rpc : refactor the RPC transport (#21998)

* rpc : refactor the RPC transport

Move all transport related code into a separate file and use the
socket_t interface to hide all transport implementation details.

* fix win32

* better socket_t construction

.devops/nix/package.nix

@@ -18,6 +18,7 @@
   vulkan-loader,
   openssl,
   shaderc,
+  spirv-headers,
   useBlas ?
     builtins.all (x: !x) [
       useCuda
@@ -145,6 +146,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       ninja
       pkg-config
       git
+      spirv-headers
     ]
     ++ optionals useCuda [
       cudaPackages.cuda_nvcc

.devops/vulkan.Dockerfile

@@ -7,7 +7,7 @@ RUN apt update && apt install -y git build-essential cmake wget xz-utils
 
 # Install SSL and Vulkan SDK dependencies
 RUN apt install -y libssl-dev curl \
-    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers
 
 # Build it
 WORKDIR /app

.github/workflows/build-android.yml

@@ -51,7 +51,7 @@ jobs:
           distribution: zulu
 
       - name: Setup Android SDK
-        uses: android-actions/setup-android@9fc6c4e9069bf8d3d10b2204b1fb8f6ef7065407 # v3
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
         with:
           log-accepted-android-sdk-licenses: false
 

.github/workflows/build-riscv.yml

@@ -47,22 +47,10 @@ jobs:
     steps:
       - name: Install dependencies
         run: |
-          sudo apt-get update
-
-          # Install necessary packages
-          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 cmake build-essential wget git-lfs
-
           # Set gcc-14 and g++-14 as the default compilers
           sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
           sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
 
-          if ! which rustc; then
-            # Install Rust stable version
-            sudo apt-get install -y rustup
-            rustup install stable
-            rustup default stable
-          fi
-
           git lfs install
 
       - name: GCC version check
@@ -74,12 +62,12 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # FIXME: Enable when ggml-org/ccache-action works on riscv64
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.21
-      #   with:
-      #     key: ubuntu-riscv64-native-sanitizer-${{ matrix.sanytizer }}-${{ matrix.build_type }}
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      - name: ccache
+        uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+        with:
+          key: ubuntu-riscv64-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build

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

@@ -97,6 +97,36 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
+  # TODO: investigate slight precision issues in some operations for test-backend-ops on the WebGPU backend.
+  #ggml-ci-nvidia-webgpu:
+  #  runs-on: [self-hosted, Linux, NVIDIA]
+
+  #  steps:
+  #    - name: Clone
+  #      id: checkout
+  #      uses: actions/checkout@v6
+
+  #    - name: Dawn Dependency
+  #      id: dawn-depends
+  #      run: |
+  #        DAWN_VERSION="v20260317.182325"
+  #        DAWN_OWNER="google"
+  #        DAWN_REPO="dawn"
+  #        DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+  #        echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+  #        curl -L -o artifact.tar.gz \
+  #          "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+  #        mkdir dawn
+  #        tar -xvf artifact.tar.gz -C dawn --strip-components=1
+
+  #    - name: Test
+  #      id: ggml-ci
+  #      run: |
+  #        GG_BUILD_WEBGPU=1 \
+  #        GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+  #        GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
+  #          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
   # TODO: provision AMX-compatible machine
   #ggml-ci-cpu-amx:
   #  runs-on: [self-hosted, Linux, CPU, AMX]
@@ -141,61 +171,59 @@ jobs:
   #         amd-smi static
   #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" 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-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="v20260317.182325"
+          DAWN_OWNER="google"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          mkdir dawn
+          tar -xvf artifact.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-vulkan.yml

@@ -93,4 +93,5 @@ jobs:
           export GGML_VK_DISABLE_F16=1
           export GGML_VK_DISABLE_COOPMAT=1
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 4800
+          # test-backend-ops is too slow on llvmpipe, skip it
+          ctest -L main -E test-backend-ops --verbose --timeout 900

.github/workflows/build.yml

@@ -267,6 +267,56 @@ jobs:
           wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
           ./bin/llama-completion -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
+  android-arm64:
+    runs-on: ubuntu-latest
+
+    env:
+      NDK_VERSION: "29.0.14206865"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: android-arm64
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Set up JDK
+        uses: actions/setup-java@v5
+        with:
+          java-version: 17
+          distribution: temurin
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Install NDK
+        run: |
+          sdkmanager "ndk;${{ env.NDK_VERSION }}"
+          echo "ANDROID_NDK=${ANDROID_SDK_ROOT}/ndk/${{ env.NDK_VERSION }}" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
   ubuntu-latest-rpc:
     runs-on: ubuntu-latest
 
@@ -318,7 +368,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev libssl-dev ninja-build
+          sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev spirv-headers libssl-dev ninja-build
           echo "CC=gcc-14" >> "$GITHUB_ENV"
           echo "CXX=g++-14" >> "$GITHUB_ENV"
 
@@ -1001,22 +1051,14 @@ jobs:
     steps:
       - name: Install dependencies
         run: |
-          sudo apt-get update
-
           # Install necessary packages
-          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 cmake build-essential libssl-dev wget git-lfs
+          sudo apt-get update
+          sudo apt-get install -y libssl-dev
 
           # Set gcc-14 and g++-14 as the default compilers
           sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
           sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
 
-          if ! which rustc; then
-            # Install Rust stable version
-            sudo apt-get install -y rustup
-            rustup install stable
-            rustup default stable
-          fi
-
           git lfs install
 
       - name: Check environment
@@ -1032,13 +1074,12 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # FIXME: Enable when ggml-org/ccache-action works on riscv64
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.21
-      #   with:
-      #     key: ubuntu-cpu-riscv64-native
-      #     evict-old-files: 1d
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      - name: ccache
+        uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+        with:
+          key: ubuntu-cpu-riscv64-native
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build

.github/workflows/close-issue.yml

@@ -17,7 +17,7 @@ jobs:
     steps:
       - uses: actions/stale@v10
         with:
-          exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap"
+          exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap,security"
           days-before-issue-stale: 30
           days-before-issue-close: 14
           stale-issue-label: "stale"

.github/workflows/release.yml

@@ -202,7 +202,7 @@ jobs:
             sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libssl-dev
           else
             sudo apt-get update -y
-            sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev libssl-dev ninja-build
+            sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev spirv-headers libssl-dev ninja-build
             echo "CC=gcc-14" >> "$GITHUB_ENV"
             echo "CXX=g++-14" >> "$GITHUB_ENV"
           fi
@@ -236,6 +236,75 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
+  android-arm64:
+    runs-on: ubuntu-latest
+
+    env:
+      NDK_VERSION: "29.0.14206865"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: android-arm64
+          evict-old-files: 1d
+
+      - name: Set up JDK
+        uses: actions/setup-java@v5
+        with:
+          java-version: 17
+          distribution: temurin
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Install NDK
+        run: |
+          sdkmanager "ndk;${{ env.NDK_VERSION }}"
+          echo "ANDROID_NDK=${ANDROID_SDK_ROOT}/ndk/${{ env.NDK_VERSION }}" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz
+          name: llama-bin-android-arm64.tar.gz
+
   ubuntu-24-openvino:
     runs-on: ubuntu-24.04
 
@@ -618,6 +687,11 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
+
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
@@ -971,6 +1045,7 @@ jobs:
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
+      - android-arm64
       - macOS-cpu
       - ios-xcode-build
       - openEuler-cann
@@ -1059,6 +1134,9 @@ jobs:
             - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
 
+            **Android:**
+            - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz)
+
             **Windows:**
             - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)
             - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-arm64.zip)

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

@@ -84,41 +84,42 @@ jobs:
           export ${{ matrix.extra_args }}
           pytest -v -x -m "not slow"
 
-  server-cuda:
-    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
-
-    name: server-cuda (${{ matrix.wf_name }})
-    strategy:
-      matrix:
-        build_type: [Release]
-        wf_name: ["GPUx1"]
-        include:
-          - build_type: Release
-            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-            wf_name:    "GPUx1, backend-sampling"
-      fail-fast: false
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
-        run: |
-          cd tools/server/tests
-          python3 -m venv venv
-          source venv/bin/activate
-          pip install -r requirements.txt
-          export ${{ matrix.extra_args }}
-          pytest -v -x -m "not slow"
+  # TODO: provision CUDA runner
+  #  server-cuda:
+  #    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
+  #
+  #    name: server-cuda (${{ matrix.wf_name }})
+  #    strategy:
+  #      matrix:
+  #        build_type: [Release]
+  #        wf_name: ["GPUx1"]
+  #        include:
+  #          - build_type: Release
+  #            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+  #            wf_name:    "GPUx1, backend-sampling"
+  #      fail-fast: false
+  #
+  #    steps:
+  #      - name: Clone
+  #        id: checkout
+  #        uses: actions/checkout@v6
+  #        with:
+  #          fetch-depth: 0
+  #          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+  #
+  #      - name: Build
+  #        id: cmake_build
+  #        run: |
+  #          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
+  #          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
+  #
+  #      - name: Tests
+  #        id: server_integration_tests
+  #        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+  #        run: |
+  #          cd tools/server/tests
+  #          python3 -m venv venv
+  #          source venv/bin/activate
+  #          pip install -r requirements.txt
+  #          export ${{ matrix.extra_args }}
+  #          pytest -v -x -m "not slow"

CMakeLists.txt

@@ -225,7 +225,7 @@ foreach(FILE_PATH ${EXTRA_LICENSES})
 endforeach()
 
 if (LLAMA_BUILD_COMMON)
-    license_generate(common)
+    license_generate(llama-common)
 endif()
 
 #
@@ -249,6 +249,10 @@ set_target_properties(llama
 
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
+if (LLAMA_BUILD_COMMON)
+    install(TARGETS llama-common LIBRARY)
+endif()
+
 configure_package_config_file(
         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/llama-config.cmake.in
         ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake

CODEOWNERS

@@ -1,5 +1,21 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
-# multiplie collaborators per item can be specified
+# multiple collaborators per item can be specified
+#
+# ggml-org/ci               : CISC, danbev, ggerganov, netrunnereve, ngxson, taronaeo
+# ggml-org/ggml-cann        : hipudding
+# ggml-org/ggml-cuda        : JohannesGaessler, am17an, IMbackK, ORippler
+# ggml-org/ggml-hexagon     : lhez, max-krasnyansky
+# ggml-org/ggml-metal       : ggerganov
+# ggml-org/ggml-opencl      : lhez, max-krasnyansky
+# ggml-org/ggml-rpc         : rgerganov
+# ggml-org/ggml-sycl        : arthw
+# ggml-org/ggml-vulkan      : 0cc4m, jeffbolznv
+# ggml-org/ggml-webgpu      : reeselevine
+# ggml-org/ggml-zdnn        : taronaeo
+# ggml-org/llama-common     : ggerganov, aldehir, angt, danbev, ngxson, pwilkin
+# ggml-org/llama-mtmd       : ngxson
+# ggml-org/llama-server     : ggerganov, ngxson, allozaur, angt, ServeurpersoCom
+# ggml-org/llama-webui      : allozaur
 
 /.devops/*.Dockerfile                   @ngxson
 /.github/actions/                       @ggml-org/ci

cmake/arm64-linux-clang.cmake

@@ -0,0 +1,17 @@
+set( CMAKE_SYSTEM_NAME Linux )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target aarch64-linux-gnu )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
+
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )
+

common/CMakeLists.txt

@@ -1,9 +1,11 @@
-# common
-
 find_package(Threads REQUIRED)
 
 llama_add_compile_flags()
 
