVAILABLE
+            // FIXME replace macros in vector FA kernel with templating and use FP32 for BF16
+            ggml_cuda_mad(sum, ggml_cuda_cast<float2>(tmp[k_KQ_1]), __half22float2(((const half2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]));
+#else
+            ggml_cuda_mad(sum, ggml_cuda_cast<float2>(tmp[k_KQ_1]), ((const float2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
+#endif // V_DOT2_F32_F16_AVAILABLE
+        }
+    }
+
+    return sum;
+}
+
 template<int D, int nthreads>
 static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q4_0(
     const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
@@ -321,6 +352,19 @@ static __device__ __forceinline__ void dequantize_V_f16(const void * __restrict_
     }
 }
 
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_bf16(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
+    static_assert(std::is_same_v<T, float>, "BF16 V dequantization only supports float output");
+    static_assert(ne % 2 == 0, "bad ne");
+    __align__(16) nv_bfloat162 tmp[ne/2];
+    ggml_cuda_memcpy_1<ne*sizeof(nv_bfloat16)>(tmp, (const nv_bfloat16 *) vx + i0);
+    float2 * dst_f2 = (float2 *) dst;
+#pragma unroll
+    for (int l = 0; l < ne/2; ++l) {
+        dst_f2[l] = ggml_cuda_cast<float2>(tmp[l]);
+    }
+}
+
 template <typename T, int ne>
 static __device__ __forceinline__ void dequantize_V_q4_0(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
     const block_q4_0 * x = (const block_q4_0 *) vx;
@@ -547,6 +591,8 @@ constexpr __device__ vec_dot_KQ_t get_vec_dot_KQ() {
         return vec_dot_fattn_vec_KQ_q5_1<D, nthreads>;
     } else if constexpr (type_K == GGML_TYPE_Q8_0) {
         return vec_dot_fattn_vec_KQ_q8_0<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_BF16) {
+        return vec_dot_fattn_vec_KQ_bf16<D, nthreads>;
     } else {
         static_assert(type_K == -1, "bad type");
         return nullptr;
@@ -567,6 +613,8 @@ constexpr __device__ dequantize_V_t get_dequantize_V() {
         return dequantize_V_q5_1<T, ne>;
     } else if constexpr (type_V == GGML_TYPE_Q8_0) {
         return dequantize_V_q8_0<T, ne>;
+    } else if constexpr (type_V == GGML_TYPE_BF16) {
+        return dequantize_V_bf16<float, ne>;
     } else {
         static_assert(type_V == -1, "bad type");
         return nullptr;

ggml/src/ggml-cuda/fattn-vec.cuh

@@ -75,17 +75,17 @@ static __global__ void flash_attn_ext_vec(
 #endif // GGML_USE_HIP
 
     constexpr int nthreads    = ggml_cuda_fattn_vec_get_nthreads_device();
-    constexpr int nthreads_KQ = type_K == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_KQ_q;
-    constexpr int nthreads_V  = type_V == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_V_q;
+    constexpr int nthreads_KQ = (type_K == GGML_TYPE_F16 || type_K == GGML_TYPE_BF16) ? 128 / cpy_nb : nthreads_KQ_q;
+    constexpr int nthreads_V  = (type_V == GGML_TYPE_F16 || type_V == GGML_TYPE_BF16) ? 128 / cpy_nb : nthreads_V_q;
 
     static_assert(WARP_SIZE % nthreads_KQ == 0, "bad nthreads_K");
     static_assert(WARP_SIZE % nthreads_V  == 0, "bad nthreads_V");
 
-    constexpr int V_rows_per_thread = type_V == GGML_TYPE_F16 ? 2*cpy_ne : 4;
+    constexpr int V_rows_per_thread = (type_V == GGML_TYPE_F16 || type_V == GGML_TYPE_BF16) ? 2*cpy_ne : 4;
     constexpr int V_cols_per_iter   = WARP_SIZE / nthreads_V;
 
     constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
-    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16 && type_K != GGML_TYPE_BF16;
 #ifdef V_DOT2_F32_F16_AVAILABLE
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half,  V_rows_per_thread>();
 #else
@@ -323,8 +323,18 @@ static __global__ void flash_attn_ext_vec(
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
                 half2 tmp[V_rows_per_thread/2];
-                dequantize_V(V + k*nb21, tmp,
-                    2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
+                if constexpr (type_V == GGML_TYPE_BF16) {
+                    float2 tmp_f[V_rows_per_thread/2];
+                    dequantize_V(V + k*nb21, tmp_f,
+                        2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
+#pragma unroll
+                    for (int i_VKQ_1 = 0; i_VKQ_1 < V_rows_per_thread/2; ++i_VKQ_1) {
+                        tmp[i_VKQ_1] = __float22half2_rn(tmp_f[i_VKQ_1]);
+                    }
+                } else {
+                    dequantize_V(V + k*nb21, tmp,
+                        2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
+                }
 #pragma unroll
                 for (int i_VKQ_1 = 0; i_VKQ_1 < V_rows_per_thread/2; ++i_VKQ_1) {
 #pragma unroll
@@ -563,6 +573,7 @@ void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_ten
     extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_0); \
     extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_1); \
     extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q8_0); \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_BF16); \
 
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_F16)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_0)
@@ -570,6 +581,7 @@ EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_1)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q8_0)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_BF16)
 
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_F16)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_0)
@@ -577,6 +589,7 @@ EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_1)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q8_0)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_BF16)
 
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_F16)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_0)
@@ -584,3 +597,4 @@ EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_1)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q8_0)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_BF16)

ggml/src/ggml-cuda/fattn.cu

@@ -224,6 +224,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_F16)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_F16)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_F16)
 
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q4_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
@@ -231,6 +232,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q4_0)
 
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q4_1)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
@@ -238,6 +240,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q4_1)
 
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q5_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
@@ -245,6 +248,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q5_0)
 
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q5_1)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
@@ -252,6 +256,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q5_1)
 
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q8_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
@@ -259,10 +264,20 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_BF16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_BF16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_BF16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_BF16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_BF16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_BF16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_BF16)
 #else
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_F16)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
     FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_BF16)
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
     GGML_ABORT("fatal error");
@@ -355,6 +370,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 #endif // GGML_CUDA_FA_ALL_QUANTS
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_BF16:
             break;
         default:
             return BEST_FATTN_KERNEL_NONE;

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1297,7 +1297,12 @@ static void ggml_cuda_op_mul_mat_cublas(
     const bool supports_bf16 = GGML_CUDA_CC_IS_NVIDIA(cc) || GGML_CUDA_CC_IS_AMD(cc) ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 
-    const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;
+    const bool use_fp16 =
+        src0->type != GGML_TYPE_NVFP4 &&
+        (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
+        ggml_is_contiguous(src0) &&
+        row_diff == src0->ne[1] &&
+        dst->op_params[0] == GGML_PREC_DEFAULT;
 
     if (supports_bf16 && src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
         ggml_cuda_pool_alloc<nv_bfloat16> src1_as_bf16(ctx.pool(id));
@@ -4781,6 +4786,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
                     case GGML_TYPE_MXFP4:
+#ifdef FP8_AVAILABLE
+                    case GGML_TYPE_NVFP4:
+#endif // FP8_AVAILABLE
                     case GGML_TYPE_Q2_K:
                     case GGML_TYPE_Q3_K:
                     case GGML_TYPE_Q4_K:

ggml/src/ggml-cuda/mmvq.cu

@@ -15,6 +15,7 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
         case GGML_TYPE_Q5_1:    return vec_dot_q5_1_q8_1;
         case GGML_TYPE_Q8_0:    return vec_dot_q8_0_q8_1;
         case GGML_TYPE_MXFP4:   return vec_dot_mxfp4_q8_1;
+        case GGML_TYPE_NVFP4:   return vec_dot_nvfp4_q8_1;
         case GGML_TYPE_Q2_K:    return vec_dot_q2_K_q8_1;
         case GGML_TYPE_Q3_K:    return vec_dot_q3_K_q8_1;
         case GGML_TYPE_Q4_K:    return vec_dot_q4_K_q8_1;
@@ -33,7 +34,7 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
     }
 }
 
-static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
+static constexpr __host__ __device__ int get_vdr_mmvq(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
@@ -41,6 +42,7 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
         case GGML_TYPE_Q5_1:    return VDR_Q5_1_Q8_1_MMVQ;
         case GGML_TYPE_Q8_0:    return VDR_Q8_0_Q8_1_MMVQ;
         case GGML_TYPE_MXFP4:   return VDR_MXFP4_Q8_1_MMVQ;
+        case GGML_TYPE_NVFP4:   return VDR_NVFP4_Q8_1_MMVQ;
         case GGML_TYPE_Q2_K:    return VDR_Q2_K_Q8_1_MMVQ;
         case GGML_TYPE_Q3_K:    return VDR_Q3_K_Q8_1_MMVQ;
         case GGML_TYPE_Q4_K:    return VDR_Q4_K_Q8_1_MMVQ;
@@ -173,11 +175,11 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
     return 1;
 }
 
-static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id) {
+static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
     if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
         switch (ncols_dst) {
             case 1:
-                return 1;
+                return small_k ? nwarps : 1;
             case 2:
             case 3:
             case 4:
@@ -193,7 +195,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false, bool small_k = false>
 __launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
         const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
@@ -208,7 +210,7 @@ static __global__ void mul_mat_vec_q(
     constexpr int vdr = get_vdr_mmvq(type);
     constexpr mmvq_parameter_table_id table_id = get_device_table_id();
     constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
-    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
+    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
@@ -414,14 +416,16 @@ static __global__ void mul_mat_vec_q(
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
         const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
-        const int warp_size, const mmvq_parameter_table_id table_id) {
-    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
+        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false) {
+    const int nwarps = calc_nwarps(type, ncols_dst, table_id);
+    const int rpb = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
+    const int64_t nblocks = (nrows_x + rpb - 1) / rpb;
     const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
-    const dim3 block_dims(warp_size, calc_nwarps(type, ncols_dst, table_id), 1);
+    const dim3 block_dims(warp_size, nwarps, 1);
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false>
+template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false, bool small_k = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -434,7 +438,7 @@ static void mul_mat_vec_q_switch_fusion(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
                 (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -444,7 +448,7 @@ static void mul_mat_vec_q_switch_fusion(
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
         (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -488,11 +492,33 @@ static void mul_mat_vec_q_switch_ncols_dst(
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
-            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, ids_stride, stream);
+
+            // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
+            // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
+            constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+            constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+            constexpr int vdr = get_vdr_mmvq(type);
+            const int blocks_per_row_x = ncols_x / qk;
+            const int blocks_per_iter_1warp = vdr * warp_size / qi;
+            const int nwarps = calc_nwarps(type, c_ncols_dst, table_id);
+            const bool use_small_k = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+            if (use_small_k) {
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
+                                                                    warp_size, table_id, true);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, false, true>(
+                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                    dims.first, dims.second, 0, ids_stride, stream);
+            } else {
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
+                                                                    warp_size, table_id);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
+                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                    dims.first, dims.second, 0, ids_stride, stream);
+            }
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
@@ -602,6 +628,12 @@ static void mul_mat_vec_q_switch_type(
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                  nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
+        case GGML_TYPE_NVFP4:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_NVFP4>
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
+            break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_F16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q4_0);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q4_1);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q5_0);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q5_1);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q8_0);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_BF16);

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -5,7 +5,7 @@ import os
 
 HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 576]
 
-TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0"]
+TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", "GGML_TYPE_BF16"]
 
 SOURCE_FATTN_TILE = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
 

ggml/src/ggml-cuda/vecdotq.cuh

@@ -322,6 +322,38 @@ static __device__ __forceinline__ float vec_dot_mxfp4_q8_1(
     return d * sumi;
 }
 
+#define VDR_NVFP4_Q8_1_MMVQ 4
+#define VDR_NVFP4_Q8_1_MMQ  8
+
+static __device__ __forceinline__ float vec_dot_nvfp4_q8_1(
+                                        const void * __restrict__ vbq,
+                                        const block_q8_1 * __restrict__ bq8_1,
+                                        const int32_t & kbx,
+                                        const int32_t & iqs) {
+
+    const block_nvfp4 * bq4 = (const block_nvfp4 *) vbq + kbx;
+    float sum = 0.0f;
+#pragma unroll
+    for (int i = 0; i < VDR_NVFP4_Q8_1_MMVQ/2; i++) {
+        const int32_t iqs0 = iqs + 2*i;
+        const int32_t iqs1 = iqs0 + 1;
+        const int32_t is = iqs0 >> 1;
+        const int2 v0 = get_int_from_table_16(get_int_b4(bq4->qs, iqs0), kvalues_mxfp4);
+        const int2 v1 = get_int_from_table_16(get_int_b4(bq4->qs, iqs1), kvalues_mxfp4);
+        const block_q8_1 * bq8 = bq8_1 + (is >> 1);
+        const int32_t i8 = ((is & 1) << 2);
+
+        int sumi = ggml_cuda_dp4a(v0.x, get_int_b4(bq8->qs, i8 + 0), 0);
+        sumi = ggml_cuda_dp4a(v0.y, get_int_b4(bq8->qs, i8 + 2), sumi);
+        sumi = ggml_cuda_dp4a(v1.x, get_int_b4(bq8->qs, i8 + 1), sumi);
+        sumi = ggml_cuda_dp4a(v1.y, get_int_b4(bq8->qs, i8 + 3), sumi);
+
+        const float d = ggml_cuda_ue4m3_to_fp32(bq4->d[is]) * __low2float(bq8->ds);
+        sum += d * float(sumi);
+    }
+
+    return sum;
+}
 #define VDR_Q2_K_Q8_1_MMVQ 1
 #define VDR_Q2_K_Q8_1_MMQ  4
 

ggml/src/ggml-cuda/vendors/cuda.h

@@ -6,9 +6,10 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 
-#if CUDART_VERSION >= 12050
+#if CUDART_VERSION >= 11080
 #include <cuda_fp8.h>
-#endif // CUDART_VERSION >= 12050
+#define FP8_AVAILABLE
+#endif // CUDART_VERSION >= 11080
 
 #if CUDART_VERSION >= 12080
 #include <cuda_fp4.h>

ggml/src/ggml-cuda/vendors/hip.h

@@ -235,6 +235,12 @@
 typedef __hip_bfloat16 nv_bfloat16;
 typedef __hip_bfloat162 nv_bfloat162;
 
+#if HIP_VERSION >= 60200000
+#include <hip/hip_fp8.h>
+typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
+#define FP8_AVAILABLE
+#endif // HIP_VERSION >= 60200000
+
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
 static __device__ __forceinline__ int __vsubss4(const int a, const int b) {

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -461,7 +461,7 @@ static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
         d[7]          = x[i * 8 + 7].d;
     }
 
-    if (opt_verbose > 1) {
+    if (opt_verbose > 2) {
         for (int i = 0; i < nb; i++) {
             dump_packed_block_q4x4x2(y, i, k);
         }
@@ -480,7 +480,7 @@ static void unpack_row_q4x4x2(block_q4_0 * x, const uint8_t * y, int64_t k) {
     const uint8_t * y_q = y + 0;              // quants first
     const uint8_t * y_d = y + qrow_size;      // then scales
 
-    if (opt_verbose > 1) {
+    if (opt_verbose > 2) {
         for (int i = 0; i < nb; i++) {
             dump_packed_block_q4x4x2(y, i, k);
         }
@@ -796,7 +796,7 @@ static void repack_row_q8x4x2(uint8_t * y, const block_q8_0 * x, int64_t k) {
         d[7]          = x[i * 8 + 7].d;
     }
 
-    if (opt_verbose > 1) {
+    if (opt_verbose > 2) {
         for (int i = 0; i < nb; i++) {
             dump_packed_block_q8x4x2(y, i, k);
         }
@@ -814,7 +814,7 @@ static void unpack_row_q8x4x2(block_q8_0 * x, const uint8_t * y, int64_t k) {
     const uint8_t * y_q = y + 0;              // quants first
     const uint8_t * y_d = y + qrow_size;      // then scales
 