+#
+# llama-common-base
+#
+
 # Build info header
 
 if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
@@ -33,17 +35,25 @@ endif()
 
 set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
 set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+
 configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
 
-set(TARGET build_info)
-add_library(${TARGET} OBJECT ${OUTPUT_FILE})
+set(TARGET llama-common-base)
+add_library(${TARGET} STATIC ${OUTPUT_FILE})
+
+target_include_directories(${TARGET} PUBLIC .)
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-set(TARGET common)
+#
+# llama-common
+#
 
-add_library(${TARGET} STATIC
+set(TARGET llama-common)
+
+add_library(${TARGET}
     arg.cpp
     arg.h
     base64.hpp
@@ -106,17 +116,24 @@ add_library(${TARGET} STATIC
     jinja/caps.h
     )
 
+set_target_properties(${TARGET} PROPERTIES
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
+    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
+)
+
 target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+    # TODO: make fine-grained exports in the future
+    set_target_properties(${TARGET} PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON)
 endif()
 
-target_link_libraries(${TARGET} PRIVATE
-    build_info
-    cpp-httplib
-)
+target_link_libraries(${TARGET} PUBLIC  llama-common-base)
+target_link_libraries(${TARGET} PRIVATE cpp-httplib)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)

common/arg.cpp

@@ -1,5 +1,6 @@
 #include "arg.h"
 
+#include "build-info.h"
 #include "chat.h"
 #include "common.h"
 #include "download.h"
@@ -1044,8 +1045,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--version"},
         "show version and build info",
         [](common_params &) {
-            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            fprintf(stderr, "version: %d (%s)\n", llama_build_number(), llama_commit());
+            fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
             exit(0);
         }
     ));

common/build-info.cpp.in

@@ -1,4 +1,35 @@
+#include "build-info.h"
+
+#include <cstdio>
+#include <string>
+
 int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
-char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
-char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
+char const * LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
+char const * LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const * LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
+
+int llama_build_number(void) {
+    return LLAMA_BUILD_NUMBER;
+}
+
+const char * llama_commit(void) {
+    return LLAMA_COMMIT;
+}
+
+const char * llama_compiler(void) {
+    return LLAMA_COMPILER;
+}
+
+const char * llama_build_target(void) {
+    return LLAMA_BUILD_TARGET;
+}
+
+const char * llama_build_info(void) {
+    static std::string s = "b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT;
+    return s.c_str();
+}
+
+void llama_print_build_info(void) {
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, llama_build_number(), llama_commit());
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, llama_compiler(), llama_build_target());
+}

common/build-info.h

@@ -0,0 +1,11 @@
+#pragma once
+
+int llama_build_number(void);
+
+const char * llama_commit(void);
+const char * llama_compiler(void);
+
+const char * llama_build_target(void);
+const char * llama_build_info(void);
+
+void llama_print_build_info(void);

common/chat-auto-parser-generator.cpp

@@ -69,6 +69,10 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
                 auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
                 builder.resolve_refs(schema);
             });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 
@@ -194,10 +198,19 @@ common_peg_parser analyze_tools::build_tool_parser_json_native(parser_build_cont
         args_field = format.function_field + "." + args_field;
     }
 
-    auto tools_parser = p.standard_json_tools(
-        format.section_start, format.section_end, inputs.tools, inputs.parallel_tool_calls,
-        inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED, name_field, args_field, format.tools_array_wrapped,
-        format.fun_name_is_key, format.id_field, format.gen_id_field, format.parameter_order);
+    auto tools_parser = p.eps();
+    if (format.section_start.empty() && !format.per_call_start.empty()) {
+        auto single_tool_parser = p.standard_json_tools(
+            format.per_call_start, format.per_call_end, inputs.tools, inputs.parallel_tool_calls,
+            inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED, name_field, args_field, format.tools_array_wrapped,
+            format.fun_name_is_key, format.id_field, format.gen_id_field, format.parameter_order);
+        tools_parser = p.trigger_rule("tool-calls", p.one_or_more(single_tool_parser + p.space()));
+    } else {
+        tools_parser = p.standard_json_tools(
+            format.section_start, format.section_end, inputs.tools, inputs.parallel_tool_calls,
+            inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED, name_field, args_field, format.tools_array_wrapped,
+            format.fun_name_is_key, format.id_field, format.gen_id_field, format.parameter_order);
+    }
 
     // Handle content wrappers if present
     if (ctx.content && ctx.content->is_always_wrapped()) {

common/chat-auto-parser.h

@@ -308,19 +308,23 @@ struct analyze_tools : analyze_base {
 
   private:
     // Extract tool calling 'haystack' for further analysis and delegate further analysis based on format
-    void analyze_tool_calls(const analyze_reasoning & reasoning);
+    void analyze_tool_calls(const analyze_reasoning & reasoning, bool supports_parallel_tool_calls);
 
     // Analyze format based on position of function and argument name in needle
     void analyze_tool_call_format(const std::string &       haystack,
                                   const std::string &       fun_name_needle,
                                   const std::string &       arg_name_needle,
-                                  const analyze_reasoning & reasoning);
+                                  const analyze_reasoning & reasoning,
+                                  bool                      supports_parallel_tool_calls);
 
     // Analyze specifics of JSON native format (entire tool call is a JSON object)
     void analyze_tool_call_format_json_native(const std::string & clean_haystack,
                                               const std::string & fun_name_needle,
                                               const std::string & arg_name_needle);
 
+    // Check if parallel calls in JSON native format array wrapped or tag wrapped
+    void analyze_json_native_parallel_calls();
+
     // Analyze specifics of non-JSON native format (tags for function name or for function name and arguments)
     void analyze_tool_call_format_non_json(const std::string & clean_haystack,
                                            const std::string & fun_name_needle);

common/chat-diff-analyzer.cpp

@@ -558,7 +558,7 @@ analyze_tools::analyze_tools(const common_chat_template & tmpl,
     : analyze_base(tmpl) {
     LOG_DBG(ANSI_ORANGE "Phase 3: Tool call analysis\n" ANSI_RESET);
 
-    analyze_tool_calls(reasoning);
+    analyze_tool_calls(reasoning, caps.supports_parallel_tool_calls);
 
     if (format.mode != tool_format::NONE && format.mode != tool_format::JSON_NATIVE) {
         if (caps.supports_parallel_tool_calls) {
@@ -577,7 +577,7 @@ analyze_tools::analyze_tools(const common_chat_template & tmpl,
     }
 }
 
-void analyze_tools::analyze_tool_calls(const analyze_reasoning & reasoning) {
+void analyze_tools::analyze_tool_calls(const analyze_reasoning & reasoning, bool supports_parallel_tool_calls) {
     json assistant_no_tools = json{
         { "role",    "assistant"   },
         { "content", ASSISTANT_MSG }
@@ -611,13 +611,14 @@ void analyze_tools::analyze_tool_calls(const analyze_reasoning & reasoning) {
         return;
     }
 
-    analyze_tool_call_format(tool_section, FUN_FIRST, ARG_FIRST, reasoning);
+    analyze_tool_call_format(tool_section, FUN_FIRST, ARG_FIRST, reasoning, supports_parallel_tool_calls);
 }
 
 void analyze_tools::analyze_tool_call_format(const std::string &       haystack,
                                              const std::string &       fun_name_needle,
                                              const std::string &       arg_name_needle,
-                                             const analyze_reasoning & reasoning) {
+                                             const analyze_reasoning & reasoning,
+                                             bool                      supports_parallel_tool_calls) {
     if (fun_name_needle.empty() || arg_name_needle.empty() || haystack.empty()) {
         return;
     }
@@ -660,6 +661,9 @@ void analyze_tools::analyze_tool_call_format(const std::string &       haystack,
 
     if (format.mode == tool_format::JSON_NATIVE) {
         analyze_tool_call_format_json_native(clean_haystack, fun_name_needle, arg_name_needle);
+        if (supports_parallel_tool_calls) {
+            analyze_json_native_parallel_calls();
+        }
     } else {
         analyze_tool_call_format_non_json(clean_haystack, fun_name_needle);
     }
@@ -668,6 +672,42 @@ void analyze_tools::analyze_tool_call_format(const std::string &       haystack,
     format.per_call_end = trim_whitespace(format.per_call_end);
 }
 
+void analyze_tools::analyze_json_native_parallel_calls() {
+    json assistant_one_tool = json{
+        { "role",       "assistant" },
+        { "content",    ""          },
+        { "tool_calls", json::array({ first_tool_call }) }
+    };
+
+    json assistant_two_tools = json{
+        { "role",       "assistant" },
+        { "content",    ""          },
+        { "tool_calls", json::array({ first_tool_call, second_tool_call }) }
+    };
+
+    template_params params;
+    params.messages              = json::array({ user_msg, assistant_one_tool });
+    params.tools                 = tools;
+    params.add_generation_prompt = false;
+    params.enable_thinking       = true;
+
+    auto comparison = compare_variants(
+        *tmpl, params, [&](template_params & p) { p.messages = json::array({ user_msg, assistant_two_tools }); });
+
+    if (!comparison) {
+        LOG_DBG(ANSI_ORANGE "%s: Template application failed\n" ANSI_RESET, __func__);
+        return;
+    }
+
+    std::string & second_call = comparison->diff.right;
+    if (!format.section_start.empty() && second_call.find(format.section_start) != std::string::npos) {
+        format.per_call_start = format.section_start;
+        format.per_call_end = format.section_end;
+        format.section_start.clear();
+        format.section_end.clear();
+    }
+}
+
 void analyze_tools::analyze_tool_call_format_json_native(const std::string & clean_haystack,
                                                          const std::string & fun_name_needle,
                                                          const std::string & arg_name_needle) {

common/chat-peg-parser.cpp

@@ -676,7 +676,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
         ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         auto nested_name = literal("\"" + nested_name_field + "\"") + space() + literal(":") + space() +
-                          literal("\"") + tool_name(literal(name)) + literal("\"");
+                          atomic(literal("\"") + tool_name(literal(name)) + literal("\""));
         auto nested_args = literal("\"" + nested_args_field + "\"") + space() + literal(":") + space() +
                           tool_args(schema(json(), "tool-" + name + "-schema", params));
 
@@ -744,7 +744,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
         ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         auto tool_name_ = name_key_parser + space() + literal(":") + space() +
-                         literal("\"") + tool_name(literal(name)) + literal("\"");
+                         atomic(literal("\"") + tool_name(literal(name)) + literal("\""));
         auto tool_args_ = args_key_parser + space() + literal(":") + space() +
                          tool_args(schema(json(), "tool-" + name + "-schema", params));
 

common/chat.cpp

@@ -865,9 +865,10 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
         adjusted_messages.push_back(adjusted);
     }
 
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar   = true;
+    auto has_tools            = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format  = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning    = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar      = true;
 
     data.supports_thinking  = true;
     data.thinking_start_tag = "[THINK]";
@@ -887,7 +888,7 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
             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()) {
+        if (has_response_format) {
             // Ministral wants to emit json surrounded by code fences
             return generation_prompt + (reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```");
         }
@@ -928,6 +929,10 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
                 auto         schema   = function.at("parameters");
                 builder.resolve_refs(schema);
             });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 
@@ -1063,6 +1068,10 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
                 auto         schema   = function.at("parameters");
                 builder.resolve_refs(schema);
             });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 
@@ -1082,6 +1091,14 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
     common_chat_params data;
 
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+
+    if (inputs.add_generation_prompt && string_ends_with(data.prompt, "<turn|>\n")) {
+        // This may happen if the model generates content + tool_call, the
+        // template does not add the model's next turn and confuses the model
+        // from emitting its proper reasoning token sequence.
+        data.prompt += "<|turn>model\n";
+    }
+
     data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
     data.supports_thinking  = true;
     data.thinking_start_tag = "<|channel>thought";
@@ -1109,7 +1126,8 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
             p.rule("thought", p.content(p.literal("<|channel>thought") + p.space() + p.until("<channel|>") + p.literal("<channel|>")));
         }
 
-        auto thought = (p.peek(p.literal("<|channel>")) + p.ref("thought")) | p.negate(p.literal("<|channel>"));
+        auto consume_empty_channels = p.gbnf(p.zero_or_more(p.literal("<|channel>") + p.negate(p.literal("thought"))), "");
+        auto thought = (p.peek(p.literal("<|channel>")) + consume_empty_channels + p.ref("thought")) | p.negate(p.literal("<|channel>"));
 
         if (has_response_format) {
             auto response_format = p.literal("```json") <<
@@ -1173,12 +1191,16 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
                 /* max = */ inputs.parallel_tool_calls ? -1 : 1
             ));
 
-            auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<|tool_call>"})));
+            auto scan_to_toolcall = p.rule("scan-to-toolcall", p.until("<|tool_call>"));
+            auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>", "<|tool_call>"})));
             auto message = p.rule("message", thought + content);
-            return start + p.zero_or_more(message) + tool_call;
+            return start + p.zero_or_more(message) + scan_to_toolcall + tool_call;
         }
 
-        auto content = p.rule("content", p.content(p.until("<|channel>")));
+        // Gemma 4 may emit an extra <|channel>thought\n<channel|> at the end of the content. It may
+        // also emit a single trailing <channel|> token. Consume all complete reasoning blocks and
+        // then stop at the first unmatched <channel|> token.
+        auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>"})));
         auto message = p.rule("message", thought + content);
         return start + p.one_or_more(message);
     });
@@ -1193,6 +1215,10 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
                 auto         schema   = function.at("parameters");
                 builder.resolve_refs(schema);
             });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 
@@ -1643,6 +1669,173 @@ static common_chat_params common_chat_params_init_gigachat_v3(
     return data;
 }
 
+static common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_template &    tmpl,
+                                                                 const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+    data.thinking_start_tag = "<think>";
+    data.thinking_end_tag   = "</think>";
+    data.preserved_tokens  = {
+        "｜DSML｜",
+        "<think>",
+        "</think>",
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    const std::string DSML         = "｜DSML｜";
+    const std::string THINK_START  = "<think>";
+    const std::string THINK_END    = "</think>";
+    const std::string FC_START     = "<" + DSML + "function_calls>";
+    const std::string FC_END       = "</" + DSML + "function_calls>";
+    const std::string INVOKE_START = "<" + DSML + "invoke";
+    const std::string INVOKE_END   = "</" + DSML + "invoke>";
+    const std::string PARAM_START  = "<" + DSML + "parameter";
+    const std::string PARAM_END    = "</" + DSML + "parameter>";
+
+    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();
+        if (extract_reasoning && inputs.enable_thinking) {
+            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+        } else if (extract_reasoning) {
+            // Thinking disabled but reasoning extraction requested: the generation prompt
+            // contains an empty <think></think> pair that must still be consumed.
+            reasoning = p.optional(p.literal(THINK_START) + p.until(THINK_END) + p.literal(THINK_END));
+        }
+
+        if (has_response_format) {
+            auto response_format = p.rule("response-format",
+                p.literal("```json") + p.space() +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
+                p.space() + p.literal("```"));
+            return generation_prompt + reasoning + response_format + end;
+        }
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        auto tool_choice = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            auto params   = function.contains("parameters") ? function.at("parameters") : json::object();
+            const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
+
+            std::set<std::string> required;
+            if (params.contains("required")) {
+                params.at("required").get_to(required);
+            }
+
+            auto schema_info = common_schema_info();
+            schema_info.resolve_refs(params);
+
+            std::vector<common_peg_parser> required_parsers;
+            std::vector<common_peg_parser> optional_parsers;
+            for (const auto & [param_name, param_schema] : props.items()) {
+                bool is_required = required.find(param_name) != required.end();
+                bool is_string   = schema_info.resolves_to_string(param_schema);
+
+                auto arg = p.tool_arg(
+                    p.tool_arg_open(
+                        p.literal(PARAM_START + " name=\"") +
+                        p.tool_arg_name(p.literal(param_name)) +
+                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
+                    (is_string
+                         ? p.tool_arg_string_value(p.until(PARAM_END))
+                         : p.tool_arg_json_value(p.schema(p.json(),
+                                                          "tool-" + name + "-arg-" + param_name + "-schema",
+                                                          param_schema, false))) +
+                    p.tool_arg_close(p.literal(PARAM_END)));
+
+                auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
+                if (is_required) {
+                    required_parsers.push_back(named_arg);
+                } else {
+                    optional_parsers.push_back(named_arg);
+                }
+            }
+
+            common_peg_parser args_seq = p.eps();
+            for (size_t i = 0; i < required_parsers.size(); i++) {
+                if (i > 0) {
+                    args_seq = args_seq + p.space();
+                }
+                args_seq = args_seq + required_parsers[i];
+            }
+
+            if (!optional_parsers.empty()) {
+                common_peg_parser any_opt = p.choice();
+                for (const auto & opt : optional_parsers) {
+                    any_opt |= opt;
+                }
+                args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
+            }
+
+            common_peg_parser invoke_body = args_seq;
+            auto func_parser = p.tool(
+                p.tool_open(p.literal(INVOKE_START + " name=\"") +
+                            p.tool_name(p.literal(name)) + p.literal("\">\n")) +
+                invoke_body + p.space() +
+                p.tool_close(p.literal(INVOKE_END)));
+
+            tool_choice |= p.rule("tool-" + name, func_parser);
+        });
+
+        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
+        auto content_before_tools = p.content(p.until(FC_START));
+        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
+        };
+    }
+
+    return data;
+}
+
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -1914,9 +2107,23 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_gigachat_v3(tmpl, params);
     }
 
+    // DeepSeek V3.2 format detection: template defines dsml_token and uses it for tool calls.
+    // The template source contains the token as a variable assignment, not as a literal in markup.
+    if (src.find("dsml_token") != std::string::npos &&
+        src.find("function_calls") != std::string::npos &&
+        src.find("DSML") != std::string::npos) {
+        LOG_DBG("Using specialized template: DeepSeek V3.2\n");
+        return common_chat_params_init_deepseek_v3_2(tmpl, params);
+    }
+
     // Gemma4 format detection
     if (src.find("'<|tool_call>call:'") != std::string::npos) {
-        workaround::convert_tool_responses_gemma4(params.messages);
+        if (src.find("{#- OpenAI Chat Completions:") == std::string::npos) {
+            // apply workarounds if using the older gemma4 templates
+            LOG_WRN("%s: detected an outdated gemma4 chat template, applying compatibility workarounds. "
+                    "Consider updating to the official template.\n", __func__);
+            workaround::convert_tool_responses_gemma4(params.messages);
+        }
         return common_chat_params_init_gemma4(tmpl, params);
     }
 

common/common.cpp

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "gguf.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "llama.h"
@@ -372,7 +373,7 @@ void common_init() {
     const char * build_type = " (debug)";
 #endif
 
-    LOG_DBG("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
+    LOG_DBG("build: %d (%s) with %s for %s%s\n", llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 }
 
 std::string common_params_get_system_info(const common_params & params) {

common/common.h

@@ -2,9 +2,10 @@
 
 #pragma once
 
+#include "llama-cpp.h"
+
 #include "ggml-opt.h"
 #include "ggml.h"
-#include "llama-cpp.h"
 
 #include <set>
 #include <sstream>
@@ -27,11 +28,6 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
-#define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
-} while(0)
-
 struct common_time_meas {
     common_time_meas(int64_t & t_acc, bool disable = false);
     ~common_time_meas();
@@ -53,14 +49,6 @@ struct common_adapter_lora_info {
 
 using llama_tokens = std::vector<llama_token>;
 
-// build info
-extern int LLAMA_BUILD_NUMBER;
-extern const char * LLAMA_COMMIT;
-extern const char * LLAMA_COMPILER;
-extern const char * LLAMA_BUILD_TARGET;
-
-const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
-
 struct common_control_vector_load_info;
 
 //

common/download.cpp

@@ -1,5 +1,6 @@
 #include "arg.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "download.h"
@@ -258,6 +259,9 @@ static bool common_pull_file(httplib::Client & cli,
             if (progress_step >= p.total / 1000 || p.downloaded == p.total) {
                 if (callback) {
                     callback->on_update(p);
+                    if (callback->is_cancelled()) {
+                        return false;
+                    }
                 }
                 progress_step = 0;
             }
@@ -300,7 +304,7 @@ static int common_download_file_single_online(const std::string & url,
         headers.emplace(h.first, h.second);
     }
     if (headers.find("User-Agent") == headers.end()) {
-        headers.emplace("User-Agent", "llama-cpp/" + build_info);
+        headers.emplace("User-Agent", "llama-cpp/" + std::string(llama_build_info()));
     }
     if (!opts.bearer_token.empty()) {
         headers.emplace("Authorization", "Bearer " + opts.bearer_token);
@@ -373,6 +377,9 @@ static int common_download_file_single_online(const std::string & url,
     }
 
     for (int i = 0; i < max_attempts; ++i) {
+        if (opts.callback && opts.callback->is_cancelled()) {
+            break;
+        }
         if (i) {
             LOG_WRN("%s: retrying after %d seconds...\n", __func__, delay);
             std::this_thread::sleep_for(std::chrono::seconds(delay));
@@ -412,6 +419,12 @@ static int common_download_file_single_online(const std::string & url,
     if (opts.callback) {
         opts.callback->on_done(p, success);
     }
+    if (opts.callback && opts.callback->is_cancelled() &&
+        std::filesystem::exists(path_temporary)) {
+        if (remove(path_temporary.c_str()) != 0) {
+            LOG_ERR("%s: unable to delete temporary file: %s\n", __func__, path_temporary.c_str());
+        }
+    }
     if (!success) {
         LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
         return -1; // max attempts reached
@@ -429,7 +442,7 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
         headers.emplace(h.first, h.second);
     }
     if (headers.find("User-Agent") == headers.end()) {
-        headers.emplace("User-Agent", "llama-cpp/" + build_info);
+        headers.emplace("User-Agent", "llama-cpp/" + std::string(llama_build_info()));
     }
 
     if (params.timeout > 0) {

common/download.h

@@ -21,6 +21,7 @@ public:
     virtual void on_start(const common_download_progress & p) = 0;
     virtual void on_update(const common_download_progress & p) = 0;
     virtual void on_done(const common_download_progress & p, bool ok) = 0;
+    virtual bool is_cancelled() const { return false; }
 };
 
 struct common_remote_params {

common/hf-cache.cpp

@@ -1,5 +1,6 @@
 #include "hf-cache.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "http.h"
@@ -200,7 +201,7 @@ static nl::json api_get(const std::string & url,
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
-        {"User-Agent", "llama-cpp/" + build_info},
+        {"User-Agent", "llama-cpp/" + std::string(llama_build_info())},
         {"Accept", "application/json"}
     };
 

common/log.cpp

@@ -23,6 +23,10 @@
 
 int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
 
+int common_log_get_verbosity_thold(void) {
+    return common_log_verbosity_thold;
+}
+
 void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }

common/log.h

@@ -38,7 +38,7 @@ enum log_colors {
 
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
 // set via common_log_set_verbosity()
-extern int common_log_verbosity_thold;
+int  common_log_get_verbosity_thold(void);
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
@@ -98,7 +98,7 @@ void common_log_flush         (struct common_log * log);                    // f
 