-    if (opt_verbose > 1) {
+    if (opt_verbose > 2) {
         for (int i = 0; i < nb; i++) {
             dump_packed_block_q8x4x2(y, i, k);
         }
@@ -1149,7 +1149,7 @@ static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k)
         e[7]        = x[i * 8 + 7].e;
     }
 
-    if (opt_verbose > 1) {
+    if (opt_verbose > 2) {
         for (int i = 0; i < nb; i++) {
             dump_packed_block_mxfp4x4x2(y, i, k);
         }
@@ -1168,7 +1168,7 @@ static void unpack_row_mxfp4x4x2(block_mxfp4 * x, const uint8_t * y, int64_t k)
     const uint8_t * y_q = y + 0;              // quants first
     const uint8_t * y_e = y + qrow_size;      // then scales
 
-    if (opt_verbose > 1) {
+    if (opt_verbose > 2) {
         for (int i = 0; i < nb; i++) {
             dump_packed_block_mxfp4x4x2(y, i, k);
         }

ggml/src/ggml-hexagon/htp/binary-ops.c

@@ -24,28 +24,26 @@
 // Context for binary operations
 struct htp_binary_context {
     struct htp_ops_context * octx;
-    struct fastdiv_values dim1_div;
-    struct fastdiv_values dim2_div;
-    struct fastdiv_values dim12_div;
+
+    struct fastdiv_values src0_dim1_div; // ne01
+    struct fastdiv_values src0_dim2_div; // ne02
+    struct fastdiv_values src0_dim12_div;// ne03
 
     struct fastdiv_values src1_dim1_div; // ne11
     struct fastdiv_values src1_dim2_div; // ne12
     struct fastdiv_values src1_dim3_div; // ne13
 
-    uint32_t nrows_per_thread;
-    bool split_at_ne01;
-    bool split_at_ne02;
-
-    // Precomputed values
     uint32_t block_max;
+    uint32_t nrows_per_thread;
     size_t   src0_row_size_aligned;
     size_t   src1_row_size_aligned;
     size_t   dst_row_size_aligned;
-    uint32_t src1_fetch_rows; // 1 or block_max
-    uint32_t src1_dma_stride; // 0 or stride
+
+    bool split_at_ne01;
+    bool split_at_ne02;
 };
 
-#define htp_binary_preamble            \
+#define htp_binary_preamble                       \
     const struct htp_tensor * src0 = &octx->src0; \
     const struct htp_tensor * src1 = &octx->src1; \
     struct htp_tensor *       dst  = &octx->dst;  \
@@ -72,12 +70,11 @@ struct htp_binary_context {
     const uint32_t nb2 = dst->nb[2];   \
     const uint32_t nb3 = dst->nb[3];
 
-static inline uint32_t calc_block_size(struct htp_binary_context * bctx, uint32_t ir, uint32_t end_row,
-                                uint32_t ne01, uint32_t ne02) {
+static inline uint32_t calc_block_size(struct htp_binary_context * bctx, uint32_t ir, uint32_t end_row, uint32_t ne01, uint32_t ne02) {
     uint32_t i03, i02, i01, rem;
-    i03 = fastdiv(ir, &bctx->dim12_div);
+    i03 = fastdiv(ir, &bctx->src0_dim12_div);
     rem = ir - i03 * (ne02 * ne01);
-    i02 = fastdiv(rem, &bctx->dim1_div);
+    i02 = fastdiv(rem, &bctx->src0_dim1_div);
     i01 = rem - i02 * ne01;
 
     uint32_t rows_left = end_row - ir;
@@ -191,6 +188,8 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
     if (start_row >= end_row) return;
 
+    FARF(HIGH, "binary-scalar: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
+
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
     uint8_t * dst_spad_base  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
     size_t src0_spad_half    = octx->src0_spad.size_per_thread / 2;
@@ -204,9 +203,9 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
     for (int k = 0; k < 2 && ir_prefetch < end_row; k++) {
         uint32_t current_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
         uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir_prefetch, &bctx->dim12_div);
+        i03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
         rem = ir_prefetch - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
+        i02 = fastdiv(rem, &bctx->src0_dim1_div);
         i01 = rem - i02 * ne01;
 
         uint8_t * src0_curr = (uint8_t *)src0->data + i03 * nb03 + i02 * nb02 + i01 * nb01;
@@ -215,7 +214,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * s0_spad = src0_spad_base + spad_idx * src0_spad_half;
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
-        dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, 0);
         dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
@@ -229,9 +228,9 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
 
         uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir, &bctx->dim12_div);
+        i03 = fastdiv(ir, &bctx->src0_dim12_div);
         rem = ir - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
+        i02 = fastdiv(rem, &bctx->src0_dim1_div);
         i01 = rem - i02 * ne01;
 
         // src1 indices (broadcast/repeat)
@@ -255,9 +254,9 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
              uint32_t p03, p02, p01, prem;
-             p03 = fastdiv(ir_prefetch, &bctx->dim12_div);
+             p03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
              prem = ir_prefetch - p03 * (ne02 * ne01);
-             p02 = fastdiv(prem, &bctx->dim1_div);
+             p02 = fastdiv(prem, &bctx->src0_dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
 
@@ -282,6 +281,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
     if (start_row >= end_row) return;
 
+    FARF(HIGH, "binary-same-shape: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
+
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
     uint8_t * src1_spad_base = octx->src1_spad.data + (ith * octx->src1_spad.size_per_thread);
     uint8_t * dst_spad_base  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
@@ -297,9 +298,9 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
     for (int k = 0; k < 2 && ir_prefetch < end_row; k++) {
         uint32_t current_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
         uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir_prefetch, &bctx->dim12_div);
+        i03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
         rem = ir_prefetch - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
+        i02 = fastdiv(rem, &bctx->src0_dim1_div);
         i01 = rem - i02 * ne01;
 
         uint32_t i13 = (ne13 == 1) ? 0 : i03;
@@ -307,23 +308,23 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         uint32_t i11 = (ne11 == 1) ? 0 : i01;
 
         uint8_t * src0_curr = (uint8_t *)src0->data + i03 * nb03 + i02 * nb02 + i01 * nb01;
-        uint8_t * src1_base = (uint8_t *)src1->data + i13 * nb13 + i12 * nb12 + i11 * nb11;
+        uint8_t * src1_curr = (uint8_t *)src1->data + i13 * nb13 + i12 * nb12 + i11 * nb11;
         uint8_t * dst_curr  = (uint8_t *)dst->data  + i03 * nb3  + i02 * nb2  + i01 * nb1;
 
         uint8_t * s0_spad = src0_spad_base + spad_idx * src0_spad_half;
         uint8_t * s1_spad = src1_spad_base + spad_idx * src1_spad_half;
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
-        dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, 0);
         dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
-        dma_queue_push(q, dma_make_ptr(s1_spad, src1_base), bctx->src1_row_size_aligned, bctx->src1_dma_stride, row_size_bytes, current_block_size);
+        dma_queue_push(q, dma_make_ptr(s1_spad, src1_curr), bctx->src1_row_size_aligned, nb11, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
 
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
-        uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
+        uint8_t * d_spad  = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
         uint8_t * s1_spad = (uint8_t *) dma_queue_pop(q).dst;
 
@@ -335,9 +336,9 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         }
 
         uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir, &bctx->dim12_div);
+        i03 = fastdiv(ir, &bctx->src0_dim12_div);
         rem = ir - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
+        i02 = fastdiv(rem, &bctx->src0_dim1_div);
         i01 = rem - i02 * ne01;
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
         dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
@@ -345,9 +346,9 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
              uint32_t p03, p02, p01, prem;
-             p03 = fastdiv(ir_prefetch, &bctx->dim12_div);
+             p03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
              prem = ir_prefetch - p03 * (ne02 * ne01);
-             p02 = fastdiv(prem, &bctx->dim1_div);
+             p02 = fastdiv(prem, &bctx->src0_dim1_div);
              p01 = prem - p02 * ne01;
 
              uint32_t p13 = (ne13 == 1) ? 0 : p03;
@@ -358,7 +359,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
              uint8_t * s1_next = (uint8_t *)src1->data + p13 * nb13 + p12 * nb12 + p11 * nb11;
 
              dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
-             dma_queue_push(q, dma_make_ptr(s1_spad, s1_next), bctx->src1_row_size_aligned, bctx->src1_dma_stride, row_size_bytes, next_block_size);
+             dma_queue_push(q, dma_make_ptr(s1_spad, s1_next), bctx->src1_row_size_aligned, nb11, row_size_bytes, next_block_size);
 
              ir_prefetch += next_block_size;
         }
@@ -373,15 +374,17 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type  = octx->src0.type;
     const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row = bctx->nrows_per_thread * ith;
-    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row  = bctx->nrows_per_thread * ith;
+    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, total_rows);
     if (start_row >= end_row) return;
 
+    FARF(HIGH, "binary-row-bcast: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
+
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
-    uint8_t * src1_spad = octx->src1_spad.data + (ith * octx->src1_spad.size_per_thread);
+    uint8_t * src1_spad_base = octx->src1_spad.data + (ith * octx->src1_spad.size_per_thread);
     uint8_t * dst_spad_base  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
 
     size_t src0_spad_half = octx->src0_spad.size_per_thread / 2;
@@ -391,15 +394,14 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
     uint32_t ir_prefetch = start_row;
     int spad_idx = 0;
 
-    void * s1_ptr = (void *) src1_spad;
+    void * s1_ptr = (void *) src1_spad_base;
 
     for (int k = 0; k < 2 && ir_prefetch < end_row; k++) {
         uint32_t current_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-        rem = ir_prefetch - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+        uint32_t rem = ir_prefetch - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         uint8_t * src0_curr = (uint8_t *)src0->data + i03 * nb03 + i02 * nb02 + i01 * nb01;
         uint8_t * dst_curr  = (uint8_t *)dst->data  + i03 * nb3  + i02 * nb2  + i01 * nb1;
@@ -407,7 +409,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         uint8_t * s0_spad = src0_spad_base + spad_idx * src0_spad_half;
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
-        dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, 0);
         dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
@@ -415,7 +417,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
 
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
-        uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
+        uint8_t * d_spad  = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
 
         for (uint32_t r = 0; r < current_block_size; r++) {
@@ -425,21 +427,19 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
             COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir, &bctx->dim12_div);
-        rem = ir - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir, &bctx->src0_dim12_div);
+        uint32_t rem = ir - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
         dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-             uint32_t p03, p02, p01, prem;
-             p03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-             prem = ir_prefetch - p03 * (ne02 * ne01);
-             p02 = fastdiv(prem, &bctx->dim1_div);
-             p01 = prem - p02 * ne01;
+             uint32_t p03  = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+             uint32_t prem = ir_prefetch - p03 * (ne02 * ne01);
+             uint32_t p02  = fastdiv(prem, &bctx->src0_dim1_div);
+             uint32_t p01  = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
              dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
@@ -458,14 +458,16 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
     const uint32_t src0_type = octx->src0.type;
     const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row = bctx->nrows_per_thread * ith;
-    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row  = bctx->nrows_per_thread * ith;
+    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, total_rows);
     if (start_row >= end_row) return;
 
+    FARF(HIGH, "binary-complex: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
+
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
     uint8_t * dst_spad_base  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
-    size_t src0_spad_half = octx->src0_spad.size_per_thread / 2;
-    size_t dst_spad_half  = octx->dst_spad.size_per_thread  / 2;
+    size_t src0_spad_half    = octx->src0_spad.size_per_thread / 2;
+    size_t dst_spad_half     = octx->dst_spad.size_per_thread  / 2;
 
     dma_queue * q = octx->ctx->dma[ith];
     uint32_t ir_prefetch = start_row;
@@ -473,11 +475,10 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
 
     for (int k = 0; k < 2 && ir_prefetch < end_row; k++) {
         uint32_t current_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-        rem = ir_prefetch - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+        uint32_t rem = ir_prefetch - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         uint8_t * src0_curr = (uint8_t *)src0->data + i03 * nb03 + i02 * nb02 + i01 * nb01;
         uint8_t * dst_curr  = (uint8_t *)dst->data  + i03 * nb3  + i02 * nb2  + i01 * nb1;
@@ -485,7 +486,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         uint8_t * s0_spad = src0_spad_base + spad_idx * src0_spad_half;
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
-        dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, 0);
         dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
@@ -496,11 +497,10 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
 
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir, &bctx->dim12_div);
-        rem = ir - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir, &bctx->src0_dim12_div);
+        uint32_t rem = ir - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint32_t r_i01 = i01 + r;
@@ -521,11 +521,10 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-             uint32_t p03, p02, p01, prem;
-             p03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-             prem = ir_prefetch - p03 * (ne02 * ne01);
-             p02 = fastdiv(prem, &bctx->dim1_div);
-             p01 = prem - p02 * ne01;
+             uint32_t p03  = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+             uint32_t prem = ir_prefetch - p03 * (ne02 * ne01);
+             uint32_t p02  = fastdiv(prem, &bctx->src0_dim1_div);
+             uint32_t p01  = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
              dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
@@ -545,14 +544,16 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
     const uint32_t elem_size_bytes = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
     const uint32_t row_size_bytes = ne00 * elem_size_bytes;;
     const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row = bctx->nrows_per_thread * ith;
-    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row  = bctx->nrows_per_thread * ith;
+    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, total_rows);
     if (start_row >= end_row) return;
 
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
     uint8_t * dst_spad_base  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
-    size_t src0_spad_half = octx->src0_spad.size_per_thread / 2;
-    size_t dst_spad_half  = octx->dst_spad.size_per_thread  / 2;
+    size_t src0_spad_half    = octx->src0_spad.size_per_thread / 2;
+    size_t dst_spad_half     = octx->dst_spad.size_per_thread  / 2;
+
+    FARF(HIGH, "binary-repeat: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
 
     dma_queue * q = octx->ctx->dma[ith];
     uint32_t ir_prefetch = start_row;
@@ -560,11 +561,10 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
 
     for (int k = 0; k < 2 && ir_prefetch < end_row; k++) {
         uint32_t current_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-        rem = ir_prefetch - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+        uint32_t rem = ir_prefetch - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         uint8_t * src0_curr = (uint8_t *)src0->data + i03 * nb03 + i02 * nb02 + i01 * nb01;
         uint8_t * dst_curr  = (uint8_t *)dst->data  + i03 * nb3  + i02 * nb2  + i01 * nb1;
@@ -572,7 +572,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         uint8_t * s0_spad = src0_spad_base + spad_idx * src0_spad_half;
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
-        dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, 0);
         dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
@@ -583,11 +583,10 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
 
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir, &bctx->dim12_div);
-        rem = ir - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir, &bctx->src0_dim12_div);
+        uint32_t rem = ir - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint32_t r_i01 = i01 + r;
@@ -612,11 +611,10 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-             uint32_t p03, p02, p01, prem;
-             p03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-             prem = ir_prefetch - p03 * (ne02 * ne01);
-             p02 = fastdiv(prem, &bctx->dim1_div);
-             p01 = prem - p02 * ne01;
+             uint32_t p03  = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+             uint32_t prem = ir_prefetch - p03 * (ne02 * ne01);
+             uint32_t p02  = fastdiv(prem, &bctx->src0_dim1_div);
+             uint32_t p01  = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
              dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
@@ -646,6 +644,7 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t nb02 = src0->nb[2];
     const uint32_t nb03 = src0->nb[3];
     const uint32_t nb11 = src1->nb[1]; // src1 row stride
+
     const uint32_t nb1 = dst->nb[1];
     const uint32_t nb2 = dst->nb[2];
     const uint32_t nb3 = dst->nb[3];
@@ -657,8 +656,8 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
 