 #define LOG_TMPL(level, verbosity, ...) \
     do { \
-        if ((verbosity) <= common_log_verbosity_thold) { \
+        if ((verbosity) <= common_log_get_verbosity_thold()) { \
             common_log_add(common_log_main(), (level), __VA_ARGS__); \
         } \
     } while (0)

common/peg-parser.cpp

@@ -890,6 +890,10 @@ struct parser_executor {
         }
         return result;
     }
+
+    common_peg_parse_result operator()(const common_peg_gbnf_parser & p) {
+        return arena.parse(p.child, ctx, start_pos);
+    }
 };
 
 common_peg_parse_result common_peg_arena::parse(common_peg_parse_context & ctx, size_t start) const {
@@ -957,7 +961,8 @@ void common_peg_arena::resolve_refs() {
                                  std::is_same_v<T, common_peg_and_parser> ||
                                  std::is_same_v<T, common_peg_not_parser> ||
                                  std::is_same_v<T, common_peg_tag_parser> ||
-                                 std::is_same_v<T, common_peg_atomic_parser>) {
+                                 std::is_same_v<T, common_peg_atomic_parser> ||
+                                 std::is_same_v<T, common_peg_gbnf_parser>) {
                 p.child = resolve_ref(p.child);
             } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
                 p.child = resolve_ref(p.child);
@@ -1036,6 +1041,8 @@ std::string common_peg_arena::dump_impl(common_peg_parser_id
             return "Not(" + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_atomic_parser>) {
             return "Atomic(" + dump_impl(p.child, visited) + ")";
+        } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
+            return "Gbnf(" + p.grammar + ", " + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_any_parser>) {
             return "Any";
         } else if constexpr (std::is_same_v<T, common_peg_space_parser>) {
@@ -1565,6 +1572,7 @@ static std::unordered_set<std::string> collect_reachable_rules(
                                  std::is_same_v<T, common_peg_not_parser> ||
                                  std::is_same_v<T, common_peg_tag_parser> ||
                                  std::is_same_v<T, common_peg_atomic_parser> ||
+                                 std::is_same_v<T, common_peg_gbnf_parser> ||
                                  std::is_same_v<T, common_peg_schema_parser>) {
                 visit(p.child);
             } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1651,10 +1659,13 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
             } else if constexpr (std::is_same_v<T, common_peg_sequence_parser>) {
                 std::string s;
                 for (const auto & child : p.children) {
+                    auto child_gbnf = to_gbnf(child);
+                    if (child_gbnf.empty()) {
+                        continue;
+                    }
                     if (!s.empty()) {
                         s += " ";
                     }
-                    auto child_gbnf = to_gbnf(child);
                     const auto & child_parser = effective_parser(child);
                     if (std::holds_alternative<common_peg_choice_parser>(child_parser) ||
                         std::holds_alternative<common_peg_sequence_parser>(child_parser)) {
@@ -1754,6 +1765,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                 return to_gbnf(p.child);
             } else if constexpr (std::is_same_v<T, common_peg_atomic_parser>) {
                 return to_gbnf(p.child);
+            } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
+                return p.grammar;
             } else {
                 static_assert(is_always_false_v<T>);
             }
@@ -1888,6 +1901,8 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
                 {"child", p.child},
                 {"tag", p.tag}
             };
+        } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
+            return json{{"type", "gbnf"}, {"child", p.child}, {"grammar", p.grammar}};
         }
     }, variant);
 }
@@ -2050,6 +2065,16 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         };
     }
 
+    if (type == "gbnf") {
+        if (!j.contains("child") || !j.contains("grammar")) {
+            throw std::runtime_error("gbnf parser missing required fields");
+        }
+        return common_peg_gbnf_parser{
+            j["child"].get<common_peg_parser_id>(),
+            j["grammar"].get<std::string>(),
+        };
+    }
+
     throw std::runtime_error("Unknown parser type: " + type);
 }
 

common/peg-parser.h

@@ -270,6 +270,11 @@ struct common_peg_tag_parser {
     std::string tag;
 };
 
+struct common_peg_gbnf_parser {
+    common_peg_parser_id child;
+    std::string grammar;
+};
+
 // Variant holding all parser types
 using common_peg_parser_variant = std::variant<
     common_peg_epsilon_parser,
@@ -290,7 +295,8 @@ using common_peg_parser_variant = std::variant<
     common_peg_rule_parser,
     common_peg_ref_parser,
     common_peg_atomic_parser,
-    common_peg_tag_parser
+    common_peg_tag_parser,
+    common_peg_gbnf_parser
 >;
 
 class common_peg_arena {
@@ -504,6 +510,10 @@ class common_peg_parser_builder {
     // Unlike rules, you can tag multiple nodes with the same tag.
     common_peg_parser tag(const std::string & tag, const common_peg_parser & p) { return add(common_peg_tag_parser{p.id(), tag}); }
 
+    // Wraps a child parser but emits a custom GBNF grammar string instead of
+    // the child's grammar. Parsing delegates entirely to the child.
+    common_peg_parser gbnf(const common_peg_parser & p, const std::string & grammar) { return add(common_peg_gbnf_parser{p, grammar}); }
+
     void set_root(const common_peg_parser & p);
 
     common_peg_arena build();

common/sampling.cpp

@@ -287,8 +287,8 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         }
     }
 
-    // reasoning budget sampler
-    if (!params.reasoning_budget_start.empty() && !params.reasoning_budget_end.empty()) {
+    // reasoning budget sampler (skip when budget is unlimited unless a lazy grammar is active, which needs rbudget for thinking-block suppression)
+    if (!params.reasoning_budget_start.empty() && !params.reasoning_budget_end.empty() && (params.grammar_lazy || params.reasoning_budget_tokens >= 0)) {
         rbudget = common_reasoning_budget_init(
             vocab,
             params.reasoning_budget_start,

convert_hf_to_gguf.py

@@ -4258,9 +4258,7 @@ class Qwen2VLVisionModel(MmprojModel):
                 yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Qwen2_5OmniModel")
-class Qwen25OmniModel(Qwen2VLVisionModel):
-    has_vision_encoder = True
+class Qwen25AudioModel(MmprojModel):
     has_audio_encoder = True
 
     def __init__(self, *args, **kwargs):
@@ -4276,12 +4274,6 @@ class Qwen25OmniModel(Qwen2VLVisionModel):
         self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["num_mel_bins"])
         self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-5))
 
-    def get_vision_config(self) -> dict[str, Any] | None:
-        return self.global_config["thinker_config"].get("vision_config")
-
-    def get_audio_config(self) -> dict[str, Any] | None:
-        return self.global_config["thinker_config"].get("audio_config")
-
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         # SinusoidsPositionEmbedding
         assert self.hparams_audio is not None
@@ -4312,7 +4304,32 @@ class Qwen25OmniModel(Qwen2VLVisionModel):
                 # this tensor is left unused in transformers code
                 # https://github.com/huggingface/transformers/blob/6e3063422c4b1c014aa60c32b9254fd2902f0f28/src/transformers/models/qwen2_5_omni/modular_qwen2_5_omni.py#L1809
                 return
-        yield from super().modify_tensors(data_torch, name, bid)
+            yield from MmprojModel.modify_tensors(self, data_torch, name, bid)
+
+        return  # skip other tensors
+
+
+@ModelBase.register("Qwen2_5OmniModel")
+class Qwen25OmniModel(Qwen2VLVisionModel, Qwen25AudioModel):
+    has_audio_encoder = True
+    has_vision_encoder = True
+
+    def get_vision_config(self) -> dict[str, Any] | None:
+        return self.global_config["thinker_config"].get("vision_config")
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config["thinker_config"].get("audio_config")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN25O)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "visual." in name:
+            yield from Qwen2VLVisionModel.modify_tensors(self, data_torch, name, bid)
+        elif "audio_tower." in name:
+            yield from Qwen25AudioModel.modify_tensors(self, data_torch, name, bid)
+        return  # skip other tensors
 
 
 @ModelBase.register("InternVisionModel")
@@ -4816,7 +4833,10 @@ class RND1Model(Qwen2MoeModel):
 class Qwen3VLVisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        assert self.hparams_vision is not None
+        if self.hparams_vision is None:
+            logger.info("No vision config found, skipping vision tensor processing")
+            return
+
         # Compute image_size if not present
         if "image_size" not in self.hparams_vision:
             # For Qwen3VL/Qwen3VLMoe, compute from num_position_embeddings
@@ -4837,7 +4857,9 @@ class Qwen3VLVisionModel(MmprojModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN3VL)
+        # in case mixed modalities, the arch will be handled by subclass
+        if not self.has_audio_encoder:
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN3VL)
         self.gguf_writer.add_vision_use_gelu(True)
 
         if self.hparams_vision is not None:
@@ -4925,11 +4947,64 @@ class Qwen3VLVisionModel(MmprojModel):
             return
 
         if name.startswith("visual."):
-            yield from super().modify_tensors(data_torch, name, bid)
-            return
+            yield from MmprojModel.modify_tensors(self, data_torch, name, bid)
+        return  # skip other tensors
 
-        # Fall back to parent class for other tensors
-        yield from super().modify_tensors(data_torch, name, bid)
+
+@ModelBase.register("Qwen3OmniMoeForConditionalGeneration")
+class Qwen3OmniMmprojModel(Qwen3VLVisionModel, Qwen25AudioModel):
+    has_audio_encoder = True
+    has_vision_encoder = True
+
+    def get_vision_config(self) -> dict[str, Any] | None:
+        if self.has_vision_encoder:
+            return self.global_config["thinker_config"].get("vision_config")
+        else:
+            return None
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        if self.has_audio_encoder:
+            return self.global_config["thinker_config"].get("audio_config")
+        else:
+            return None
+
+    def set_gguf_parameters(self):
+        if self.has_vision_encoder:
+            Qwen3VLVisionModel.set_gguf_parameters(self)
+            self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.QWEN3VL)
+        if self.has_audio_encoder:
+            Qwen25AudioModel.set_gguf_parameters(self)
+            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.QWEN3A)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "visual." in name:
+            if not self.has_vision_encoder:
+                raise ValueError(f"Model does not have vision encoder, but found tensor {name}")
+            # need to transform vision tensor naming, so that modify_tensors() logic can be used correctly
+            name = name.replace("thinker.visual.", "model.visual.")
+            if ".merger_list." in name:
+                name = name.replace(".merger_list.", ".deepstack_merger_list.")
+                name = name.replace(".ln_q", ".norm")
+                name = name.replace(".mlp.0", ".linear_fc1")
+                name = name.replace(".mlp.2", ".linear_fc2")
+            elif ".merger." in name:
+                name = name.replace(".ln_q", ".norm")
+                name = name.replace(".mlp.0", ".linear_fc1")
+                name = name.replace(".mlp.2", ".linear_fc2")
+            yield from Qwen3VLVisionModel.modify_tensors(self, data_torch, name, bid)
+        elif "audio_tower." in name:
+            if not self.has_audio_encoder:
+                raise ValueError(f"Model does not have audio encoder, but found tensor {name}")
+            if "conv2d" in name and name.endswith(".bias"):
+                # transform conv2d bias [n_embd] --> [1, 1, n_embd]
+                data_torch = data_torch.unsqueeze(-1).unsqueeze(-1)
+            yield from Qwen25AudioModel.modify_tensors(self, data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3ASRForConditionalGeneration")
+class Qwen3ASRMmprojModel(Qwen3OmniMmprojModel):
+    has_audio_encoder = True
+    has_vision_encoder = False
 
 
 @ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration")
@@ -4992,6 +5067,8 @@ class Step3VLVisionModel(MmprojModel):
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         if ".position_embd." in new_name:
             return gguf.GGMLQuantizationType.F32
+        if ("mm.0." in new_name or "mm.1." in new_name) and new_name.endswith(".weight"):
+            return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else 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]]:
@@ -5030,9 +5107,10 @@ class Qwen3VLTextModel(Qwen3Model):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-
-        # Handle MRoPE (Multi-axis Rotary Position Embedding) for Qwen3-VL
-        vision_config = self.hparams.get("vision_config", {})
+        if "thinker_config" in self.hparams:
+            vision_config = self.hparams["thinker_config"].get("vision_config", {})
+        else:
+            vision_config = self.hparams.get("vision_config", {})
         deepstack_layer_num = len(vision_config.get("deepstack_visual_indexes", []))
         self.gguf_writer.add_num_deepstack_layers(deepstack_layer_num)
 
@@ -5101,6 +5179,70 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Qwen3OmniMoeForConditionalGeneration")
+class Qwen3OmniMoeTextModel(Qwen3VLMoeTextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3VLMOE
+
+    def set_vocab(self):
+        super().set_vocab()
+        # correct BOS/EOS tokens
+        with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_config = json.load(f)
+            added_tokens = tokenizer_config.get("added_tokens_decoder", {})
+            for token_id, data in added_tokens.items():
+                if data.get("content") == "<|im_end|>":
+                    self.gguf_writer.add_bos_token_id(int(token_id))
+                    self.gguf_writer.add_eos_token_id(int(token_id))
+                    break
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_num_deepstack_layers(0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision and audio tensors - they go in the mmproj file
+        if "visual." in name or "audio_tower." in name \
+                or "talker." in name or "code2wav." in name:
+            return
+
+        name = name.replace("thinker.", "")
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3ASRForConditionalGeneration")
+class Qwen3ASRTextModel(Qwen3VLTextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3VL
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_num_deepstack_layers(0)
+
+    def set_vocab(self):
+        super().set_vocab()
+        # fix chat template, use correct chatml format
+        self.gguf_writer.add_chat_template("{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}")
+        # correct BOS/EOS tokens
+        with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_config = json.load(f)
+            added_tokens = tokenizer_config.get("added_tokens_decoder", {})
+            for token_id, data in added_tokens.items():
+                if data.get("content") == "<|im_end|>":
+                    self.gguf_writer.add_bos_token_id(int(token_id))
+                    self.gguf_writer.add_eos_token_id(int(token_id))
+                    break
+
+    def modify_tensors(self, data_torch, name, bid):
+        # qwen3-omni
+        name = name.replace("thinker.", "")
+
+        # Skip vision and audio tensors - they go in the mmproj file
+        if "visual." in name or "audio_tower." in name \
+                or "talker." in name or "code2wav." in name:
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 class _LinearAttentionVReorderBase(Qwen3NextModel):
     model_arch = gguf.MODEL_ARCH.QWEN3NEXT  # overridden by subclasses
     """reorders V heads from grouped to tiled order for ggml broadcast
@@ -10751,7 +10893,64 @@ class NemotronHModel(GraniteHybridModel):
                 self.gguf_writer.add_moe_latent_size(latent_size)
 
     def set_vocab(self):
-        super().set_vocab()
+        # The NemotronH config uses pattern characters (e.g. '-') that may not
+        # be supported by the installed transformers version. AutoTokenizer
+        # internally calls AutoConfig which triggers this parsing failure.
+        # Using trust_remote_code=True to load the model's own config class.
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+
+        # Pad vocab size (from Mamba2Model/GraniteHybridModel)
+        self.hparams["pad_vocab_size_multiple"] = 8 # Setting this here since GraniteHybridModel.set_vocab() isn't being invoked now.
+        # From Mamba2Model.set_vocab():
+        vocab_size = self.hparams["vocab_size"]
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+
+        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    if not added_tokens_decoder[i].normalized:
+                        previous_token = token
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
+                        if previous_token != token:
+                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
+
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        # From TextModel.set_vocab_gpt2():
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
 
         # The tokenizer _does_ add a BOS token (via post_processor type
         # TemplateProcessing) but does not set add_bos_token to true in the
@@ -11279,6 +11478,48 @@ class UltravoxWhisperEncoderModel(WhisperEncoderModel):
         self.gguf_writer.add_audio_stack_factor(self.global_config["stack_factor"])
 
 
+@ModelBase.register("MERaLiON2ForConditionalGeneration")
+class MERaLiONWhisperEncoderModel(WhisperEncoderModel):
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("speech_config")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MERALION)
+        self.gguf_writer.add_audio_stack_factor(self.global_config.get("speech_mlp_scale_factor", 15))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("text_decoder."):
+            return
+
+        if name.startswith("speech_encoder."):
+            name = name.replace("speech_encoder.", "audio_tower.")
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
+
+        suffix = "." + name.rsplit(".", 1)[-1]
+
+        if name.startswith("ln_speech."):
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MM_NORM_PRE, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("speech_audio_adapter."):
+            if ".mlp_adapter.0." in name:
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 0, suffix=suffix), data_torch)
+            elif ".gate_proj." in name:
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 1, suffix=suffix), data_torch)
+            elif ".pool_proj." in name:
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 2, suffix=suffix), data_torch)
+            elif ".out_proj." in name:
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 3, suffix=suffix), data_torch)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("VoxtralForConditionalGeneration")
 class VoxtralWhisperEncoderModel(WhisperEncoderModel):
     has_vision_encoder = False # no vision encoder

docs/backend/SYCL.md

@@ -689,6 +689,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
 | GGML_SYCL_GRAPH    | OFF *(default)* \|ON *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
+| GGML_SYCL_HOST_MEM_FALLBACK | ON *(default)* \|OFF *(Optional)* | Allow host memory fallback when device memory is full during quantized weight reorder. Enables inference to continue at reduced speed (reading over PCIe) instead of failing. Requires Linux kernel 6.8+. |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 

docs/backend/snapdragon/CMakeUserPresets.json

@@ -52,10 +52,39 @@
         }
     },
 
+    {
+        "name": "arm64-linux-snapdragon",
+        "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "cmake/arm64-linux-clang.cmake",
+            "CMAKE_C_FLAGS":   "-march=armv8 -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8 -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
+            "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
+            "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
+            "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
+            "CMAKE_PREFIX_PATH":  "$env{OPENCL_SDK_ROOT}",
+            "HEXAGON_SDK_ROOT":   "$env{HEXAGON_SDK_ROOT}",
+            "HEXAGON_TOOLS_ROOT": "$env{HEXAGON_TOOLS_ROOT}",
+            "PREBUILT_LIB_DIR": "linux_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "OFF",
+            "GGML_HEXAGON":     "ON",
+            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
+            "LLAMA_OPENSSL":    "OFF"
+        }
+    },
+
     { "name": "arm64-android-snapdragon-debug"  , "inherits": [ "base", "arm64-android-snapdragon", "debug" ] },
     { "name": "arm64-android-snapdragon-release", "inherits": [ "base", "arm64-android-snapdragon", "release" ] },
 
     { "name": "arm64-windows-snapdragon-debug"  , "inherits": [ "base", "arm64-windows-snapdragon", "debug" ] },
-    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] }
+    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] },
+
+    { "name": "arm64-linux-snapdragon-debug"  , "inherits": [ "base", "arm64-linux-snapdragon", "debug" ] },
+    { "name": "arm64-linux-snapdragon-release", "inherits": [ "base", "arm64-linux-snapdragon", "release" ] }
   ]
 }

docs/backend/snapdragon/README.md

@@ -236,10 +236,6 @@ build: 6a8cf8914 (6733)
   Controls whether the Hexagon backend allocates host buffers. By default, all buffers except for REPACK are host buffers.
   This option is required for testing Ops that require REPACK buffers (MUL_MAT and MUL_MAT_ID).
 
-- `GGML_HEXAGON_EXPERIMENTAL=1`
-  Controls whether the Hexagon backend enables experimental features.
-  This option is required for enabling/testing experimental Ops (FLASH_ATTN_EXT).
-
 - `GGML_HEXAGON_VERBOSE=1`
   Enables verbose logging of Ops from the backend. Example output:
 
@@ -259,11 +255,17 @@ build: 6a8cf8914 (6733)
   Allows enabling specific stages of the processing pipeline:
 
   - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
-  - `0x2` Enable Dynamic Quantizer (if needed for the Op)
-  - `0x4` Enable Op Compute (MUL_MAT, etc.)
+  - `0x2` Enable Op Compute (MUL_MAT, etc.)
 
   Examples:
 
       `GGML_HEXAGON_OPMASK=0x1 llama-completion ...` - Ops are enqueued but NPU-side processing is stubbed out
-      `GGML_HEXAGON_OPMASK=0x3 llama-completion ...` - NPU performs dynamic quantization and skips the rest
-      `GGML_HEXAGON_OPMASK=0x7 llama-completion ...` - Full queuing and processing of Ops (default)
+      `GGML_HEXAGON_OPMASK=0x3 llama-completion ...` - Full queuing and processing of Ops (default)
+
+- `GGML_HEXAGON_OPFILTER=regex`
+  Allows filtering (disabling) Ops that match the regex pattern:
+
+  Examples:
+
+      `GGML_HEXAGON_OPFILTER="FLASH_ATTN_EXT" llama-completion ...` - Disable Flash Attention on Hexagon (falls back to CPU or GPU)
+      `GGML_HEXAGON_OPFILTER="ADD\|SUB" llama-completion ...` - Disable ADD and SUB on Hexagon (fall back to CPU or GPU)

docs/backend/snapdragon/linux.md

@@ -0,0 +1,58 @@
+# Snapdragon-based Linux devices
+
+## Docker Setup
+
+The easiest way to build llama.cpp for a Snapdragon-based Linux device is using the toolchain Docker image (see [github.com/snapdragon-toolchain](https://github.com/snapdragon-toolchain)).
+This image includes OpenCL SDK, Hexagon SDK, CMake, and the ARM64 Linux cross-compilation toolchain.
+
+Cross-compilation is supported on **Linux X86** hosts. The resulting binaries are deployed to and run on the target **Qualcomm Snapdragon ARM64 Linux** device.
+
+```
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-linux:v0.1
+[d]/> cd /workspace
+```
+
+Note: The rest of the **Linux** build process assumes that you're running inside the toolchain container.
+
+
+## How to Build
+
+Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
+
+```
+[d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
+
+[d]/workspace> cmake --preset arm64-linux-snapdragon-release -B build-snapdragon
+
+[d]/workspace> cmake --build build-snapdragon -j $(nproc)
+```
+
+To generate an installable "package" simply use cmake --install, then zip it:
+
+```
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon
+[d]/workspace> zip -r pkg-snapdragon.zip pkg-snapdragon
+```
+
+## How to Install
+
+For this step, you will deploy the built binaries and libraries to the target Linux device. Transfer `pkg-snapdragon.zip` to the target device, then unzip it and set up the environment variables:
+
+```
+$ unzip pkg-snapdragon.zip
+$ cd pkg-snapdragon
+$ export LD_LIBRARY_PATH=./lib
+$ export ADSP_LIBRARY_PATH=./lib
+```
+
+At this point, you should also download some models onto the device:
+
+```
+$ wget https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-Q4_0.gguf
+```
+
+## How to Run
+Next, since we have setup the environment variables, we can run the llama-cli with the Hexagon backends:
+```
+$ ./bin/llama-cli -m Llama-3.2-3B-Instruct-Q4_0.gguf --device HTP0 -ngl 99 -p "what is the most popular cookie in the world?"
+```

docs/build.md

@@ -281,6 +281,12 @@ Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F` environment variable to force use FP16
 