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
     uint8_t * dst_spad_base  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
-    size_t src0_spad_half = octx->src0_spad.size_per_thread / 2;
-    size_t dst_spad_half  = octx->dst_spad.size_per_thread  / 2;
+    size_t src0_spad_half    = octx->src0_spad.size_per_thread / 2;
+    size_t dst_spad_half     = octx->dst_spad.size_per_thread  / 2;
 
     dma_queue * q = octx->ctx->dma[ith];
     uint32_t ir_prefetch = start_row;
@@ -666,11 +665,10 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
 
     for (int k = 0; k < 2 && ir_prefetch < end_row; k++) {
         uint32_t current_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-        rem = ir_prefetch - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+        uint32_t rem = ir_prefetch - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         uint8_t * src0_curr = (uint8_t *)src0->data + i03 * nb03 + i02 * nb02 + i01 * nb01;
         uint8_t * dst_curr  = (uint8_t *)dst->data  + i03 * nb3  + i02 * nb2  + i01 * nb1;
@@ -678,7 +676,7 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * s0_spad = src0_spad_base + spad_idx * src0_spad_half;
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
-        dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), 0);
         dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
@@ -689,11 +687,10 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
 
-        uint32_t i03, i02, i01, rem;
-        i03 = fastdiv(ir, &bctx->dim12_div);
-        rem = ir - i03 * (ne02 * ne01);
-        i02 = fastdiv(rem, &bctx->dim1_div);
-        i01 = rem - i02 * ne01;
+        uint32_t i03 = fastdiv(ir, &bctx->src0_dim12_div);
+        uint32_t rem = ir - i03 * (ne02 * ne01);
+        uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
+        uint32_t i01 = rem - i02 * ne01;
 
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint32_t r_i01 = i01 + r; // linear within block since we split at ne01
@@ -712,11 +709,10 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
-             uint32_t p03, p02, p01, prem;
-             p03 = fastdiv(ir_prefetch, &bctx->dim12_div);
-             prem = ir_prefetch - p03 * (ne02 * ne01);
-             p02 = fastdiv(prem, &bctx->dim1_div);
-             p01 = prem - p02 * ne01;
+             uint32_t p03  = fastdiv(ir_prefetch, &bctx->src0_dim12_div);
+             uint32_t prem = ir_prefetch - p03 * (ne02 * ne01);
+             uint32_t p02  = fastdiv(prem, &bctx->src0_dim1_div);
+             uint32_t p01  = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
              dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
              ir_prefetch += next_block_size;
@@ -739,40 +735,36 @@ static int execute_op_binary(struct htp_ops_context * octx) {
     const size_t elem_size = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
     const size_t src0_row_size = src0->ne[0] * elem_size;
     const size_t src1_row_size = src1->ne[0] * elem_size;
-    const size_t dst_row_size  = dst->ne[0] * elem_size;
+    const size_t dst_row_size  = dst->ne[0]  * elem_size;
 
-    // Align to VLEN
-    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
-    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+    size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
     size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
+    size_t dst_row_size_aligned  = hex_round_up(dst_row_size,  VLEN);
 
     bool is_add_id = (octx->op == HTP_OP_ADD_ID);
     bool is_scalar = !is_add_id && (src1->ne[0] == 1);
 
-    // Determine which kernel we will use to alloc memory and dispatch
-    bool use_vector_same = !is_add_id && !is_scalar && ((src0->nb[1] % VLEN) == 0) && (src1->ne[0] == src0->ne[0]) &&
+    bool is_transposed = (src0->nb[1] < src0_row_size || src1->nb[1] < src1_row_size || dst->nb[1] < dst_row_size);
+
+    bool is_same_shape = !is_add_id && !is_scalar && !is_transposed &&
+               (src1->ne[0] == src0->ne[0] && src0->ne[0] % VLEN == 0) &&
                (src1->ne[1] == src0->ne[1] || src1->ne[1] == 1) &&
                (src1->ne[2] == src0->ne[2] || src1->ne[2] == 1) &&
                (src1->ne[3] == src0->ne[3] || src1->ne[3] == 1);
 
-    bool is_row_bcast = use_vector_same && (src1->ne[1] == 1 && src1->ne[2] == 1 && src1->ne[3] == 1);
-    bool use_complex = !is_add_id && !is_scalar && !use_vector_same && (src1->ne[0] == src0->ne[0]);
-    bool use_repeat  = !is_add_id && !is_scalar && !use_vector_same && (src1->ne[0] != src0->ne[0]);
+    bool is_row_bcast = is_same_shape && (src1->ne[1] == 1 && src1->ne[2] == 1 && src1->ne[3] == 1);
+    bool is_complex   = !is_add_id && !is_scalar && !is_same_shape && (src1->ne[0] == src0->ne[0]);
+    bool is_repeat    = !is_add_id && !is_scalar && !is_same_shape && (src1->ne[0] != src0->ne[0]);
 
     size_t spad_row_total;
-    if (is_scalar) {
-        spad_row_total = 2 * (src0_row_size_aligned + dst_row_size_aligned);
-    } else if (is_row_bcast) {
-        spad_row_total = 2 * (src0_row_size_aligned + dst_row_size_aligned);
-    } else if (use_vector_same) {
+    if (is_same_shape) {
         spad_row_total = 2 * (src0_row_size_aligned + src1_row_size_aligned + dst_row_size_aligned);
-    } else if (is_add_id) {
-        spad_row_total = 2 * (src0_row_size_aligned + dst_row_size_aligned); // src1 read directly
     } else {
         spad_row_total = 2 * (src0_row_size_aligned + dst_row_size_aligned);
     }
 
     size_t rows_per_buffer = octx->ctx->vtcm_size / (n_threads * spad_row_total);
+
     // Adjust for static src1 in row_bcast case
     if (is_row_bcast) {
         size_t needed_static = src1_row_size_aligned;
@@ -782,28 +774,26 @@ static int execute_op_binary(struct htp_ops_context * octx) {
     }
 
     if (rows_per_buffer < 1) {
-         FARF(ERROR, "binary: VTCM too small\n");
-         return HTP_STATUS_VTCM_TOO_SMALL;
+        FARF(ERROR, "binary: VTCM too small\n");
+        return HTP_STATUS_VTCM_TOO_SMALL;
     }
 
     octx->src0_spad.size_per_thread = rows_per_buffer * 2 * src0_row_size_aligned;
     octx->dst_spad.size_per_thread  = rows_per_buffer * 2 * dst_row_size_aligned;
 
-    if (is_scalar || use_complex || use_repeat || is_add_id) {
-        octx->src1_spad.size_per_thread = 0;
-    } else if (is_row_bcast) {
+    if (is_add_id || is_scalar || is_complex || is_repeat || is_row_bcast) {
         octx->src1_spad.size_per_thread = 0;
     } else {
         octx->src1_spad.size_per_thread = rows_per_buffer * 2 * src1_row_size_aligned;
     }
 
+    octx->dst_spad.size  = n_threads * octx->dst_spad.size_per_thread;
     octx->src0_spad.size = n_threads * octx->src0_spad.size_per_thread;
     if (is_row_bcast) {
         octx->src1_spad.size = src1_row_size_aligned;
     } else {
         octx->src1_spad.size = n_threads * octx->src1_spad.size_per_thread;
     }
-    octx->dst_spad.size  = n_threads * octx->dst_spad.size_per_thread;
 
     if (octx->ctx->vtcm_size < (octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size)) {
         return HTP_STATUS_VTCM_TOO_SMALL;
@@ -823,46 +813,37 @@ static int execute_op_binary(struct htp_ops_context * octx) {
     }
 
     struct htp_binary_context bctx;
-    bctx.octx = octx;
-    bctx.nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
-    bctx.block_max = rows_per_buffer;
+    bctx.octx                  = octx;
+    bctx.nrows_per_thread      = (src0_nrows + n_threads - 1) / n_threads;
+    bctx.block_max             = rows_per_buffer;
     bctx.src0_row_size_aligned = src0_row_size_aligned;
     bctx.src1_row_size_aligned = src1_row_size_aligned;
     bctx.dst_row_size_aligned  = dst_row_size_aligned;
 
-    bctx.dim1_div = init_fastdiv_values(src0->ne[1]);
-    bctx.dim2_div = init_fastdiv_values(src0->ne[2]);
-    bctx.dim12_div = init_fastdiv_values(src0->ne[1] * src0->ne[2]);
+    bctx.src0_dim1_div  = init_fastdiv_values(src0->ne[1]);
+    bctx.src0_dim2_div  = init_fastdiv_values(src0->ne[2]);
+    bctx.src0_dim12_div = init_fastdiv_values(src0->ne[1] * src0->ne[2]);
 
-    bctx.src1_dim1_div = init_fastdiv_values(src1->ne[1]);
-    bctx.src1_dim2_div = init_fastdiv_values(src1->ne[2]);
-    bctx.src1_dim3_div = init_fastdiv_values(src1->ne[3]);
+    bctx.src1_dim1_div  = init_fastdiv_values(src1->ne[1]);
+    bctx.src1_dim2_div  = init_fastdiv_values(src1->ne[2]);
+    bctx.src1_dim3_div  = init_fastdiv_values(src1->ne[3]);
 
     bool src0_contig_dim1 = (src0->nb[2] == src0->ne[1] * src0->nb[1]);
-    bool dst_contig_dim1  = (dst->nb[2] == src0->ne[1] * dst->nb[1]);
+    bool dst_contig_dim1  = (dst->nb[2]  == src0->ne[1] * dst->nb[1]);
 
     bool src0_contig_dim2 = (src0->nb[3] == src0->ne[2] * src0->nb[2]);
-    bool dst_contig_dim2  = (dst->nb[3] == src0->ne[2] * dst->nb[2]);
+    bool dst_contig_dim2  = (dst->nb[3]  == src0->ne[2] * dst->nb[2]);
 
-    bctx.split_at_ne01 = (src0->ne[2] > 1) &&
-                         ((src1->ne[1] > 1) || (src1->ne[2] > 1) || !src0_contig_dim1 || !dst_contig_dim1);
-
-    bctx.split_at_ne02 = (src0->ne[3] > 1) &&
-                         ((src1->ne[2] > 1) || (src1->ne[3] > 1) || !src0_contig_dim2 || !dst_contig_dim2);
-
-    // Precompute specific kernel parameters
-    if (use_vector_same) {
-        bctx.src1_dma_stride = (src1->ne[1] == 1) ? 0 : src1->nb[1];
-        bctx.src1_fetch_rows = (src1->ne[1] == 1) ? 1 : rows_per_buffer;
-    }
+    bctx.split_at_ne01 = (src0->ne[2] > 1) && ((src1->ne[1] > 1) || (src1->ne[2] > 1) || !src0_contig_dim1 || !dst_contig_dim1);
+    bctx.split_at_ne02 = (src0->ne[3] > 1) && ((src1->ne[2] > 1) || (src1->ne[3] > 1) || !src0_contig_dim2 || !dst_contig_dim2);
 
     worker_callback_t worker_func;
-    if (is_add_id) worker_func = binary_job_add_id;
-    else if (is_scalar) worker_func = binary_job_scalar;
-    else if (is_row_bcast) worker_func = binary_job_vector_row_broadcast;
-    else if (use_vector_same) worker_func = binary_job_vector_same_shape;
-    else if (use_complex) worker_func = binary_job_vector_complex;
-    else worker_func = binary_job_element_repeat;
+    if (is_add_id)          worker_func = binary_job_add_id;
+    else if (is_scalar)     worker_func = binary_job_scalar;
+    else if (is_row_bcast)  worker_func = binary_job_vector_row_broadcast;
+    else if (is_same_shape) worker_func = binary_job_vector_same_shape;
+    else if (is_complex)    worker_func = binary_job_vector_complex;
+    else                    worker_func = binary_job_element_repeat;
 
     if (is_row_bcast) {
         dma_queue_pop(q);

ggml/src/ggml-hexagon/htp/hex-dma.c

@@ -31,8 +31,8 @@ dma_queue * dma_queue_create(size_t capacity) {
     q->capacity = capacity;
     q->idx_mask = capacity - 1;
 
-    q->desc = (hexagon_udma_descriptor_type1_t *) memalign(64, capacity * sizeof(hexagon_udma_descriptor_type1_t));
-    memset(q->desc, 0, capacity * sizeof(hexagon_udma_descriptor_type1_t));
+    q->desc = (dma_descriptor_2d *) memalign(64, capacity * sizeof(dma_descriptor_2d));
+    memset(q->desc, 0, capacity * sizeof(dma_descriptor_2d));
 
     q->dptr = (dma_ptr *) memalign(4, capacity * sizeof(dma_ptr));
     memset(q->dptr, 0, capacity * sizeof(dma_ptr));

ggml/src/ggml-hexagon/htp/hex-dma.h

@@ -10,19 +10,84 @@
 extern "C" {
 #endif
 
+// Define the HW descriptor structs here since the ones in HexSDK are a bit out of date
+typedef struct dma_descriptor_1d_s {
+    void *   next;
+    uint32_t size:24;
+    uint32_t desc_size:2;
+    uint32_t dst_comp:1;
+    uint32_t src_comp:1;
+    uint32_t dst_bypass:1;
+    uint32_t src_bypass:1;
+    uint32_t order:1;
+    uint32_t done:1;
+    void *   src;
+    void *   dst;
+} dma_descriptor_1d;
+
+#if __HVX_ARCH__ < 75
+
+typedef struct dma_descriptor_2d_s {
+    void *   next;
+    uint32_t reserved0:24;
+    uint32_t desc_size:2;
+    uint32_t dst_comp:1;
+    uint32_t src_comp:1;
+    uint32_t dst_bypass:1;
+    uint32_t src_bypass:1;
+    uint32_t order:1;
+    uint32_t done:1;
+    void *   src;
+    void *   dst;
+    uint32_t desc_type:8;
+    uint32_t reserved1:24;
+    uint32_t row_size:16;
+    uint32_t nrows:16;
+    uint32_t src_stride:16;
+    uint32_t dst_stride:16;
+    uint32_t src_offset:16;
+    uint32_t dst_offset:16;
+} dma_descriptor_2d;
+
+#else
+
+typedef struct dma_descriptor_2d_s {
+    void *   next;
+    uint32_t dst_stride:24;
+    uint32_t desc_size:2;
+    uint32_t dst_comp:1;
+    uint32_t src_comp:1;
+    uint32_t dst_bypass:1;
+    uint32_t src_bypass:1;
+    uint32_t order:1;
+    uint32_t done:1;
+    void *   src;
+    void *   dst;
+    uint32_t desc_type:8;
+    uint32_t reserved0:24;
+    uint32_t row_size:24;
+    uint32_t nrows_lo:8;
+    uint32_t nrows_hi:8;
+    uint32_t src_stride:24;
+    uint32_t offset:24;
+    uint32_t reserved1:8;
+} dma_descriptor_2d;
+
+#endif
+
 typedef struct {
-    void *dst;
+    void       *dst;
     const void *src;
 } dma_ptr;
 
 typedef struct {
-    hexagon_udma_descriptor_type1_t * desc;  // descriptor pointers
-    hexagon_udma_descriptor_type1_t * tail;  // tail pointer
-    dma_ptr                         * dptr;  // dst/src pointers
-    uint32_t                          push_idx;
-    uint32_t                          pop_idx;
-    uint32_t                          capacity;
-    uint32_t                          idx_mask;
+    dma_descriptor_2d * desc;  // descriptor pointers
+    dma_descriptor_2d * tail;  // tail pointer
+    dma_ptr           * dptr;  // dst/src pointers
+    uint32_t            push_idx;
+    uint32_t            pop_idx;
+    uint32_t            capacity;
+    uint32_t            idx_mask;
 } dma_queue;
 
 dma_queue * dma_queue_create(size_t capacity);
@@ -59,71 +124,87 @@ static inline dma_ptr dma_make_ptr(void *dst, const void *src)
     return p;
 }
 