 The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
 
+### Peer Access
+
+The environment variable `GGML_CUDA_P2P` can be set to enable peer-to-peer access between multiple GPUs, allowing them to transfer data directly rather than to go through system memory.
+Requires driver support (usually restricted to workstation/datacenter GPUs).
+May cause crashes or corrupted outputs for some motherboards and BIOS settings (e.g. IOMMU).
+
 ### Performance Tuning
 
 The following compilation options are also available to tweak performance:
@@ -456,7 +462,8 @@ pacman -S git \
     mingw-w64-ucrt-x86_64-gcc \
     mingw-w64-ucrt-x86_64-cmake \
     mingw-w64-ucrt-x86_64-vulkan-devel \
-    mingw-w64-ucrt-x86_64-shaderc
+    mingw-w64-ucrt-x86_64-shaderc \
+    mingw-w64-ucrt-x86_64-spirv-headers
 ```
 
 Switch into the `llama.cpp` directory and build using CMake.
@@ -490,9 +497,11 @@ First, follow the official LunarG instructions for the installation and setup of
 
 On Debian / Ubuntu, you can install the required dependencies using:
 ```sh
-sudo apt-get install libvulkan-dev glslc
+sudo apt-get install libvulkan-dev glslc spirv-headers
 ```
 
+SPIRV-Headers (`spirv/unified1/spirv.hpp`) are required for the Vulkan backend and are **not** always pulled in by the Vulkan loader dev package alone. Other distros use names such as `spirv-headers` (Ubuntu / Debian / Arch), or `spirv-headers-devel` (Fedora / openSUSE). On Windows, the LunarG Vulkan SDK’s `Include` directory already contains these headers.
+
 #### Common steps
 
 Second, after verifying that you have followed all of the SDK installation/setup steps, use this command to make sure before proceeding:

docs/development/HOWTO-add-model.md

@@ -5,6 +5,7 @@ Adding a model requires few steps:
 1. Convert the model to GGUF
 2. Define the model architecture in `llama.cpp`
 3. Build the GGML graph implementation
+4. Optional: Add multimodal encoder implementation
 
 After following these steps, you can open PR.
 
@@ -114,6 +115,38 @@ Some `ggml` backends do not support all operations. Backend implementations can
 
 Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).
 
+### 4. Optional: Add multimodal encoder implementation
+
+If the new model supports multimodal inputs, you will need to add a new encoder definition in `libmtmd`. You can find more information about llama.cpp's multimodal support in [the docs](../multimodal.md) and in the `tools/mtmd` source directory.
+
+1. In the conversion script, make sure you add a subclass that extends `MmprojModel` or another class that inherits from the same base class.
+2. Add the encoder definition in `clip.cpp`.
+3. Implement the preprocessor in `mtmd.cpp`. In most cases, you can reuse an existing preprocessor.
+4. Implement the encoder GGML graph, either in a dedicated file if the model is truly different from existing ones, or by reusing an existing implementation (for example: siglip, pixtral, or qwen) and adding a model-specific projector.
+
+Note:
+- Many multimodal encoders are based on models that are already supported. Make sure to read the existing encoder definitions in `tools/mtmd/models` before adding a new one. In `libmtmd`, it is generally better to extend an existing model than to duplicate code.
+- To debug the multimodal preprocessor and encoder, you can use [llama-mtmd-debug](tools/mtmd/debug/mtmd-debug.cpp).
+- Adding a model-specific API or CLI is an anti-pattern in `libmtmd`. The goal of `libmtmd` is to provide an easy-to-use, model-agnostic library for multimodal pipeline.
+- In most cases, `llama-mtmd-cli` should not be modified. If a model requires a specific prompt, either let the user provide it or bake it into the Jinja chat template.
+
+## Tips and tricks
+
+### Working with ggml_rope_ext
+
+PyTorch implementations usually prefer explicitly calculating `freq_cis`/`sin`/`cos` components. However, in llama.cpp, most RoPE operations can be handled via `ggml_rope_ext`, which does not require a sin/cos matrix. This saves memory while allowing the GGML RoPE kernel to be fused with other ops.
+
+However, since `ggml_rope_ext` only provides a subset of the RoPE implementations that models use, converting models from PyTorch to llama.cpp may require some creative adaptations.
+
+For more information about `ggml_rope_ext`, please refer to the in-code documentation in `ggml.h`.
+
+Examples:
+- `libmtmd` implements 2D RoPE with `GGML_ROPE_TYPE_NORMAL` ordering by splitting the input tensor in half, applying `ggml_rope_ext` separately to each half, then joining them back together using `ggml_concat`.
+- The [Kimi-K2.5](https://github.com/ggml-org/llama.cpp/pull/19170) vision encoder uses vision RoPE with interleaved frequencies. The weights must be permuted during conversion in order to reuse the `build_rope_2d()` function.
+- [Gemma 4](https://github.com/ggml-org/llama.cpp/pull/21309) uses "proportional" RoPE. We employ a trick where `rope_freqs` is set to a very large value in the last dimensions to prevent those dimensions from being rotated. See the `Gemma4Model` class in `convert_hf_to_gguf.py`.
+- Some models require scaling the input position. For example, `[0, 1, 2, ...]` becomes `[0, 0.5, 1, ...]`. In this case, you can provide the scaling via `freq_scale = 0.5f`.
+- Some models use learned RoPE frequencies instead of relying on `powf(freq_base, -2.0 * i / n_dims)`. In this case, you can provide the learned frequencies via the `rope_freqs` tensor (corresponding to the `c` argument in `ggml_rope_ext`), then set `freq_base = 1.0f`. An important note is that `rope_freqs` in GGML is the **inverse** (`theta = pos[i] / rope_freqs`), so you may need to invert `rope_freqs` during conversion.
+
 ## GGUF specification
 
 https://github.com/ggml-org/ggml/blob/master/docs/gguf.md

docs/multimodal.md

@@ -94,6 +94,11 @@ NOTE: some models may require large context window, for example: `-c 8192`
 # Moondream2 20250414 version
 (tool_name) -hf ggml-org/moondream2-20250414-GGUF
 
+# Gemma 4
+(tool_name) -hf ggml-org/gemma-4-E2B-it-GGUF
+(tool_name) -hf ggml-org/gemma-4-E4B-it-GGUF
+(tool_name) -hf ggml-org/gemma-4-26B-A4B-it-GGUF
+(tool_name) -hf ggml-org/gemma-4-31B-it-GGUF
 ```
 
 **Audio models**:
@@ -109,6 +114,10 @@ NOTE: some models may require large context window, for example: `-c 8192`
 
 # Mistral's Voxtral
 (tool_name) -hf ggml-org/Voxtral-Mini-3B-2507-GGUF
+
+# Qwen3-ASR
+(tool_name) -hf ggml-org/Qwen3-ASR-0.6B-GGUF
+(tool_name) -hf ggml-org/Qwen3-ASR-1.7B-GGUF
 ```
 
 **Mixed modalities**:
@@ -118,6 +127,16 @@ NOTE: some models may require large context window, for example: `-c 8192`
 # Capabilities: audio input, vision input
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
+
+# Qwen3 Omni
+# Capabilities: audio input, vision input
+(tool_name) -hf ggml-org/Qwen3-Omni-30B-A3B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen3-Omni-30B-A3B-Thinking-GGUF
+
+# Gemma 4
+# Capabilities: audio input, vision input
+(tool_name) -hf ggml-org/gemma-4-E2B-it-GGUF
+(tool_name) -hf ggml-org/gemma-4-E4B-it-GGUF
 ```
 
 ## Finding more models:

docs/ops.md

@@ -22,13 +22,13 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
@@ -46,7 +46,7 @@ Legend:
 |                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                             FILL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -84,10 +84,10 @@ Legend:
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROPE | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -116,6 +116,6 @@ Legend:
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
+|                            XIELU | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |

examples/batched/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-batched)
 add_executable(${TARGET} batched.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/convert-llama2c-to-ggml/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-convert-llama2c-to-ggml)
 add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/debug/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-debug)
 add_executable(${TARGET} debug.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/diffusion/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-diffusion-cli)
 add_executable(${TARGET} diffusion-cli.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama llama-common ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/diffusion/diffusion-cli.cpp

@@ -602,8 +602,8 @@ int main(int argc, char ** argv) {
 
     int n_input = input_tokens.size();
 
-    if (n_input >= params.n_ctx) {
-        LOG_ERR("error: input too long (%d tokens), max context is %d\n", n_input, params.n_ctx);
+    if (static_cast<uint32_t>(n_input) >= llama_n_ctx(ctx)) {
+        LOG_ERR("error: input too long (%d tokens), max context is %d\n", n_input, llama_n_ctx(ctx));
         llama_free(ctx);
         llama_model_free(model);
         return 1;

examples/embedding/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-embedding)
 add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/eval-callback/CMakeLists.txt

@@ -1,7 +1,7 @@
 set(TARGET llama-eval-callback)
 add_executable(${TARGET} eval-callback.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_BUILD_TESTS)

examples/gen-docs/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-gen-docs)
 add_executable(${TARGET} gen-docs.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/idle/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-idle)
 add_executable(${TARGET} idle.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama llama-common ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/llama.android/lib/src/main/cpp/CMakeLists.txt

@@ -51,6 +51,6 @@ target_include_directories(${CMAKE_PROJECT_NAME} PRIVATE
 
 target_link_libraries(${CMAKE_PROJECT_NAME}
         llama
-        common
+        llama-common
         android
         log)

examples/lookahead/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-lookahead)
 add_executable(${TARGET} lookahead.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/lookup/CMakeLists.txt

@@ -1,23 +1,23 @@
 set(TARGET llama-lookup)
 add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-create)
 add_executable(${TARGET} lookup-create.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-merge)
 add_executable(${TARGET} lookup-merge.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-stats)
 add_executable(${TARGET} lookup-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/parallel/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-parallel)
 add_executable(${TARGET} parallel.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/passkey/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-passkey)
 add_executable(${TARGET} passkey.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/retrieval/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-retrieval)
 add_executable(${TARGET} retrieval.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/save-load-state/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-save-load-state)
 add_executable(${TARGET} save-load-state.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/speculative-simple/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-speculative-simple)
 add_executable(${TARGET} speculative-simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/speculative/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-speculative)
 add_executable(${TARGET} speculative.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/sycl/CMakeLists.txt

@@ -5,5 +5,5 @@
 set(TARGET llama-ls-sycl-device)
 add_executable(${TARGET} ls-sycl-device.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/training/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-finetune)
 add_executable(${TARGET} finetune.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

ggml/CMakeLists.txt

@@ -1,4 +1,11 @@
 cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
+
+# ref: https://cmake.org/cmake/help/latest/policy/CMP0194.html
+# MSVC is not a valid assembler for the ASM language.
+# Set to NEW to avoid a warning on CMake 4.1+ with MSVC.
+if (POLICY CMP0194)
+    cmake_policy(SET CMP0194 NEW)
+endif()
 project("ggml" C CXX ASM)
 
 ### GGML Version
@@ -247,6 +254,7 @@ option(GGML_RPC                             "ggml: use RPC"
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 option(GGML_SYCL_GRAPH                      "ggml: enable graphs in the SYCL backend"         ON)
+option(GGML_SYCL_HOST_MEM_FALLBACK          "ggml: allow host memory fallback in SYCL reorder (requires kernel 6.8+)" ON)
 option(GGML_SYCL_DNN                        "ggml: enable oneDNN in the SYCL backend"         ON)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")

ggml/include/ggml-backend.h

@@ -202,8 +202,11 @@ extern "C" {
 
     // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // AllReduce operation for tensor parallelism (meta backend)
-    typedef bool                         (*ggml_backend_allreduce_tensor_t)(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);
+    // Context management and operations for faster communication between backends, used for tensor parallelism (meta backend)
+    typedef void * (*ggml_backend_comm_init_t)(ggml_backend_t * backends, size_t n_backends);
+    typedef void   (*ggml_backend_comm_free_t)(void * comm_ctx);
+    typedef bool   (*ggml_backend_comm_allreduce_tensor_t)(void * comm_ctx, struct ggml_tensor ** tensors);
+
     // Split buffer type for tensor parallelism (old)
     typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
     // Set the number of threads for the backend
@@ -348,6 +351,53 @@ extern "C" {
     // Set a callback to be called for each resulting node during graph compute
     GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
 
+    //
+    // Meta backend
+    //
+
+#define GGML_BACKEND_META_MAX_DEVICES 16
+
+    enum ggml_backend_meta_split_axis {
+        // tensor split by tensor dimensions:
+        GGML_BACKEND_SPLIT_AXIS_0 = 0,
+        GGML_BACKEND_SPLIT_AXIS_1 = 1,
+        GGML_BACKEND_SPLIT_AXIS_2 = 2,
+        GGML_BACKEND_SPLIT_AXIS_3 = 3,
+
+        GGML_BACKEND_SPLIT_AXIS_MIRRORED = 10, // all values on all backends
+        GGML_BACKEND_SPLIT_AXIS_PARTIAL  = 11, // each backend has a partial sum
+
+        // for internal bookkeeping only:
+        GGML_BACKEND_SPLIT_AXIS_NONE    = 98,
+        GGML_BACKEND_SPLIT_AXIS_UNKNOWN = 99,
+    };
+    GGML_API const char * ggml_backend_meta_split_axis_name(enum ggml_backend_meta_split_axis split_axis);
+
+    struct ggml_backend_meta_split_state {
+        enum ggml_backend_meta_split_axis axis;
+
+        // for tensors with axis >= 0 && axis < GGML_MAX_DIMS:
+        //   - each device has a slice of the tensor along the split axis
+        //   - most tensors have n_segments == 1 and a contiguous slice of the tensor data
+        //   - some tensors have an inhomogenenous data layout along the split axis,
+        //     those tensors are divided into segments which are each individually split across devices
+        //   - ne has one entry per segment and device that add up to ggml_tensor::ne for that axis,
+        //     the outer/inner loops are over segments/devices like [seg0_dev0, seg0_dev1, seg1_dev0, seg1_dev1],
+        //   - for example, a transformer may have a fused QKV matrix rather than 3 matrices, those would be 3 separate segments
+        //     that each need to be split individually across devices so that each device gets a slice of Q, K, and V
+        int64_t  ne[16*GGML_BACKEND_META_MAX_DEVICES];
+        uint32_t n_segments;
+    };
+
+    // function to assign split states for statically allocated tensors, compute tensor split states will be assigned to be compatible:
+    typedef struct ggml_backend_meta_split_state(*ggml_backend_meta_get_split_state_t)(const struct ggml_tensor * tensor, void * userdata);
+
+    // create a new meta device from "simple" devices, meta buffer type/buffer/backend is then derived from this:
+    // TODO: this looks a bit strange - a backend API creates a device. I think we should try
+    //       express this as a backend registry functionality instead
+    GGML_API ggml_backend_dev_t ggml_backend_meta_device(
+        ggml_backend_dev_t * devs, size_t n_devs, ggml_backend_meta_get_split_state_t get_split_state, void * get_split_state_ud);
+
     //
     // Utils
     //

ggml/include/ggml-rpc.h

@@ -6,9 +6,9 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    6
-#define RPC_PROTO_PATCH_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    4
+#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    0
 
 #ifdef  __cplusplus
 static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");

ggml/include/ggml.h

@@ -1773,8 +1773,32 @@ extern "C" {
             int                   n_dims,
             int                   mode);
 
-    // custom RoPE
+    // RoPE operations with extended options
+    // a is the input tensor to apply RoPE to, shape [n_embd, n_head, n_token]
+    // b is an int32 vector with size n_token
     // c is freq factors (e.g. phi3-128k), (optional)
+    // mode can be GGML_ROPE_TYPE_NORMAL or NEOX; for MROPE and VISION mode, use ggml_rope_multi
+    //
+    // pseudo-code for computing theta:
+    //   for i in [0, n_dims/2):
+    //     theta[i] = b[i] * powf(freq_base, -2.0 * i / n_dims);
+    //     theta[i] = theta[i] / c[i];  # if c is provided, divide theta by c
+    //     theta[i] = rope_yarn(theta[i], ...);  # note: theta = theta * freq_scale is applied here
+    //
+    // other params are used by YaRN RoPE scaling, these default values will disable YaRN:
+    //   freq_scale  = 1.0f
+    //   ext_factor  = 0.0f
+    //   attn_factor = 1.0f
+    //   beta_fast   = 0.0f
+    //   beta_slow   = 0.0f
+    //
+    // example:
+    //   (marking: c = cos, s = sin, 0 = unrotated)
+    //   given a single head with size = 8 --> [00000000]
+    //   GGML_ROPE_TYPE_NORMAL  n_dims = 4 --> [cscs0000]
+    //   GGML_ROPE_TYPE_NORMAL  n_dims = 8 --> [cscscscs]
+    //   GGML_ROPE_TYPE_NEOX    n_dims = 4 --> [ccss0000]
+    //   GGML_ROPE_TYPE_NEOX    n_dims = 8 --> [ccccssss]
     GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1790,6 +1814,36 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    // multi-dimensional RoPE, for Qwen-VL and similar vision models
+    // mode can be either VISION, MROPE, IMROPE, cannot be combined with NORMAL or NEOX
+    // sections specify how many dimensions to rotate in each section:
+    //   section length is equivalent to number of cos/sin pairs, NOT the number of dims
+    //   (i.e. sum of 4 sections are expected to be n_dims/2)
+    //   last sections can be 0, means ignored
+    // all other options are identical to ggml_rope_ext
+    //
+    // important note:
+    //   - NEOX ordering is automatically applied and cannot be disabled for MROPE and VISION
+    //     if you need normal ordering, there are 2 methods:
+    //     (1) split the tensor manually using ggml_view
+    //     (2) permute the weight upon conversion
+    //   - for VISION, n_dims must be head_size/2
+    //
+    // example M-RoPE:
+    //  given sections = [t=4, y=2, x=2, 0]
+    //  given a single head with size = 18 --> [000000000000000000]
+    //  GGML_ROPE_TYPE_MROPE   n_dims = 16 --> [ttttyyxxttttyyxx00] (cos/sin are applied in NEOX ordering)
+    //  GGML_ROPE_TYPE_IMROPE  n_dims = 16 --> [ttyxttyxttyxttyx00] (interleaved M-RoPE, still NEOX ordering)
+    //  note: the theta for each dim is computed the same way as ggml_rope_ext, no matter the section
+    //        in other words, idx used for theta: [0123456789... until n_dims/2], not reset for each section
+    //
+    // example vision RoPE:
+    //  given sections = [y=4, x=4, 0, 0] (last 2 sections are ignored)
+    //  given a single head with size = 8 --> [00000000]
+    //  GGML_ROPE_TYPE_VISION  n_dims = 4 --> [yyyyxxxx]
+    //  other values of n_dims are untested and is undefined behavior
+    //  note: unlike MROPE, the theta for each dim is computed differently for each section
+    //        in other words, idx used for theta: [0123] for y section, then [0123] for x section
     GGML_API struct ggml_tensor * ggml_rope_multi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/ggml-alloc.c

@@ -2,6 +2,7 @@
 #include "ggml-backend-impl.h"
 #include "ggml.h"
 #include "ggml-impl.h"
+
 #include <assert.h>
 #include <limits.h>
 #include <stdarg.h>

ggml/src/ggml-backend-meta.cpp

@@ -5,9 +5,6 @@
 #include "ggml-alloc.h"
 #include "ggml-cpp.h"
 
-// TODO: tmp
-#include "ggml-ext.h"
-
 #include <algorithm>
 #include <cassert>
 #include <cmath>
@@ -1273,7 +1270,45 @@ static void ggml_backend_meta_buffer_get_tensor(ggml_backend_buffer_t buffer, co
     GGML_ASSERT(ggml_is_contiguous(tensor));
 
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
-    GGML_ASSERT(split_state.n_segments == 1);
+
+    if (split_state.n_segments != 1) {
+        GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
+        GGML_ASSERT(offset == 0);
+        GGML_ASSERT(size == ggml_nbytes(tensor));
+        GGML_ASSERT(tensor->ne[3] == 1);
+        size_t offset_data = 0;
+        std::vector<size_t> simple_offsets(n_bufs, 0);
+        if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_0) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+            const int64_t blck_size = ggml_blck_size(tensor->type);
+            for (size_t s = 0; s < split_state.n_segments; s++) {
+                for (size_t j = 0; j < n_bufs; j++) {
+                    const ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                    GGML_ASSERT(split_state.ne[s*n_bufs + j] % blck_size == 0);
+                    const size_t nbytes = split_state.ne[s*n_bufs + j]/blck_size * tensor->nb[0];
+                    ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
+                        tensor->ne[1], simple_tensor->nb[1], tensor->nb[1]);
+                    offset_data       += nbytes;
+                    simple_offsets[j] += nbytes;
+                }
+            }
+            GGML_ASSERT(offset_data*tensor->ne[1] == size);
+            return;
+        }
+        GGML_ASSERT(split_state.axis == GGML_BACKEND_SPLIT_AXIS_1);
+        for (size_t s = 0; s < split_state.n_segments; s++) {
+            for (size_t j = 0; j < n_bufs; j++) {
+                const ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                const size_t nbytes = split_state.ne[s*n_bufs + j] * tensor->nb[1];
+                ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
+                    tensor->ne[2], simple_tensor->nb[2], tensor->nb[2]);
+                offset_data       += nbytes;
+                simple_offsets[j] += nbytes;
+            }
+        }
+        GGML_ASSERT(offset_data*tensor->ne[2] == size);
+        return;
+    }
 
     switch (split_state.axis) {
         case GGML_BACKEND_SPLIT_AXIS_0:
@@ -1422,22 +1457,48 @@ struct ggml_backend_meta_context {
     size_t                      max_tmp_size  = 0;
     size_t                      max_subgraphs = 0;
 