-static inline bool dma_queue_push(dma_queue * q,
-                                  dma_ptr     dptr,
-                                  size_t      dst_row_size,
-                                  size_t      src_row_size,
-                                  size_t      width, // width in bytes. number of bytes to transfer per row
-                                  size_t      nrows) {
+#if __HVX_ARCH__ < 73
+static const uint32_t dma_src_l2_bypass_on = 1;
+static const uint32_t dma_dst_l2_bypass_on = 0;
+#else
+static const uint32_t dma_src_l2_bypass_on = 1;
+static const uint32_t dma_dst_l2_bypass_on = 1;
+#endif
+
+static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t size) {
     if (((q->push_idx + 1) & q->idx_mask) == q->pop_idx) {
-        FARF(ERROR, "dma-push: queue full\n");
+        FARF(HIGH, "dma-push: queue full\n");
         return false;
     }
 
-    hexagon_udma_descriptor_type1_t * desc = &q->desc[q->push_idx];
+    dma_descriptor_1d * desc = (dma_descriptor_1d *) &q->desc[q->push_idx];
+    desc->next       = NULL;
+    desc->desc_size  = 0; // 1D mode
+    desc->src_bypass = dma_src_l2_bypass_on;
+    desc->dst_bypass = dma_dst_l2_bypass_on;
+    desc->order      = 1;
+    desc->done       = 0;
+    desc->src        = (void *) dptr.src;
+    desc->dst        = (void *) dptr.dst;
+    desc->size       = size;
+
+    q->dptr[q->push_idx] = dptr;
+
+    dmlink(q->tail, desc);
+    q->tail = (dma_descriptor_2d *) desc;
+
+    // FARF(ERROR, "dma-push: i %u row-size %u nrows %d dst %p src %p\n", q->push_idx, row_size, nrows, dptr.dst, dptr.src);
+    q->push_idx = (q->push_idx + 1) & q->idx_mask;
+    return true;
+}
+
+static inline bool dma_queue_push_single_2d(dma_queue * q, dma_ptr dptr, size_t dst_stride, size_t src_stride, size_t row_size, size_t nrows) {
+    if (((q->push_idx + 1) & q->idx_mask) == q->pop_idx) {
+        FARF(HIGH, "dma-push: queue full\n");
+        return false;
+    }
+
+    dma_descriptor_2d * desc = &q->desc[q->push_idx];
 
     desc->next           = NULL;
-    desc->length         = 0;
-    desc->desctype       = HEXAGON_UDMA_DESC_DESCTYPE_TYPE1;
-    desc->dstbypass      = 1;
-    desc->srcbypass      = 1;
-#if __HVX_ARCH__ >= 73
-    desc->dstbypass      = 1;
-    desc->srcbypass      = 1;
-#else
-    desc->dstbypass      = 0;
-    desc->srcbypass      = 1;
-#endif
-    desc->order          = 0;
-    desc->dstate         = HEXAGON_UDMA_DESC_DSTATE_INCOMPLETE;
+    desc->reserved0      = 0;
+    desc->reserved1      = 0;
+    desc->desc_size      = 1; // 2d mode
+    desc->src_bypass     = dma_src_l2_bypass_on;
+    desc->dst_bypass     = dma_dst_l2_bypass_on;
+    desc->src_comp       = 0;
+    desc->dst_comp       = 0;
+    desc->order          = 1;
+    desc->done           = 0;
+    desc->src_stride     = src_stride;
+    desc->dst_stride     = dst_stride;
     desc->src            = (void *) dptr.src;
     desc->dst            = (void *) dptr.dst;
-    desc->allocation     = 0;
-    desc->padding        = 0;
-    desc->roiwidth       = width;
-    desc->roiheight      = nrows;
-    desc->srcstride      = src_row_size;
-    desc->dststride      = dst_row_size;
-    desc->srcwidthoffset = 0;
-    desc->dstwidthoffset = 0;
+    desc->row_size       = row_size;
+
+#if __HVX_ARCH__ < 75
+    desc->desc_type      = 0; // 2d (16-bit) mode
+    desc->nrows          = nrows;
+    desc->src_offset     = 0;
+    desc->dst_offset     = 0;
+#else
+    desc->desc_type      = 9; // 2d (24-bit) mode
+    desc->nrows_lo       = (nrows & 0xff);
+    desc->nrows_hi       = (nrows >> 8);
+    desc->offset         = 0;
+#endif
 
     q->dptr[q->push_idx] = dptr;
 
     dmlink(q->tail, desc);
     q->tail = desc;
 
-    // FARF(ERROR, "dma-push: i %u width %u nrows %d dst %p src %p\n", q->push_idx, width, nrows, dptr.dst, dptr.src);
+    // FARF(ERROR, "dma-push: i %u row-size %u nrows %d dst %p src %p\n", q->push_idx, row_size, nrows, dptr.dst, dptr.src);
     q->push_idx = (q->push_idx + 1) & q->idx_mask;
     return true;
 }
 
-static inline bool dma_queue_push_ddr_to_vtcm(dma_queue * q,
-                                              dma_ptr     dptr,
-                                              size_t      dst_row_size,
-                                              size_t      src_row_size,
-                                              size_t      nrows) {
-    return dma_queue_push(q, dptr, dst_row_size, src_row_size, src_row_size, nrows);
-}
-
-
-static inline bool dma_queue_push_vtcm_to_ddr(dma_queue * q,
-                                              dma_ptr     dptr,
-                                              size_t      dst_row_size,
-                                              size_t      src_row_size,
-                                              size_t      nrows) {
-    return dma_queue_push(q, dptr, dst_row_size, src_row_size, dst_row_size, nrows);
-}
-
 static inline dma_ptr dma_queue_pop(dma_queue * q) {
     dma_ptr dptr  = { NULL };
 
@@ -131,12 +212,12 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {
         return dptr;
     }
 
-    hexagon_udma_descriptor_type1_t * desc = &q->desc[q->pop_idx];
+    dma_descriptor_2d * desc = &q->desc[q->pop_idx];
 
     // Wait for desc to complete
     while (1) {
         dmpoll();
-        if (desc->dstate == HEXAGON_UDMA_DESC_DSTATE_COMPLETE) {
+        if (desc->done) {
             break;
         }
         // FARF(ERROR, "dma-pop: waiting for DMA : %u\n", q->pop_idx);
@@ -175,86 +256,62 @@ static inline uint32_t dma_queue_capacity(dma_queue * q) {
     return q->capacity;
 }
 
-// ---------------------------------------------------------------------------
-// Overflow-safe DMA push: all UDMA type1 descriptor fields (roiwidth,
-// roiheight, srcstride, dststride) are 16-bit, max 65535.  This helper
-// transparently handles values that exceed the 16-bit limit and submits
-// chained DMA transtions.
-//
-// Case 1 (fast path): all params fit in 16 bits -> direct dma_queue_push.
-// Case 2 (contiguous block): width == srcstride == dststride.  Reshape the
-//   flat transfer into a 2D descriptor with sub_width <= 65535.  Produces a
-//   single descriptor, preserving async DMA behavior.
-// Case 3 (stride overflow): srcstride or dststride > 65535.  Issue rows
-//   one at a time.  The first N-1 rows are pushed+popped synchronously;
-//   the last row is left async so the caller can pop it.
-// ---------------------------------------------------------------------------
-#define UDMA_MAX_FIELD_VAL 65535u
+#if __HVX_ARCH__ < 75
 
-static inline bool dma_queue_push_chained(dma_queue *q, dma_ptr dptr, size_t dst_stride, size_t src_stride, size_t width, size_t nrows) {
-    // Fast path: everything fits in 16 bits.
-    if (__builtin_expect(
-            width      <= UDMA_MAX_FIELD_VAL &&
-            nrows      <= UDMA_MAX_FIELD_VAL &&
-            src_stride <= UDMA_MAX_FIELD_VAL &&
-            dst_stride <= UDMA_MAX_FIELD_VAL, 1)) {
-        return dma_queue_push(q, dptr, dst_stride, src_stride, width, nrows);
+// Overflow-safe DMA push: all 2d descriptor fields (row_size, nrows, src_stride, dst_stride) are 16-bit, max 65535.
+// This version transparently handles values that exceed the 16-bit limit and submits chained DMA transtions.
+
+#define DMA_MAX_FIELD_VAL 65535u
+
+static inline bool dma_queue_push(dma_queue *q, dma_ptr dptr, size_t dst_stride, size_t src_stride, size_t row_size, size_t nrows) {
+    // Fast path: everything fits in 16 bits
+    if (nrows == 0 || __builtin_expect(
+            row_size   <= DMA_MAX_FIELD_VAL &&
+            nrows      <= DMA_MAX_FIELD_VAL &&
+            src_stride <= DMA_MAX_FIELD_VAL &&
+            dst_stride <= DMA_MAX_FIELD_VAL, 1)) {
+        return dma_queue_push_single_2d(q, dptr, dst_stride, src_stride, row_size, nrows);
     }
 
-    // Case 2: contiguous block (width == src_stride == dst_stride).
-    // Reshape total bytes into sub_width * sub_nrows where sub_width <= 65535.
-    if (width == src_stride && width == dst_stride) {
-        size_t total = width * nrows;
-
-        // Pick the largest 128-byte-aligned sub_width that divides total evenly.
-        size_t sub_width = UDMA_MAX_FIELD_VAL & ~(size_t)127;  // 65408
-        while (sub_width > 0 && total % sub_width != 0) {
-            sub_width -= 128;
-        }
-        if (sub_width == 0) {
-            // Fallback: use original width (must fit) with adjusted nrows.
-            // This shouldn't happen for 128-aligned DMA sizes.
-            sub_width = width;
-        }
-        size_t sub_nrows = total / sub_width;
-
-        // Handle sub_nrows > 65535 by issuing chunked descriptors.
-        const uint8_t *src = (const uint8_t *)dptr.src;
-        uint8_t       *dst = (uint8_t *)dptr.dst;
-        size_t rows_done = 0;
-        while (rows_done < sub_nrows) {
-            size_t chunk = sub_nrows - rows_done;
-            if (chunk > UDMA_MAX_FIELD_VAL) chunk = UDMA_MAX_FIELD_VAL;
-
-            dma_ptr p = dma_make_ptr(dst + rows_done * sub_width, src + rows_done * sub_width);
-            if (!dma_queue_push(q, p, sub_width, sub_width, sub_width, chunk))
-                return false;
-
-            rows_done += chunk;
-            // Complete all chunks without waiting except the last one, so the
-            // caller's single dma_queue_pop drains the final descriptor.
-            if (rows_done < sub_nrows)
-                dma_queue_pop_nowait(q);
-        }
-        return true;
+    // Contiguous block
+    // Use 1d DMA mode which supports sizes up to 24-bits (16MB)
+    if (nrows == 1 || (row_size == src_stride && row_size == dst_stride)) {
+        size_t total = row_size * nrows;
+        return dma_queue_push_single_1d(q, dptr, total);
     }
 
-    // Case 3: stride overflow — fall back to row-by-row.
+    // Stride overflow — fall back to row-by-row.
     {
-        const uint8_t *src = (const uint8_t *)dptr.src;
-        uint8_t       *dst = (uint8_t *)dptr.dst;
+        const uint8_t *src = (const uint8_t *) dptr.src;
+        uint8_t       *dst = (uint8_t *)       dptr.dst;
         for (size_t r = 0; r < nrows; ++r) {
-          dma_ptr p = dma_make_ptr(dst + r * dst_stride,
-                                   src + r * src_stride);
-          if (!dma_queue_push(q, p, 0, 0, width, 1))
-            return false;
-          if (r + 1 < nrows)
-            dma_queue_pop_nowait(q);
+            dma_ptr p = dma_make_ptr(dst + r * dst_stride, src + r * src_stride);
+            if (!dma_queue_push_single_1d(q, p, row_size))
+                return false;
+            if (r + 1 < nrows)
+                dma_queue_pop(q);
         }
         return true;
     }
 }
 
+#else // HVX_ARCH >= 75
+
+static inline bool dma_queue_push(dma_queue *q, dma_ptr dptr, size_t dst_stride, size_t src_stride, size_t row_size, size_t nrows) {
+    // On v75 and up we always use 2d 24-bit mode
+    return dma_queue_push_single_2d(q, dptr, dst_stride, src_stride, row_size, nrows);
+}
+
+#endif
+
+static inline bool dma_queue_push_ddr_to_vtcm(dma_queue * q, dma_ptr dptr, size_t dst_row_size, size_t src_row_size, size_t nrows) {
+    return dma_queue_push(q, dptr, dst_row_size, src_row_size, src_row_size, nrows);
+}
+
+static inline bool dma_queue_push_vtcm_to_ddr(dma_queue * q, dma_ptr dptr, size_t dst_row_size, size_t src_row_size, size_t nrows) {
+    return dma_queue_push(q, dptr, dst_row_size, src_row_size, dst_row_size, nrows);
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif

ggml/src/ggml-hexagon/htp/hex-dump.h

@@ -21,6 +21,15 @@ static inline void hex_dump_uint8_line(char * pref, const uint8_t * x, uint32_t
     FARF(HIGH, "%s\n", str);
 }
 
+static inline void hex_dump_uint32_line(char * pref, const uint32_t * x, uint32_t n) {
+    char str[1024], *p = str, *p_end = str + sizeof(str);
+    p += snprintf(p, p_end - p, "%s: ", pref);
+    for (int i = 0; i < n; i++) {
+        p += snprintf(p, p_end - p, "%u, ", (unsigned int) x[i]);
+    }
+    FARF(HIGH, "%s\n", str);
+}
+
 static inline void hex_dump_int32_line(char * pref, const int32_t * x, uint32_t n) {
     char str[1024], *p = str, *p_end = str + sizeof(str);
     p += snprintf(p, p_end - p, "%s: ", pref);

ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c

@@ -727,7 +727,7 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                     if (use_dma_activation) {
                         const size_t row_bytes    = (size_t) params->k * sizeof(float);
                         const size_t stride_bytes = (size_t) params->act_stride * sizeof(float);
-                        dma_queue_push_chained(ctx->dma[0],
+                        dma_queue_push(ctx->dma[0],
                                           dma_make_ptr(vtcm_f32_act, activation_chunk),
                                           row_bytes, stride_bytes, row_bytes, n_rows);
                         dma_queue_pop(ctx->dma[0]);
@@ -747,7 +747,7 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
 
                 {
                     const size_t n_cols_first = hex_smin((size_t) params->n, n_chunk_n_cols);
-                    dma_queue_push_chained(ctx->dma[0], dma_make_ptr(buf_curr, weight_group),
+                    dma_queue_push(ctx->dma[0], dma_make_ptr(buf_curr, weight_group),
                                       fp16_row_bytes, weight_row_bytes, fp16_row_bytes, n_cols_first);
                 }
 
@@ -765,7 +765,7 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                             const size_t n_cols_next = hex_smin((size_t) params->n - nc_next, n_chunk_n_cols);
                             const __fp16 *next_weight_chunk = weight_group + nc_next * params->weight_stride;
 
-                            dma_queue_push_chained(ctx->dma[0], dma_make_ptr(buf_next, next_weight_chunk),
+                            dma_queue_push(ctx->dma[0], dma_make_ptr(buf_next, next_weight_chunk),
                                               fp16_row_bytes, weight_row_bytes, fp16_row_bytes, n_cols_next);
                         }
 
@@ -891,7 +891,7 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
             if (use_dma_activation) {
                 const size_t row_bytes    = (size_t) k * sizeof(float);
                 const size_t stride_bytes = (size_t) act_stride * sizeof(float);
-                dma_queue_push_chained(ctx->dma[0],
+                dma_queue_push(ctx->dma[0],
                                   dma_make_ptr(vtcm_f32_act, activation_chunk),
                                   row_bytes, stride_bytes, row_bytes, n_rows);
                 dma_queue_pop(ctx->dma[0]);
@@ -916,7 +916,7 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
         {
             const size_t n_cols_first = hex_smin(n, n_chunk_n_cols);
 
-            dma_queue_push_chained(ctx->dma[0], dma_make_ptr(buf_curr, permuted_weight),
+            dma_queue_push(ctx->dma[0], dma_make_ptr(buf_curr, permuted_weight),
                               fp16_row_bytes, weight_row_bytes, fp16_row_bytes, n_cols_first);
         }
 
@@ -933,7 +933,7 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
                     const size_t n_cols_next       = hex_smin(n - nc_next, n_chunk_n_cols);
                     const __fp16 *next_weight_chunk = permuted_weight + nc_next * weight_stride;
 
-                    dma_queue_push_chained(ctx->dma[0], dma_make_ptr(buf_next, next_weight_chunk),
+                    dma_queue_push(ctx->dma[0], dma_make_ptr(buf_next, next_weight_chunk),
                                       fp16_row_bytes, weight_row_bytes, fp16_row_bytes, n_cols_next);
                 }
 
@@ -1104,7 +1104,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
             // because UDMA roiwidth is 16-bit and total size can exceed 65535.
             {
                 const size_t n_cols_first = hex_smin(n, n_chunk_n_cols);
-                dma_queue_push_chained(ctx->dma[0], dma_make_ptr(buf_curr, permuted_weight), row_stride, row_stride, row_stride, n_cols_first);
+                dma_queue_push(ctx->dma[0], dma_make_ptr(buf_curr, permuted_weight), row_stride, row_stride, row_stride, n_cols_first);
             }
 
             for (size_t nc = 0; nc < n; nc += n_chunk_n_cols) {
@@ -1120,7 +1120,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
 
                         const uint8_t *next_weight_chunk = permuted_weight + nc_next * row_stride;
 
-                        dma_queue_push_chained(ctx->dma[0], dma_make_ptr(buf_next, next_weight_chunk), row_stride, row_stride, row_stride, n_cols_next);
+                        dma_queue_push(ctx->dma[0], dma_make_ptr(buf_next, next_weight_chunk), row_stride, row_stride, row_stride, n_cols_next);
                     }
 
                     // Dequant + vscatter writes directly to [K, N] transposed tiles.
@@ -1173,7 +1173,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
             {
                 // Use 2D DMA (n_cols rows x row_stride) to avoid 16-bit roiwidth overflow.
                 const uint8_t *qweight_chunk_A0 = permuted_weight;
-                dma_queue_push_chained(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_A0), row_stride, row_stride, row_stride, n_cols_A0);
+                dma_queue_push(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_A0), row_stride, row_stride, row_stride, n_cols_A0);
             }
 
             {
@@ -1191,7 +1191,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
                 const size_t n_cols_A1 = hex_smin(n - 1 * n_chunk_n_cols, n_chunk_n_cols);
                 if (1 < n_chunk_cnt) {
                     const uint8_t *qweight_chunk_A1 = permuted_weight + n_chunk_n_cols * row_stride;
-                    dma_queue_push_chained(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_A1), row_stride, row_stride, row_stride, n_cols_A1);
+                    dma_queue_push(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_A1), row_stride, row_stride, row_stride, n_cols_A1);
                 }
 
                 // C0
@@ -1218,7 +1218,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
                 // issue A_{i+2}
                 if (i + 2 < n_chunk_cnt) {
                     const uint8_t *qweight_chunk_p2 = permuted_weight + nc_p2 * row_stride;
-                    dma_queue_push_chained(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_p2), row_stride, row_stride, row_stride, n_cols_p2);
+                    dma_queue_push(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_p2), row_stride, row_stride, row_stride, n_cols_p2);
                 }
 
                 // wait for HMX (C_{i}) -- C_{i} is done
@@ -1443,7 +1443,7 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                 {
                     const float *activation_block = x + mr * k + kk;
 
-                    dma_queue_push_chained(ctx->dma[0],
+                    dma_queue_push(ctx->dma[0],
                                      dma_make_ptr(vtcm_scratch1, activation_block),
                                      k_blk_sz * sizeof(float),
                                      k * sizeof(float),
@@ -1472,10 +1472,10 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                     s.scale_width = nb_sub * HMX_X4X2_DBLK_SIZE;
 
                     // 2D DMA: quants sub-range
-                    dma_queue_push_chained(ctx->dma[0], dma_make_ptr(s.dst, s.src + s.quant_off),
+                    dma_queue_push(ctx->dma[0], dma_make_ptr(s.dst, s.src + s.quant_off),
                                       s.dst_stride, s.src_stride, s.quant_width, s.n_rows);
                     // 2D DMA: scales sub-range
-                    dma_queue_push_chained(ctx->dma[0], dma_make_ptr(s.dst + s.quant_width, s.src + s.scale_off),
+                    dma_queue_push(ctx->dma[0], dma_make_ptr(s.dst + s.quant_width, s.src + s.scale_off),
                                       s.dst_stride, s.src_stride, s.scale_width, s.n_rows);
                 }
                 TIMER_STOP(fetch);

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -15,12 +15,4 @@
 #include "hvx-div.h"
 #include "hvx-base.h"
 
-#ifndef GATHER_TYPE
-#    if defined(__hexagon__)
-#        define GATHER_TYPE(_a) (intptr_t) _a
-#    else
-#        define GATHER_TYPE(_a) (HVX_Vector *) _a
-#    endif
-#endif
-
 #endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -214,7 +214,7 @@ static int vtcm_alloc(struct htp_context * ctx) {
     HAP_compute_res_attr_init(&attr);
     HAP_compute_res_attr_set_serialize(&attr, 0);
     HAP_compute_res_attr_set_cache_mode(&attr, 1);
-    HAP_compute_res_attr_set_vtcm_param_v2(&attr, vtcm_size, 0, vtcm_size);
+    HAP_compute_res_attr_set_vtcm_param_v2(&attr, vtcm_size, vtcm_size, vtcm_size); // single page
     HAP_compute_res_attr_set_release_callback(&attr, vtcm_release_callback, (void *) ctx);
     HAP_compute_res_attr_set_hmx_param(&attr, 1);
 
@@ -319,7 +319,7 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
     ctx->n_threads = n_hvx;
     for (int i = 0; i < ctx->n_threads; i++) {
         // see discussion https://github.com/ggml-org/llama.cpp/pull/18151#discussion_r2632388541
-        ctx->dma[i] = dma_queue_create(64);
+        ctx->dma[i] = dma_queue_create(128);
     }
 
     // init worker pool

ggml/src/ggml-hexagon/htp/ssm-conv.c

@@ -151,7 +151,7 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
     const int dr = scctx->nrows_per_thread;
     const uint32_t ir0 = dr * ith;
     const uint32_t ir1 = MIN(ir0 + dr, d_inner);
-    const int      ir  = ir1 - ir0;
+    const uint32_t ir  = ir1 - ir0;
 
     if (ir0 >= ir1) {
         return;  // No work for this thread
@@ -205,10 +205,10 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
                 HVX_Vector acc_vec = Q6_V_vsplat_R(0);
 
                 for (uint32_t i0 = 0; i0 < d_conv; ++i0) {
-                    Q6_vgather_ARMVw(src0_vec, GATHER_TYPE(spad_src0 + (i0 + i1 * ncs) * sizeof(float) + i2 * (src0->nb[0])),
-                                     src0_gather_len, (*(const HVX_Vector *) src0_offsets));
-                    Q6_vgather_ARMVw(src1_vec, GATHER_TYPE(spad_src1 + (i0 + i1 * nc) * sizeof(float)),
-                                     src1_gather_len, (*(const HVX_Vector *) src1_offsets));
+                    uint32_t src0_base = (uint32_t) spad_src0 + (i0 + i1 * ncs) * sizeof(float) + i2 * (src0->nb[0]);
+                    uint32_t src1_base = (uint32_t) spad_src1 + (i0 + i1 * nc)  * sizeof(float);
+                    Q6_vgather_ARMVw(src0_vec, src0_base, src0_gather_len, (*(const HVX_Vector *) src0_offsets));
+                    Q6_vgather_ARMVw(src1_vec, src1_base, src1_gather_len, (*(const HVX_Vector *) src1_offsets));
 
                     HVX_Vector prod = Q6_Vqf32_vmpy_VsfVsf(*(const HVX_Vector *) src0_vec, *(const HVX_Vector *) src1_vec);
                     acc_vec = Q6_Vqf32_vadd_Vqf32Vqf32(acc_vec, prod);
@@ -222,10 +222,10 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
                 HVX_Vector acc_vec = Q6_V_vsplat_R(0);
 
                 for (uint32_t i0 = 0; i0 < d_conv; ++i0) {
-                    Q6_vgather_ARMVw(src0_vec, GATHER_TYPE(spad_src0 + (i0 + i1 * ncs) * sizeof(float) + i2 * (src0->nb[0])),
-                                     src0_gather_len, (*(const HVX_Vector *) src0_offsets));
-                    Q6_vgather_ARMVw(src1_vec, GATHER_TYPE(spad_src1 + (i0 + i1 * nc) * sizeof(float)),
-                                     src1_gather_len, (*(const HVX_Vector *) src1_offsets));
+                    uint32_t src0_base = (uint32_t) spad_src0 + (i0 + i1 * ncs) * sizeof(float) + i2 * (src0->nb[0]);
+                    uint32_t src1_base = (uint32_t) spad_src1 + (i0 + i1 * nc)  * sizeof(float);
+                    Q6_vgather_ARMVw(src0_vec, src0_base, src0_gather_len, (*(const HVX_Vector *) src0_offsets));
+                    Q6_vgather_ARMVw(src1_vec, src1_base, src1_gather_len, (*(const HVX_Vector *) src1_offsets));
 
                     HVX_Vector prod = Q6_Vqf32_vmpy_VsfVsf(*(const HVX_Vector *) src0_vec, *(const HVX_Vector *) src1_vec);
                     acc_vec = Q6_Vqf32_vadd_Vqf32Vqf32(acc_vec, prod);

ggml/src/ggml-hip/CMakeLists.txt

@@ -71,12 +71,11 @@ if (GGML_CUDA_FA_ALL_QUANTS)
     list(APPEND GGML_SOURCES_ROCM ${SRCS})
     add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
 else()
-    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    list(APPEND GGML_SOURCES_ROCM
+        ../ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu
+        ../ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu
+        ../ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu
+        ../ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu)
 endif()
 
 ggml_add_backend_library(ggml-hip

ggml/src/ggml-impl.h

@@ -773,6 +773,5 @@ inline bool ggml_check_edges(const struct ggml_cgraph *                cgraph,
 
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);
-GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);
 GGML_API void gguf_write_to_buf(const struct gguf_context * ctx, std::vector<int8_t> & buf, bool only_meta);
 #endif // __cplusplus

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -246,6 +246,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary(ggml_metal
                 case GGML_UNARY_OP_EXP:         op_num = OP_UNARY_NUM_EXP;         break;
                 case GGML_UNARY_OP_SOFTPLUS:    op_num = OP_UNARY_NUM_SOFTPLUS;    break;
                 case GGML_UNARY_OP_EXPM1:       op_num = OP_UNARY_NUM_EXPM1;       break;
+                case GGML_UNARY_OP_FLOOR:       op_num = OP_UNARY_NUM_FLOOR;       break;
+                case GGML_UNARY_OP_CEIL:        op_num = OP_UNARY_NUM_CEIL;        break;
+                case GGML_UNARY_OP_ROUND:       op_num = OP_UNARY_NUM_ROUND;       break;
+                case GGML_UNARY_OP_TRUNC:       op_num = OP_UNARY_NUM_TRUNC;       break;
                 default: GGML_ABORT("fatal error");
             } break;
         default: GGML_ABORT("fatal error");
@@ -1748,6 +1752,28 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_2d(ggml_met
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_3d(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_CONV_3D);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type         == GGML_TYPE_F32);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_conv_3d_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_UPSCALE);
 

ggml/src/ggml-metal/ggml-metal-device.h

@@ -148,6 +148,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_3d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1039,6 +1039,10 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_EXPM1:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
                 default:
                     return false;
@@ -1077,6 +1081,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32) &&
                 op->src[1]->type == GGML_TYPE_F32 &&
                 op->type == GGML_TYPE_F32;
+        case GGML_OP_CONV_3D:
+            return ggml_is_contiguous(op->src[0]) &&
+                   ggml_is_contiguous(op->src[1]) &&
+                   (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32) &&
+                   op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_SUM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
         case GGML_OP_TRI:
@@ -1143,6 +1152,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 op->src[0]->ne[0] != 192 &&
                 op->src[0]->ne[0] != 256 &&
                 op->src[0]->ne[0] != 320 &&
+                op->src[0]->ne[0] != 512 &&
                 op->src[0]->ne[0] != 576) {
                 return false;
             }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -120,6 +120,10 @@
 #define OP_UNARY_NUM_EXP         114
 #define OP_UNARY_NUM_SOFTPLUS    115
 #define OP_UNARY_NUM_EXPM1       116
+#define OP_UNARY_NUM_FLOOR       117
+#define OP_UNARY_NUM_CEIL        118
+#define OP_UNARY_NUM_ROUND       119
+#define OP_UNARY_NUM_TRUNC       120
 
 #define OP_SUM_ROWS_NUM_SUM_ROWS 10
 #define OP_SUM_ROWS_NUM_MEAN     11
@@ -643,6 +647,42 @@ typedef struct {
     int32_t  KHW; // KH * KW, pre-computed on CPU to save GPU resources
 } ggml_metal_kargs_im2col;
 
+typedef struct {
+    int32_t  IW;
+    int32_t  IH;
+    int32_t  ID;
+    int32_t  OW;
+    int32_t  OH;
+    int32_t  OD;
+    int32_t  KW;
+    int32_t  KH;
+    int32_t  KD;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  s2;
+    int32_t  p0;
+    int32_t  p1;
+    int32_t  p2;
+    int32_t  d0;
+    int32_t  d1;
+    int32_t  d2;
+    int32_t  IC;
+    int32_t  N;
+    int32_t  OC;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_conv_3d;
+
 typedef struct{
     int32_t  ne00;
     uint64_t nb01;

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -394,6 +394,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_conv_transpose_2d(ctx, idx);
             } break;
+        case GGML_OP_CONV_3D:
+            {
+                n_fuse = ggml_metal_op_conv_3d(ctx, idx);
+            } break;
         case GGML_OP_UPSCALE:
             {
                 n_fuse = ggml_metal_op_upscale(ctx, idx);
@@ -3697,6 +3701,77 @@ int ggml_metal_op_conv_2d(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_conv_3d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    // 1. Extract standard dimensions and byte strides
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    // 2. Extract hyperparams from op_params
+    const int32_t s0 = ((const int32_t *)(op->op_params))[0];
+    const int32_t s1 = ((const int32_t *)(op->op_params))[1];
+    const int32_t s2 = ((const int32_t *)(op->op_params))[2];
+    const int32_t p0 = ((const int32_t *)(op->op_params))[3];
+    const int32_t p1 = ((const int32_t *)(op->op_params))[4];
+    const int32_t p2 = ((const int32_t *)(op->op_params))[5];
+    const int32_t d0 = ((const int32_t *)(op->op_params))[6];
+    const int32_t d1 = ((const int32_t *)(op->op_params))[7];
+    const int32_t d2 = ((const int32_t *)(op->op_params))[8];
+    const int32_t IC = ((const int32_t *)(op->op_params))[9];
+    const int32_t N  = ((const int32_t *)(op->op_params))[10];
+    const int32_t OC = ((const int32_t *)(op->op_params))[11];
+
+    // 3. Build the parameter struct using the macro-generated variables
+    ggml_metal_kargs_conv_3d args = {
+        /*.IW =*/ (int32_t)op->src[1]->ne[0],
+        /*.IH =*/ (int32_t)op->src[1]->ne[1],
+        /*.ID =*/ (int32_t)op->src[1]->ne[2],
+        /*.OW =*/ (int32_t)op->ne[0],
+        /*.OH =*/ (int32_t)op->ne[1],
+        /*.OD =*/ (int32_t)op->ne[2],
+        /*.KW =*/ (int32_t)op->src[0]->ne[0],
+        /*.KH =*/ (int32_t)op->src[0]->ne[1],
+        /*.KD =*/ (int32_t)op->src[0]->ne[2],
+        s0, s1, s2,
+        p0, p1, p2,
+        d0, d1, d2,
+        IC, N, OC,
+        nb00, nb01, nb02, nb03, // Weight strides
+        nb10, nb11, nb12, nb13, // Input strides
+        nb0,  nb1,  nb2,  nb3   // Output strides
+    };
+
+    // 4. Fetch the JIT pipeline
+    auto pipeline = ggml_metal_library_get_pipeline_conv_3d(lib, op);
+
+    // 5. Grid mapping
+    int nth0 = 32; // Standard SIMD width for Apple Silicon
+    int nth1 = 1;
+    int nth2 = 1;
+
+    int64_t spatial_volume = args.OW * args.OH * args.OD;
+
+    int ntg0 = (spatial_volume + nth0 - 1) / nth0;
+    int ntg1 = args.OC;
+    int ntg2 = args.N;
+
+    // 6. Bind and Dispatch via the ggml C wrapper
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ntg0, ntg1, ntg2, nth0, nth1, nth2);
+
+    return 1;
+}
+
 int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -75,6 +75,7 @@ int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rope              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_im2col            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_2d           (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_conv_3d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_1d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_2d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_upscale           (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/ggml-metal.metal

@@ -1094,6 +1094,22 @@ kernel void kernel_unary_impl(
             // TODO: precise implementation
             dst_ptr[i0] = (T) (exp(x) - 1);
         }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
     }
 
 #undef FC_OP
@@ -4883,6 +4899,98 @@ kernel void kernel_upscale_bilinear_f32(
     }
 }
 
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+
 static inline float bicubic_weight1(float x) {
     const float a = -0.75f;
     return ((a + 2) * x - (a + 3)) * x * x + 1;
@@ -6177,6 +6285,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk192_dv192")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f32_dk192_dv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f32_dk256_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  256, 256>;
 template [[host_name("kernel_flash_attn_ext_f32_dk320_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  320, 256>;
+template [[host_name("kernel_flash_attn_ext_f32_dk512_dv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  512, 512>;
 template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
@@ -6192,6 +6301,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk192_dv192")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_dk192_dv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f16_dk256_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256, 256>;
 template [[host_name("kernel_flash_attn_ext_f16_dk320_dv256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  320, 256>;
+template [[host_name("kernel_flash_attn_ext_f16_dk512_dv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  512, 512>;
 template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_HAS_BF16)
@@ -6208,6 +6318,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk192_dv192")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 320, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk512_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 512, 512>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
@@ -6224,6 +6335,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk192_dv192")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 320, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk512_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 512, 512>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
@@ -6239,6 +6351,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk192_dv192")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 320, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk512_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 512, 512>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
@@ -6254,6 +6367,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk192_dv192")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 320, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk512_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 512, 512>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
@@ -6269,6 +6383,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk192_dv192")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 320, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk512_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 512, 512>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
@@ -6284,6 +6399,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_dk192_dv192")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 320, 256>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk512_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 512, 512>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
@@ -6865,6 +6981,17 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_dk320_dv256")]] kernel flas
 template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 320, 256, 2>;
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 320, 256, 2>;
 
+template [[host_name("kernel_flash_attn_ext_vec_f32_dk512_dv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES_F32, float4,     1, dequantize_f32_t4,  float4,      1, dequantize_f32_t4,  512, 512, 1>;
+template [[host_name("kernel_flash_attn_ext_vec_f16_dk512_dv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  512, 512, 1>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_dk512_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     bfloat4,    1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 512, 512, 1>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_dk512_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q4_0, 8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 512, 512, 1>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_dk512_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q4_1, 8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 512, 512, 1>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_dk512_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q5_0, 8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 512, 512, 1>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk512_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 512, 512, 1>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk512_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 512, 512, 1>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f32_dk576_dv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES_F32, float4,     1, dequantize_f32_t4,  float4,      1, dequantize_f32_t4,  576, 512, 2>;
 template [[host_name("kernel_flash_attn_ext_vec_f16_dk576_dv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  576, 512, 2>;
 #if defined(GGML_METAL_HAS_BF16)

ggml/src/ggml-musa/CMakeLists.txt

@@ -48,12 +48,11 @@ if (MUSAToolkit_FOUND)
         list(APPEND GGML_SOURCES_MUSA ${SRCS})
         add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
     else()
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        list(APPEND GGML_SOURCES_MUSA
+            ../ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu
+            ../ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu
+            ../ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu
+            ../ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu)
     endif()
 
     set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)

ggml/src/ggml-opencl/CMakeLists.txt

@@ -89,6 +89,7 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_1_f32
     mul_mv_q4_1_f32_flat
     mul_mv_q4_k_f32
+    mul_mv_q4_k_f32_flat
     mul_mv_q6_k_f32
     mul_mv_q6_k_f32_flat
     mul_mv_q8_0_f32
@@ -107,11 +108,14 @@ set(GGML_OPENCL_KERNELS
     mul_mm_q4_0_f32_l4_lm
     mul_mm_q4_1_f32_l4_lm
     mul_mm_q8_0_f32_l4_lm
+    mul_mm_q4_k_f32_l4_lm
     mul_mm_q6_k_f32_l4_lm
     mul_mm_q8_0_f32_8x4
     gemv_noshuffle_q4_1_f32
     gemm_noshuffle_q4_1_f32
     gemv_noshuffle_general_q8_0_f32
+    gemv_noshuffle_q6_k_f32
+    gemm_noshuffle_q6_k_f32
     mul
     neg
     norm

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -529,16 +529,19 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_convert_block_q4_1, kernel_restore_block_q4_1;
     cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4, kernel_restore_block_mxfp4_trans;
     cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0, kernel_restore_block_q8_0_trans;
+    cl_kernel kernel_convert_block_q6_K_noshuffle, kernel_restore_block_q6_K_noshuffle;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_restore_block_q4_0_noshuffle;
     cl_kernel kernel_convert_block_q4_1_noshuffle;
     cl_kernel kernel_restore_block_q4_1_noshuffle;
+    cl_kernel kernel_convert_block_q4_K, kernel_restore_block_q4_K;
     cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q4_1_f32;
     cl_kernel kernel_mul_mv_q4_1_f32_flat;
     cl_kernel kernel_mul_mv_q4_K_f32;
+    cl_kernel kernel_mul_mv_q4_K_f32_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
     cl_kernel kernel_mul_mv_q6_K_f32_flat;
     cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
@@ -578,6 +581,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mm_q4_0_f32_l4_lm;
     cl_kernel kernel_mul_mm_q4_1_f32_l4_lm;
     cl_kernel kernel_mul_mm_q8_0_f32_l4_lm;
+    cl_kernel kernel_mul_mm_q4_k_f32_l4_lm;
     cl_kernel kernel_mul_mm_q6_k_f32_l4_lm;
 
     std::vector<ProfilingInfo> profiling_info;
@@ -713,6 +717,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gemm_noshuffle_q4_1_f32;
     cl_kernel kernel_mul_mm_q8_0_f32_8x4;
     cl_kernel CL_mul_mat_vec_q8_0_f32;
+    cl_kernel kernel_gemv_noshuffle_q6_K_f32;
+    cl_kernel kernel_gemm_noshuffle_q6_K_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
     void free() {
@@ -917,8 +923,12 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_convert_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q8_0_trans  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0_trans", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q4_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_K", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q6_K_noshuffle  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q6_K_noshuffle  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K_noshuffle", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1209,6 +1219,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_q4_k_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_k_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_k_f32_flat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q4_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q4_K_f32_flat", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_q6_k_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1482,6 +1509,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mm_q4_k_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_q4_k_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_q4_k_f32_l4_lm.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_q4_k_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q4_k_f32_l4_lm", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mm_q6_k_f32_l4_lm
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2603,6 +2647,45 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_gemm_moe_mxfp4_f32 = clCreateKernel(backend_ctx->program_gemm_moe_mxfp4_f32, "kernel_gemm_moe_mxfp4_f32", &err), err));
         GGML_LOG_CONT(".");
     }
+
+    // gemv_noshuffle_q6_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_q6_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_q6_k_f32.cl");
+#endif
+
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q6_K_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q6_K_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemm_noshuffle_q6_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_q6_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_q6_k_f32.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_moe_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q6_K_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q6_K_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
     GGML_LOG_CONT("\n");
 }
@@ -3347,6 +3430,40 @@ struct ggml_tensor_extra_cl_q8_0 {
     }
 };
 
+struct ggml_tensor_extra_cl_q4_K {
+    // Quantized values
+    cl_mem q = nullptr;
+    // Scales for each super block.
+    cl_mem s  = nullptr;
+    // Scales
+    cl_mem d = nullptr;
+    // Min
+    cl_mem dm  = nullptr;
+
+    ~ggml_tensor_extra_cl_q4_K() {
+        reset();
+    }
+
+    void reset() {
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (s != nullptr) {
+            CL_CHECK(clReleaseMemObject(s));
+            s = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+        if (dm != nullptr) {
+            CL_CHECK(clReleaseMemObject(dm));
+            dm = nullptr;
+        }
+    }
+};
+
 struct ggml_tensor_extra_cl_q6_K {
     // Lower 4 bits of quantized weights.
     cl_mem ql = nullptr;
@@ -3956,6 +4073,12 @@ struct ggml_backend_opencl_buffer_context {
         for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
             delete e;
         }
+        for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K_in_use) {
+            delete e;
+        }
         for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K) {
             delete e;
         }
@@ -4039,6 +4162,21 @@ struct ggml_backend_opencl_buffer_context {
         return extra;
     }
 
+    ggml_tensor_extra_cl_q4_K * ggml_opencl_alloc_temp_tensor_extra_q4_K() {
+        ggml_tensor_extra_cl_q4_K * extra;
+        if (temp_tensor_extras_q4_K.empty()) {
+            extra = new ggml_tensor_extra_cl_q4_K();
+        } else {
+            extra = temp_tensor_extras_q4_K.back();
+            temp_tensor_extras_q4_K.pop_back();
+        }
+
+        temp_tensor_extras_q4_K_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
     ggml_tensor_extra_cl_q6_K * ggml_opencl_alloc_temp_tensor_extra_q6_K() {
         ggml_tensor_extra_cl_q6_K * extra;
         if (temp_tensor_extras_q6_K.empty()) {
@@ -4080,6 +4218,11 @@ struct ggml_backend_opencl_buffer_context {
         }
         temp_tensor_extras_q8_0_in_use.clear();
 
+        for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K_in_use) {
+            temp_tensor_extras_q4_K.push_back(e);
+        }
+        temp_tensor_extras_q4_K_in_use.clear();
+
         for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
             temp_tensor_extras_q6_K.push_back(e);
         }
@@ -4101,6 +4244,8 @@ struct ggml_backend_opencl_buffer_context {
     std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
+    std::vector<ggml_tensor_extra_cl_q4_K *> temp_tensor_extras_q4_K;
+    std::vector<ggml_tensor_extra_cl_q4_K *> temp_tensor_extras_q4_K_in_use;
     std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K;
     std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K_in_use;
 
@@ -4835,6 +4980,83 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
 
         return;
     }
+    if (tensor->type == GGML_TYPE_Q4_K) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q4_K * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q4_K();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_dm = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_s = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*(3 * ggml_blck_size(tensor->type) / 64);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        GGML_ASSERT(size_d + size_dm + size_s + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        cl_buffer_region region;
+
+        // Create subbuffer for d.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for mins.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_dm;
+        extra->dm = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Create subbuffer for s.
+        region.origin = align_to(previous_origin + size_dm, backend_ctx->alignment);
+        region.size = size_s;
+        extra->s = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q4_K;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->dm));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra  = extra;
+        return;
+    }
     if (tensor->type == GGML_TYPE_Q6_K) {
         ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
         GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
@@ -4851,61 +5073,58 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             "Incorrect tensor size");
 
         cl_int err;
-        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
-            ggml_nbytes(tensor), NULL, &err);
-        CL_CHECK(err);
-        CL_CHECK(clEnqueueWriteBuffer(
-            queue, data_device, CL_TRUE, 0,
-            ggml_nbytes(tensor), data, 0, NULL, NULL));
+        cl_mem data_device;
+        CL_CHECK((data_device = clCreateBuffer(context, CL_MEM_READ_WRITE, ggml_nbytes(tensor), NULL, &err), err));
+        CL_CHECK(clEnqueueWriteBuffer(queue, data_device, CL_TRUE, 0, ggml_nbytes(tensor), data, 0, NULL, NULL));
 
         cl_buffer_region region;
 
         // Subbuffer for ql
         region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
         region.size = size_ql;
-        extra->ql = clCreateSubBuffer(
-            extra_orig->data_device, CL_MEM_READ_WRITE,
-            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
-        CL_CHECK(err);
+        CL_CHECK((extra->ql = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
         auto previous_origin = region.origin;
 
         // Subbuffer for qh
         region.origin = align_to(previous_origin + size_ql, backend_ctx->alignment);
         region.size = size_qh;
-        extra->qh = clCreateSubBuffer(
-            extra_orig->data_device, CL_MEM_READ_WRITE,
-            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
-        CL_CHECK(err);
+        CL_CHECK((extra->qh = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
         previous_origin = region.origin;
 
         // Subbuffer for scales
         region.origin = align_to(previous_origin + size_qh, backend_ctx->alignment);
         region.size = size_s;
-        extra->s = clCreateSubBuffer(
-            extra_orig->data_device, CL_MEM_READ_WRITE,
-            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
-        CL_CHECK(err);
+        CL_CHECK((extra->s = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
         previous_origin = region.origin;
 
         // Create subbuffer for d.
         region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
         region.size = size_d;
-        extra->d = clCreateSubBuffer(
-            extra_orig->data_device, CL_MEM_READ_WRITE,
-            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
-        CL_CHECK(err);
+        CL_CHECK((extra->d = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
         previous_origin = region.origin;
 
         // Flatten the weights
-        cl_kernel kernel = backend_ctx->kernel_convert_block_q6_K;
+        cl_kernel kernel;
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        kernel = backend_ctx->kernel_convert_block_q6_K;
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            kernel = backend_ctx->kernel_convert_block_q6_K_noshuffle;
+        }
+#else
+        kernel = backend_ctx->kernel_convert_block_q6_K;
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql));
-        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
-        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s));
-        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d));
+        cl_uchar mask = 0xff;
+        cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &n_blk));
 
-        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t global_work_size[] = {(size_t)CEIL_DIV(n_blk, 64)*64, 1, 1};
         size_t local_work_size[] = {64, 1, 1};
 
         cl_event evt;
@@ -4919,6 +5138,29 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         extra->size_d  = size_d;
 
         tensor->extra  = extra;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            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);
+
+            // 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);
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
         return;
     }
 #endif // GGML_OPENCL_SOA_Q
@@ -5245,24 +5487,111 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
         CL_CHECK(clReleaseMemObject(data_device));
         return;
     }
-    if (tensor->type == GGML_TYPE_Q6_K) {
-        ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
+    if (tensor->type == GGML_TYPE_Q4_K) {
+        ggml_tensor_extra_cl_q4_K * extra = (ggml_tensor_extra_cl_q4_K *)tensor->extra;
 
         cl_int err;
         cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
             ggml_nbytes(tensor), NULL, &err);
         CL_CHECK(err);
 
-        cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K;
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh));
-        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
-        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q4_K;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->dm));
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
 
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {1, 1, 1};
 
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            static ggml_cl_buffer buf_trans_ql;
+            static ggml_cl_buffer buf_trans_qh;
+            static ggml_cl_buffer buf_trans_s;
+            static ggml_cl_buffer buf_trans_d;
+            static ggml_cl_buffer buf_unpacked;
+
+            cl_int M = tensor->ne[1];   // ne01
+            cl_int K = tensor->ne[0];   // ne00
+
+            GGML_ASSERT(K % ggml_blck_size(tensor->type) == 0);
+
+            size_t size_ql = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+            size_t size_qh = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/4;
+            size_t size_s  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/16;
+            size_t size_d  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+            GGML_ASSERT(size_ql + size_qh + size_s + size_d == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+            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);
+
+            // unpack
+            cl_uchar mask = 0xFF;
+            cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+            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, 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));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &n_blk));
+
+            size_t global_work_size[] = {(size_t)n_blk, 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            cl_event evt;
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+            CL_CHECK(clEnqueueReadBuffer(queue, buf_unpacked.buffer, CL_TRUE, offset, size, data, 0, NULL, NULL));
+
+            return;
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_uchar mask = 0xFF;
+        cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &n_blk));
+
+        size_t global_work_size[] = {(size_t)n_blk, 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
         cl_event evt;
         CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
             global_work_size, local_work_size, 0, NULL, &evt));
@@ -5553,6 +5882,8 @@ typedef struct {
 static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2,
     "wrong q4_0 block size/padding");
 
+#define QK_MXFP4 32
+
 #include <math.h>
 #ifdef __cplusplus
 #include "half.hpp"
@@ -5596,7 +5927,7 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
         buf_d = malloc(size_e);
 
         CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
-        CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->e, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
         CL_CHECK(clFinish(queue));
     } else {
         // Read out the tensor from GPU memory.
@@ -9331,6 +9662,196 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
 #endif
 }
 
+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);
+    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;
+
+    // subbuffer and image for activation
+    if (ne1 == 1) {
+        cl_mem ql_img = nullptr;
+        cl_mem qh_img = nullptr;
+        cl_mem b_sub_buffer = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for ql
+        img_fmt.image_channel_order = CL_R;
+        img_fmt.image_channel_data_type = CL_FLOAT;
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = ne01 * ne00 / 8;
+        img_desc.buffer = extra0_q6_K->ql;
+        CL_CHECK((ql_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // image for qh
+        img_fmt.image_channel_order = CL_R;
+        img_fmt.image_channel_data_type = CL_HALF_FLOAT;
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = ne01 * ne00 / 8;
+        img_desc.buffer = extra0_q6_K->qh;
+        CL_CHECK((qh_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        region.origin = offset1;
+        region.size = ne00 * ne1 * sizeof(float);
+        CL_CHECK((b_sub_buffer = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        img_fmt.image_channel_order = CL_RGBA;
+        img_fmt.image_channel_data_type = CL_FLOAT;
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = ne00 * ne1 / 4;
+        img_desc.buffer = b_sub_buffer;
+        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;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &ql_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &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, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(ql_img));
+        CL_CHECK(clReleaseMemObject(qh_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buffer));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+        cl_mem b_sub_buf;
+        cl_mem b_buf_trans;
+        cl_mem b_img;
+        cl_mem b_img_trans;
+
+        // subbuffer for activation
+        region.origin = offset1;
+        region.size = ne00 * ne1 * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activation
+        img_fmt.image_channel_order = CL_RGBA;
+        img_fmt.image_channel_data_type = CL_FLOAT;
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = ne00 * ne1 / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = ne1 % 8;
+        int padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activation
+        region.origin = 0;
+        region.size = ne00 * (ne1 + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activation
+        img_fmt.image_channel_order = CL_RGBA;
+        img_fmt.image_channel_data_type = CL_HALF_FLOAT;
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = ne00 * (ne1 + padding) / 4;
+        img_desc.buffer = b_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // transpose activation
+        int height_B = ne1/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = ne00/4;
+        int padded_height_B = (ne1 + padding) / 4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_size_t[2] = { 1, 16 };
+        size_t global_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_q6_K_f32;
+        int padded_N = ne1 + padding;
+
+        cl_ushort mask_f000 = 0xF000;
+        cl_uchar  mask_c0   = 0xC0;
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q6_K->ql));
+        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_trans));
+        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(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &padded_N));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ushort),&mask_f000));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_uchar), &mask_c0));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3] = {2, 128, 1};
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -9357,6 +9878,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     ggml_tensor_extra_cl_q4_1 * extra0_q4_1 = (ggml_tensor_extra_cl_q4_1 *)src0->extra;
     ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
     ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+    ggml_tensor_extra_cl_q4_K * extra0_q4_K = (ggml_tensor_extra_cl_q4_K *)src0->extra;
     ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
 #endif
 
@@ -9466,6 +9988,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             return;
     }
 
+    // q6_K x fp32
+    if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32) {
+        ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);
+        return;
+    }
+
     // q4_0 x fp32
     if(src0t == GGML_TYPE_Q4_0 && src1t == GGML_TYPE_F32) {
         // TODO: remove duplicate definitions of image description + format -- move to top
@@ -10005,6 +10533,50 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
                 return;
             }
+            case GGML_TYPE_Q4_K: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_q4_k_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_K->q));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_K->s));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q4_K->d));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q4_K->dm));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
             case GGML_TYPE_Q6_K: {
                 if (ne11 < 32) {
                     break;
@@ -10449,6 +11021,43 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q4_K_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 16;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_K->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_K->s));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q4_K->d));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q4_K->dm));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &r3));
+#else
             kernel = backend_ctx->kernel_mul_mv_q4_K_f32;
 
             if (backend_ctx->gpu_family == INTEL) {
@@ -10482,6 +11091,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),        &ne1));
             CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),        &r2));
             CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),        &r3));
+#endif // GGML_OPENCL_SOA_Q
             break;
         }
         case GGML_TYPE_Q5_K:

ggml/src/ggml-opencl/kernels/cvt.cl

@@ -28,6 +28,7 @@
 #define QK8_0                   32
 #define QR8_0                   1
 #define QK_K                    256
+#define K_SCALE_SIZE            (3 * QK_K / 64)
 #define K_QUANTS_PER_ITERATION  2
 
 typedef char int8_t;
@@ -55,6 +56,16 @@ struct block_q4_1 {
     uchar qs[QK4_1 / 2]; // nibbles / quants
 };
 
+//------------------------------------------------------------------------------
+// block_q4_k
+//------------------------------------------------------------------------------
+struct block_q4_K {
+    half d; // delta
+    half dm; // min
+    uchar s[K_SCALE_SIZE];
+    uchar q[QK_K / 2]; // nibbles / quants
+};
+
 //------------------------------------------------------------------------------
 // block_q6_K
 //------------------------------------------------------------------------------
@@ -408,6 +419,62 @@ kernel void kernel_restore_block_q8_0_trans(
     }
 }
 
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_K
+// Convert the block_q4_K format to 4 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+// Each thread processes a super block.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_K(
+    global struct block_q4_K * src0,
+    global uchar * dst_q,
+    global uchar * dst_s,
+    global half  * dst_d,
+    global half  * dst_dm
+) {
+    global struct block_q4_K * b = (global struct block_q4_K *) src0 + get_global_id(0);
+    global uchar * q  = (global uchar *) dst_q  + QK_K/2*get_global_id(0);
+    global uchar * s  = (global uchar *) dst_s  + K_SCALE_SIZE*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+    global half  * dm = (global half  *) dst_dm + get_global_id(0);
+
+    *d  = b->d;
+    *dm = b->dm;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        q[i] = b->q[i];
+    }
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        s[i] = b->s[i];
+    }
+}
+
+// Restore block_q4_K from flattened arrays.
+// Each thread processes a super block.
+kernel void kernel_restore_block_q4_K(
+    global uchar * src_q,
+    global uchar * src_s,
+    global half  * src_d,
+    global half  * src_dm,
+    global struct block_q4_K * dst
+) {
+    global struct block_q4_K * b = (global struct block_q4_K *) dst + get_global_id(0);
+    global uchar * q  = (global uchar *) src_q  + QK_K/2*get_global_id(0);
+    global uchar * s  = (global uchar *) src_s + K_SCALE_SIZE*get_global_id(0);
+    global half  * d  = (global half  *) src_d  + get_global_id(0);
+    global half  * dm = (global half  *) src_dm  + get_global_id(0);
+
+    b->d  = *d;
+    b->dm = *dm;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        b->q[i] = q[i];
+    }
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        b->s[i] = s[i];
+    }
+}
+
 //------------------------------------------------------------------------------
 // kernel_convert_block_q6_K
 // Convert the block_q6_K format to 3 separate arrays (AOS -> SOA).
@@ -419,8 +486,13 @@ kernel void kernel_convert_block_q6_K(
     global uchar * dst_ql,
     global uchar * dst_qh,
     global char  * dst_s,
-    global half  * dst_d
+    global half  * dst_d,
+    uchar          mask_lsb_8,
+    ulong          n_blk
 ) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
     global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
     global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
     global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
@@ -447,8 +519,13 @@ kernel void kernel_restore_block_q6_K(
     global uchar * dst_qh,
     global char  * dst_s,
     global half  * dst_d,
-    global struct block_q6_K * dst
+    global struct block_q6_K * dst,
+    uchar mask_lsb_8,
+    ulong n_blk
 ) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
     global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
     global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
     global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
@@ -467,3 +544,117 @@ kernel void kernel_restore_block_q6_K(
         b->scales[i] = s[i];
     }
 }
+
+kernel void kernel_convert_block_q6_K_noshuffle(
+    global struct block_q6_K * src0,
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d,
+    uchar          mask_lsb_8,
+    ulong          n_blk
+) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
+    global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK_K/2/4; ++i) {
+        uchar x0 = b->ql[i*2 + 0] & mask_lsb_8;
+        uchar x1 = b->ql[i*2 + 1] & mask_lsb_8;
+        ql[i +  0] = (x0 & 0x0F)        | ((x1 & 0x0F) << 4);
+        ql[i + 32] = ((x0 & 0xF0) >> 4) | (x1 & 0xF0);
+
+        uchar x2 = b->ql[i*2 + 0 + 64] & mask_lsb_8;
+        uchar x3 = b->ql[i*2 + 1 + 64] & mask_lsb_8;
+        ql[i + 64] = (x2 & 0x0F)        | ((x3 & 0x0F) << 4);
+        ql[i + 96] = ((x2 & 0xF0) >> 4) | (x3 & 0xF0);
+    }
+
+    for (int i = 0; i < QK_K/4/8; ++i) {
+        uchar x0 = b->qh[i*4 + 0] & mask_lsb_8;
+        uchar x1 = b->qh[i*4 + 1] & mask_lsb_8;
+        uchar x2 = b->qh[i*4 + 2] & mask_lsb_8;
+        uchar x3 = b->qh[i*4 + 3] & mask_lsb_8;
+        qh[i +  0] = (x0 & 0x03)        | ((x1 & 0x03) << 2) | ((x2 & 0x03) << 4) | ((x3 & 0x03) << 6);
+        qh[i +  8] = ((x0 & 0x0C) >> 2) | (x1 & 0x0C)        | ((x2 & 0x0C) << 2) | ((x3 & 0x0C) << 4);
+        qh[i + 16] = ((x0 & 0x30) >> 4) | ((x1 & 0x30) >> 2) | (x2 & 0x30)        | ((x3 & 0x30) << 2);
+        qh[i + 24] = ((x0 & 0xC0) >> 6) | ((x1 & 0xC0) >> 4) | ((x2 & 0xC0) >> 2) | (x3 & 0xC0);
+
+        uchar x4 = b->qh[i*4 + 0 + 32] & mask_lsb_8;
+        uchar x5 = b->qh[i*4 + 1 + 32] & mask_lsb_8;
+        uchar x6 = b->qh[i*4 + 2 + 32] & mask_lsb_8;
+        uchar x7 = b->qh[i*4 + 3 + 32] & mask_lsb_8;
+        qh[i + 32] = (x4 & 0x03)        | ((x5 & 0x03) << 2) | ((x6 & 0x03) << 4) | ((x7 & 0x03) << 6);
+        qh[i + 40] = ((x4 & 0x0C) >> 2) | (x5 & 0x0C)        | ((x6 & 0x0C) << 2) | ((x7 & 0x0C) << 4);
+        qh[i + 48] = ((x4 & 0x30) >> 4) | ((x5 & 0x30) >> 2) | (x6 & 0x30)        | ((x7 & 0x30) << 2);
+        qh[i + 56] = ((x4 & 0xC0) >> 6) | ((x5 & 0xC0) >> 4) | ((x6 & 0xC0) >> 2) | (x7 & 0xC0);
+    }
+
+    for (int i = 0; i < QK_K/16; ++i) {
+        s[i] = b->scales[i];
+    }
+}
+
+kernel void kernel_restore_block_q6_K_noshuffle(
+    global uchar * src_ql,
+    global uchar * src_qh,
+    global char  * src_s,
+    global half  * src_d,
+    global struct block_q6_K * dst,
+    uchar          mask_lsb_8,
+    ulong          n_blk
+) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
+    global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
+    global uchar * ql = (global uchar *) src_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) src_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) src_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) src_d  + get_global_id(0);
+
+    b->d = *d;
+
+    for (int i = 0; i < QK_K/2/4; ++i) {
+        uchar x0   = ql[i +  0] & mask_lsb_8;
+        uchar x1   = ql[i + 32] & mask_lsb_8;
+        b->ql[i*2 + 0] = (x0 & 0x0F)        | ((x1 & 0x0F) << 4);
+        b->ql[i*2 + 1] = ((x0 & 0xF0) >> 4) | (x1 & 0xF0);
+
+        uchar x2   = ql[i + 64] & mask_lsb_8;
+        uchar x3   = ql[i + 96] & mask_lsb_8;
+        b->ql[i*2 + 0 + 64] = (x2 & 0x0F)        | ((x3 & 0x0F) << 4);
+        b->ql[i*2 + 1 + 64] = ((x2 & 0xF0) >> 4) | (x3 & 0xF0);
+    }
+
+    for (int i = 0; i < QK_K/4/8; ++i) {
+        uchar x0 = qh[i +  0] & mask_lsb_8;
+        uchar x1 = qh[i +  8] & mask_lsb_8;
+        uchar x2 = qh[i + 16] & mask_lsb_8;
+        uchar x3 = qh[i + 24] & mask_lsb_8;
+        b->qh[i*4 + 0] = (x0 & 0x03)        | ((x1 & 0x03) << 2) | ((x2 & 0x03) << 4) | ((x3 & 0x03) << 6);
+        b->qh[i*4 + 1] = ((x0 & 0x0C) >> 2) | (x1 & 0x0C)        | ((x2 & 0x0C) << 2) | ((x3 & 0x0C) << 4);
+        b->qh[i*4 + 2] = ((x0 & 0x30) >> 4) | ((x1 & 0x30) >> 2) | (x2 & 0x30)        | ((x3 & 0x30) << 2);
+        b->qh[i*4 + 3] = ((x0 & 0xC0) >> 6) | ((x1 & 0xC0) >> 4) | ((x2 & 0xC0) >> 2) | (x3 & 0xC0);
+
+        uchar x4 = qh[i +  0 + 32] & mask_lsb_8;
+        uchar x5 = qh[i +  8 + 32] & mask_lsb_8;
+        uchar x6 = qh[i + 16 + 32] & mask_lsb_8;
+        uchar x7 = qh[i + 24 + 32] & mask_lsb_8;
+        b->qh[i*4 + 0 + 32] = (x4 & 0x03)        | ((x5 & 0x03) << 2) | ((x6 & 0x03) << 4) | ((x7 & 0x03) << 6);
+        b->qh[i*4 + 1 + 32] = ((x4 & 0x0C) >> 2) | (x5 & 0x0C)        | ((x6 & 0x0C) << 2) | ((x7 & 0x0C) << 4);
+        b->qh[i*4 + 2 + 32] = ((x4 & 0x30) >> 4) | ((x5 & 0x30) >> 2) | (x6 & 0x30)        | ((x7 & 0x30) << 2);
+        b->qh[i*4 + 3 + 32] = ((x4 & 0xC0) >> 6) | ((x5 & 0xC0) >> 4) | ((x6 & 0xC0) >> 2) | (x7 & 0xC0);
+    }
+
+    for (int i = 0; i < QK_K/16; ++i) {
+        b->scales[i] = s[i];
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl

@@ -0,0 +1,140 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+kernel void kernel_gemm_noshuffle_q6_K_f32(
+        global const ushort * src0_ql,
+        global const uchar  * src0_qh,
+        global const ushort * src0_s,
+        global const half   * src0_d,
+        read_only image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int m,
+        int n,
+        int k,
+        int n_no_padding,
+        ushort mask_f000,
+        uchar  mask_c0
+) {
+    dst = (global float *)( (global char *)dst + offsetd );
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0); // n
+    int gx = get_global_id(1); // m
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    global const ushort * ptr_ql = src0_ql + gx_2;
+    global const uchar  * ptr_qh = src0_qh + gx_2;
+    global const ushort * ptr_s  = src0_s  + gx_2;
+    global const half   * ptr_d  = src0_d  + gx_2;
+
+    for (int i = 0; i < k; i += 4) {
+        // load 4x elements (ushort) of ql on M, each ushort contains 4 weights
+        // 4x ushort correspons to 4 rows on M
+        ushort4 bits4 = vload4(0, ptr_ql + (i/4)*m); // ql packed in 4s in ushort
+        uchar4  bits2 = vload4(0, ptr_qh + (i/4)*m); // qh packed in 4s in uchar
+
+        // load 4 consecutive scales
+        char8 scale_s_8 = as_char8(vload4(0, ptr_s + (i/16/2)*m)); // 1 char scale every 16 elements, packed in 2s
+        char4   scale_s = ((i/16) % 2) == 0 ? scale_s_8.s0246 : scale_s_8.s1357; // transposed as ushort, 2 blocks
+        half4   scale_d = vload4(0, ptr_d + (i/256)*m);  // 1 half scale every 256 elements
+
+        // j=0
+        // load 2x 4 elements of activations on N, corresponding to 8 rows on N
+        B.s0123 = read_imageh(src1, gy*2 + (i + 0)*n_4 + 0);
+        B.s4567 = read_imageh(src1, gy*2 + (i + 0)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half((bits4.s0 & 0x000F) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0;
+        dequantized_weights.s1 = (convert_half((bits4.s1 & 0x000F) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s1;
+        dequantized_weights.s2 = (convert_half((bits4.s2 & 0x000F) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s2;
+        dequantized_weights.s3 = (convert_half((bits4.s3 & 0x000F) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i + 1)*n_4 + 0);
+        B.s4567 = read_imageh(src1, gy*2 + (i + 1)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half((((bits4.s0 & 0x00F0) >> 4) | ((bits2.s0 & 0x0C) << 2))) - 32.f) * scale_s.s0 * scale_d.s0;
+        dequantized_weights.s1 = (convert_half((((bits4.s1 & 0x00F0) >> 4) | ((bits2.s1 & 0x0C) << 2))) - 32.f) * scale_s.s1 * scale_d.s1;
+        dequantized_weights.s2 = (convert_half((((bits4.s2 & 0x00F0) >> 4) | ((bits2.s2 & 0x0C) << 2))) - 32.f) * scale_s.s2 * scale_d.s2;
+        dequantized_weights.s3 = (convert_half((((bits4.s3 & 0x00F0) >> 4) | ((bits2.s3 & 0x0C) << 2))) - 32.f) * scale_s.s3 * scale_d.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i + 2)*n_4 + 0);
+        B.s4567 = read_imageh(src1, gy*2 + (i + 2)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half((((bits4.s0 & 0x0F00) >> 8) | (bits2.s0 & 0x30))) - 32.f) * scale_s.s0 * scale_d.s0;
+        dequantized_weights.s1 = (convert_half((((bits4.s1 & 0x0F00) >> 8) | (bits2.s1 & 0x30))) - 32.f) * scale_s.s1 * scale_d.s1;
+        dequantized_weights.s2 = (convert_half((((bits4.s2 & 0x0F00) >> 8) | (bits2.s2 & 0x30))) - 32.f) * scale_s.s2 * scale_d.s2;
+        dequantized_weights.s3 = (convert_half((((bits4.s3 & 0x0F00) >> 8) | (bits2.s3 & 0x30))) - 32.f) * scale_s.s3 * scale_d.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i + 3)*n_4 + 0);
+        B.s4567 = read_imageh(src1, gy*2 + (i + 3)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half((((bits4.s0 & mask_f000) >> 12) | ((bits2.s0 & mask_c0) >> 2))) - 32.f) * scale_s.s0 * scale_d.s0;
+        dequantized_weights.s1 = (convert_half((((bits4.s1 & mask_f000) >> 12) | ((bits2.s1 & mask_c0) >> 2))) - 32.f) * scale_s.s1 * scale_d.s1;
+        dequantized_weights.s2 = (convert_half((((bits4.s2 & mask_f000) >> 12) | ((bits2.s2 & mask_c0) >> 2))) - 32.f) * scale_s.s2 * scale_d.s2;
+        dequantized_weights.s3 = (convert_half((((bits4.s3 & mask_f000) >> 12) | ((bits2.s3 & mask_c0) >> 2))) - 32.f) * scale_s.s3 * scale_d.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl

@@ -0,0 +1,293 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+#define dequantize_block_acc_bcast_8_hi(total_sum, bits4, bits2, scale_d, scale_s, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s7; \
+
+#define dequantize_block_acc_bcast_8_lo(total_sum, bits4, bits2, scale_d, scale_s, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s7; \
+
+#define dequantize_block_acc_bcast_1_hi(total_sum, bits4, bits2, scale_d, scale_s, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
+
+#define dequantize_block_acc_bcast_1_lo(total_sum, bits4, bits2, scale_d, scale_s, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
+
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q6_K_f32(
+    read_only image1d_buffer_t src0_ql,
+    read_only image1d_buffer_t src0_qh,
+    global half2 * src0_s,
+    global half2 * src0_d,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01
+) {
+    int grp = get_local_id(1);
+    int gid = get_global_id(0);
+    ushort slid = get_sub_group_local_id();
+
+    int nb = ne00 / 32;
+
+    uint4    reg_a_l;
+    ushort4  reg_a_h;
+    half2    reg_d;
+    char4    reg_s;
+    float8   reg_b;
+
+    float2  total_sum = 0.0f;
+
+    int line_stride_a = ne01 / 2;
+    int block_stride_a = NSUBGROUPS * ne01;
+
+    for (int k = grp; k < nb; k += NSUBGROUPS) {
+        reg_d = src0_d[gid + k/8 * line_stride_a];
+        reg_s = as_char4(src0_s[gid + k * line_stride_a]);
+
+        if (slid < 4) {
+            reg_b.s0123 = read_imagef(src1, 0 + slid*2 + k*8);
+            reg_b.s4567 = read_imagef(src1, 1 + slid*2 + k*8);
+        }
+
+        reg_a_l.s0 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*0).x;
+        reg_a_l.s1 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*1).x;
+        reg_a_l.s2 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*2).x;
+        reg_a_l.s3 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*3).x;
+
+        reg_a_h.s0 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*0).x);
+        reg_a_h.s1 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*1).x);
+        reg_a_h.s2 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*2).x);
+        reg_a_h.s3 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*3).x);
+
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantize_block_acc_bcast_8_hi(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
+#else
+        dequantize_block_acc_bcast_1_hi(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+
+        reg_a_l.s0 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*4).x;
+        reg_a_l.s1 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*5).x;
+        reg_a_l.s2 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*6).x;
+        reg_a_l.s3 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*7).x;
+
+        reg_a_h.s0 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*4).x);
+        reg_a_h.s1 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*5).x);
+        reg_a_h.s2 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*6).x);
+        reg_a_h.s3 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*7).x);
+
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantize_block_acc_bcast_8_lo(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
+#else
+        dequantize_block_acc_bcast_1_lo(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+
+    local float2 reduce_lm[SUBGROUP_SIZE * 3];
+    if (grp == 1) {
+        reduce_lm[SUBGROUP_SIZE*0 + slid] = total_sum;
+    }
+    if (grp == 2) {
+        reduce_lm[SUBGROUP_SIZE*1 + slid] = total_sum;
+    }
+    if (grp == 3) {
+        reduce_lm[SUBGROUP_SIZE*2 + slid] = total_sum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (grp == 0) {
+        total_sum += reduce_lm[SUBGROUP_SIZE*0 + slid];
+    }
+    if (grp == 0) {
+        total_sum += reduce_lm[SUBGROUP_SIZE*1 + slid];
+    }
+    if (grp == 0) {
+        total_sum += reduce_lm[SUBGROUP_SIZE*2 + slid];
+    }
+
+    if (grp == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(total_sum, 0, &(dst[gid * 2]));
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl

@@ -0,0 +1,179 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_q4_k_f32_l4_lm(
+    global uchar4 * src0_q,
+    global uchar  * src0_s,
+    global half   * src0_d,
+    global half   * src0_dm,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                int ib  = idx / 64;
+                int iqs = (idx % 64) * 2;
+
+                int n = iqs / 32;
+                int b = (iqs % 32) / 16;
+                int is = 2 * n + b;
+                int qsi = n * 32 + (iqs % 16) * 2;
+
+                char * scales = src0_s + ib * 12;
+
+                int scidx0 = (is < 4) ? is : (is + 4);
+                int scidx1 = (is < 4) ? is : (is - 4);
+                int scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+                int scidxshift1 = (is < 4) ? 0 : 2;
+                int mbidx0 = is + 4;
+                int mbidx1 = (is < 4) ? is + 4 : is;
+                int mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+                int mbidxshift0 = (is < 4) ? 0 : 4;
+                int mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+                int mbidxshift1 = (is < 4) ? 0 : 2;
+
+                uchar sc    = (scales[scidx0] & 0xF) | ((scales[scidx1] & scidxmask1) >> scidxshift1);
+                uchar mbyte = ((scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((scales[mbidx1] & mbidxmask1) >> mbidxshift1);
+
+                float d =  (float)src0_d[ib]  * (float)sc;
+                float m = -(float)src0_dm[ib] * (float)mbyte;
+
+                global uchar4 * qs = src0_q + ib*32 + (qsi >> 2);
+                uchar4 q = *qs;
+                float4 v1 = (convert_float4((uchar4)((q.s0 >> (b * 4))&0x0F, (q.s1 >> (b * 4))&0x0F, (q.s2 >> (b * 4))&0x0F, (q.s3 >> (b * 4))&0x0F)))*d + m;
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v1.s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl

@@ -0,0 +1,196 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// block_q4_K
+//------------------------------------------------------------------------------
+#define QK_K            256
+#define BLOCK_Q4K_SIZE  144
+#define K_SCALE_SIZE    12
+
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+typedef struct {
+    half d;    // super-block scale for quantized scales
+    half dmin; // super-block scale for quantized mins
+
+    uchar scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uchar qs[QK_K/2];           // 4-bit quants
+} block_q4_K;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // number of rows each SIMD group works on
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 16
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// number of (super) blocks each subgroup processes
+// each thread in a subgroup processes a block (32 weights)
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_K_f32_flat(
+    global uchar * src0_q,
+    global uchar * src0_s,
+    global half  * src0_d,
+    global half  * src0_dm,
+    global char  * src1,
+    int offset1,
+    global char  * dst,
+    int offsetd,
+    int ne00,
+    int ne01,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3
+) {
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    ushort kmask1 = 0x3f3f;
+    ushort kmask2 = 0x0f0f;
+    ushort kmask3 = 0xc0c0;
+
+    int ix = get_sub_group_local_id()/8;
+    int it = get_sub_group_local_id()%8;
+    int iq = it/4;
+    int ir = it%4;
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int offset_src0 = (first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03)/BLOCK_Q4K_SIZE;
+    uint blk = nb01 / BLOCK_Q4K_SIZE;
+    global uchar * blk_q     = (global uchar *)src0_q  + offset_src0*(QK_K/2);
+    global uchar * blk_s     = (global uchar *)src0_s  + offset_src0*K_SCALE_SIZE;
+    global half  * blk_d     = (global half  *)src0_d  + offset_src0;
+    global half  * blk_dm    = (global half  *)src0_dm + offset_src0;
+
+    int offset_src1 = r1*nb11 + (i12)*nb12 + (i13)*nb13;
+    global float * y = (global float *)(src1 + offset_src1);
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    ushort  sc16[4];
+    uchar * sc8 = (uchar *)sc16;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = y4[i+0];
+            sumy.s0 += yl[i+0];
+
+            yl[i+8] = y4[i+32];
+            sumy.s1 += yl[i+8];
+
+            yh[i+0] = y4[i+128];
+            sumy.s2 += yh[i+0];
+
+            yh[i+8] = y4[i+160];
+            sumy.s3 += yh[i+8];
+        }
+
+        global ushort * q1 = (global ushort *)(blk_q + ib * (QK_K/2)) + (16 * iq + 4 * ir);
+        global ushort * sc = (global ushort *)(blk_s + ib * K_SCALE_SIZE) + iq;
+        global half   * d  = blk_d + ib;
+        global half   * dm = blk_dm + ib;
+
+        for (int row = 0; row < N_DST; row++) {
+            sc16[0] = sc[0] & kmask1;
+            sc16[1] = sc[2] & kmask1;
+            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
+            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
+
+            global ushort * q2 = q1 + 32;
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1.s0 += yl[i+0] * (q1[i/2] & 0x000F);
+                acc1.s1 += yl[i+1] * (q1[i/2] & 0x0F00);
+                acc1.s2 += yl[i+8] * (q1[i/2] & 0x00F0);
+                acc1.s3 += yl[i+9] * (q1[i/2] & 0xF000);
+                acc2.s0 += yh[i+0] * (q2[i/2] & 0x000F);
+                acc2.s1 += yh[i+1] * (q2[i/2] & 0x0F00);
+                acc2.s2 += yh[i+8] * (q2[i/2] & 0x00F0);
+                acc2.s3 += yh[i+9] * (q2[i/2] & 0xF000);
+            }
+
+            float dall = *d;
+            float dmin = *dm;
+            sumf[row] += dall * ((acc1.s0 + 1.f/256.f * acc1.s1) * sc8[0] +
+                                 (acc1.s2 + 1.f/256.f * acc1.s3) * sc8[1] * 1.f/16.f +
+                                 (acc2.s0 + 1.f/256.f * acc2.s1) * sc8[4] +
+                                 (acc2.s2 + 1.f/256.f * acc2.s3) * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy.s0 * sc8[2] + sumy.s1 * sc8[3] + sumy.s2 * sc8[6] + sumy.s3 * sc8[7]);
+
+            q1 += blk*64;
+            sc += blk*6;
+            d  += blk;
+            dm += blk;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -97,6 +97,8 @@ struct ggml_backend_openvino_buffer_context {
             ov_buffer = std::make_shared<ov::intel_gpu::ocl::USMTensor>(std::move(usm_tensor));
         } else {
             data = ggml_aligned_malloc(size);
+            GGML_ASSERT(data);
+            memset(data, 0, size);
             ov_buffer = std::make_shared<ov::Tensor>(ov::element::u8, ov::Shape{size}, data);
         }
 

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -1443,7 +1443,9 @@ ggml_tensor * rpc_server::create_node(uint64_t id,
     const rpc_tensor * tensor = it_ptr->second;
 
     struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
-    if (result == nullptr) {
+    if (result == nullptr || result->buffer == nullptr) {
+        GGML_LOG_ERROR("[%s] invalid tensor: null %s (id=%" PRIu64 ")\n",
+                       __func__, result == nullptr ? "tensor" : "buffer", id);
         return nullptr;
     }
     tensor_map[id] = result;