+    void *                               comm_ctx       = nullptr;
+    ggml_backend_comm_allreduce_tensor_t comm_allreduce = nullptr;
+
     ggml_backend_meta_context(ggml_backend_dev_t meta_dev, const char * params) {
         const size_t n_devs = ggml_backend_meta_dev_n_devs(meta_dev);
         name = "Meta(";
+        std::vector<ggml_backend_t> simple_backends;
         backend_configs.reserve(n_devs);
+        simple_backends.reserve(n_devs);
         for (size_t i = 0; i < n_devs; i++) {
             ggml_backend_dev_t simple_dev = ggml_backend_meta_dev_simple_dev(meta_dev, i);
             if (i > 0) {
                 name += ",";
             }
             name += ggml_backend_dev_name(simple_dev);
-            backend_configs.emplace_back(ggml_backend_dev_init(simple_dev, params));
+            simple_backends.push_back(ggml_backend_dev_init(simple_dev, params));
+            backend_configs.emplace_back(simple_backends.back());
         }
         name += ")";
+
+        if (n_devs > 1) {
+            ggml_backend_comm_init_t comm_init = (ggml_backend_comm_init_t) ggml_backend_reg_get_proc_address(
+                ggml_backend_dev_backend_reg(ggml_backend_get_device(simple_backends[0])), "ggml_backend_comm_init");
+            if (comm_init != nullptr) {
+                comm_ctx = comm_init(simple_backends.data(), simple_backends.size());
+            }
+        }
+        if (comm_ctx != nullptr) {
+            comm_allreduce = (ggml_backend_comm_allreduce_tensor_t)
+                ggml_backend_reg_get_proc_address(ggml_backend_dev_backend_reg(
+                    ggml_backend_get_device(simple_backends[0])), "ggml_backend_comm_allreduce_tensor");
+            GGML_ASSERT(comm_allreduce != nullptr);
+        }
     }
 
     ~ggml_backend_meta_context() {
+        if (comm_ctx != nullptr) {
+            ggml_backend_comm_free_t comm_free = (ggml_backend_comm_free_t) ggml_backend_reg_get_proc_address(
+                ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_configs[0].backend)), "ggml_backend_comm_free");
+            GGML_ASSERT(comm_free != nullptr);
+            comm_free(comm_ctx);
+        }
         for (auto & bc : backend_configs) {
             ggml_backend_free(bc.backend);
         }
@@ -1848,20 +1909,15 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
 
         if (n_backends > 1 && i < n_subgraphs - 1) {
             bool backend_allreduce_success = false;
-            ggml_backend_allreduce_tensor_t allreduce_tensor = (ggml_backend_allreduce_tensor_t) ggml_backend_reg_get_proc_address(
-                ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_ctx->backend_configs[0].backend)), "ggml_backend_allreduce_tensor");
-            if (allreduce_tensor) {
-                std::vector<ggml_backend_t> backends;
-                backends.reserve(n_backends);
+            if (backend_ctx->comm_ctx) {
                 std::vector<ggml_tensor *> nodes;
                 nodes.reserve(n_backends);
                 for (size_t j = 0; j < n_backends; j++) {
                     auto & bcj = backend_ctx->backend_configs[j];
-                    backends.push_back(bcj.backend);
                     ggml_cgraph * cgraph_ij = bcj.cgraphs[i].cgraph_main;
                     nodes.push_back(cgraph_ij->nodes[cgraph_ij->n_nodes-1]);
                 }
-                backend_allreduce_success = allreduce_tensor(backends.data(), nodes.data(), n_backends);
+                backend_allreduce_success = backend_ctx->comm_allreduce(backend_ctx->comm_ctx, nodes.data());
             }
 
             if (!backend_allreduce_success) {

ggml/src/ggml-backend.cpp

@@ -1030,6 +1030,8 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         GGML_ABORT("%s: failed to initialize context\n", __func__);
     }
 
+    graph->uid = ggml_graph_next_uid();
+
     // pass 1: assign backends to ops with pre-allocated inputs
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
@@ -1477,6 +1479,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         assert(graph_copy->size > graph_copy->n_leafs);
         graph_copy->leafs[graph_copy->n_leafs++] = leaf;
     }
+
+    // set ids for all splits
+    for (int i = 0; i < sched->n_splits; ++i) {
+        sched->splits[i].graph.uid = ggml_graph_next_uid();
+    }
 }
 
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -783,6 +783,7 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const int8x16_t q4_lo_1 = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits_1, m4b));
         const int8x16_t q4_hi_1 = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits_1, 4));
 
+#if defined(__ARM_FEATURE_DOTPROD)
         const int8x16_t q8_0a = vld1q_s8(y[2*ib].qs);
         const int8x16_t q8_0b = vld1q_s8(y[2*ib].qs + 16);
         const int8x16_t q8_lo_0 = vcombine_s8(vget_low_s8(q8_0a), vget_low_s8(q8_0b));
@@ -794,15 +795,40 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const int8x16_t q8_hi_1 = vcombine_s8(vget_high_s8(q8_1a), vget_high_s8(q8_1b));
 
         const int32x4_t p0 = vaddq_s32(
-            ggml_vdotq_s32(vdupq_n_s32(0), q4_lo_0, q8_lo_0),
-            ggml_vdotq_s32(vdupq_n_s32(0), q4_hi_0, q8_hi_0));
+            vdotq_s32(vdupq_n_s32(0), q4_lo_0, q8_lo_0),
+            vdotq_s32(vdupq_n_s32(0), q4_hi_0, q8_hi_0));
         const int32x4_t p1 = vaddq_s32(
-            ggml_vdotq_s32(vdupq_n_s32(0), q4_lo_1, q8_lo_1),
-            ggml_vdotq_s32(vdupq_n_s32(0), q4_hi_1, q8_hi_1));
+            vdotq_s32(vdupq_n_s32(0), q4_lo_1, q8_lo_1),
+            vdotq_s32(vdupq_n_s32(0), q4_hi_1, q8_hi_1));
 
-        const int32x4_t sums = vpaddq_s32(p0, p1);
+        const int32x4_t sumi = vpaddq_s32(p0, p1);
+#else
+        const int8x8_t q4_0_lo = vget_low_s8(q4_lo_0);
+        const int8x8_t q4_0_hi = vget_low_s8(q4_hi_0);
+        const int8x8_t q4_1_lo = vget_high_s8(q4_lo_0);
+        const int8x8_t q4_1_hi = vget_high_s8(q4_hi_0);
+        const int8x8_t q4_2_lo = vget_low_s8(q4_lo_1);
+        const int8x8_t q4_2_hi = vget_low_s8(q4_hi_1);
+        const int8x8_t q4_3_lo = vget_high_s8(q4_lo_1);
+        const int8x8_t q4_3_hi = vget_high_s8(q4_hi_1);
+
+        const int8x8_t q8_0_lo = vld1_s8(y[2*ib].qs);
+        const int8x8_t q8_0_hi = vld1_s8(y[2*ib].qs + 8);
+        const int8x8_t q8_1_lo = vld1_s8(y[2*ib].qs + 16);
+        const int8x8_t q8_1_hi = vld1_s8(y[2*ib].qs + 24);
+        const int8x8_t q8_2_lo = vld1_s8(y[2*ib+1].qs);
+        const int8x8_t q8_2_hi = vld1_s8(y[2*ib+1].qs + 8);
+        const int8x8_t q8_3_lo = vld1_s8(y[2*ib+1].qs + 16);
+        const int8x8_t q8_3_hi = vld1_s8(y[2*ib+1].qs + 24);
+
+        const int32x4_t sumi = (int32x4_t){
+            vaddvq_s32(ggml_nvfp4_dot8(q4_0_lo, q8_0_lo, q4_0_hi, q8_0_hi)),
+            vaddvq_s32(ggml_nvfp4_dot8(q4_1_lo, q8_1_lo, q4_1_hi, q8_1_hi)),
+            vaddvq_s32(ggml_nvfp4_dot8(q4_2_lo, q8_2_lo, q4_2_hi, q8_2_hi)),
+            vaddvq_s32(ggml_nvfp4_dot8(q4_3_lo, q8_3_lo, q4_3_hi, q8_3_hi)),
+        };
+#endif
 
-        // Decode 4 UE4M3 scales to f32 and multiply with q8 scales
         const float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);
         const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);
         const float32x4_t nvsc = {
@@ -813,7 +839,7 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         };
         const float32x4_t scales = vmulq_f32(nvsc, (float32x4_t){dy0, dy0, dy1, dy1});
 
-        acc = vfmaq_f32(acc, vcvtq_f32_s32(sums), scales);
+        acc = vfmaq_f32(acc, vcvtq_f32_s32(sumi), scales);
     }
     sumf = vaddvq_f32(acc);
 #else

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -306,6 +306,7 @@ inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
 
 #if !defined(__ARM_FEATURE_DOTPROD)
 
+// NOTE: this fallback produces the same total sum as native vdotq_s32 but with different per-lane grouping — do not use when individual lane values matter.
 inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
     const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
     const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
@@ -319,6 +320,15 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 
 #endif // !defined(__ARM_FEATURE_DOTPROD)
 
+static inline int32x4_t ggml_nvfp4_dot8(const int8x8_t q4_lo, const int8x8_t q8_lo,
+                                         const int8x8_t q4_hi, const int8x8_t q8_hi) {
+    const int16x8_t p_lo = vmull_s8(q4_lo, q8_lo);
+    const int16x8_t p_hi = vmull_s8(q4_hi, q8_hi);
+    const int32x4_t sum_lo = vpaddlq_s16(p_lo);
+    const int32x4_t sum_hi = vpaddlq_s16(p_hi);
+    return vaddq_s32(sum_lo, sum_hi);
+}
+
 #endif // defined(__ARM_NEON)
 
 #ifdef __wasm_simd128__

ggml/src/ggml-cpu/ops.cpp

@@ -664,6 +664,7 @@ void ggml_compute_forward_add(
             {
                 ggml_compute_forward_add_non_quantized(params, dst);
             } break;
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1113,6 +1114,7 @@ void ggml_compute_forward_add1(
                     GGML_ABORT("fatal error");
                 }
             } break;
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1242,6 +1244,7 @@ void ggml_compute_forward_acc(
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4331,6 +4334,7 @@ void ggml_compute_forward_out_prod(
     const ggml_tensor * src0 = dst->src[0];
 
     switch (src0->type) {
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4606,6 +4610,7 @@ void ggml_compute_forward_set(
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:

ggml/src/ggml-cpu/simd-gemm.h

@@ -109,6 +109,96 @@ static void simd_gemm(
         C += N;
     }
 }
+#elif defined(GGML_SIMD) && defined(__riscv_v_intrinsic)
+// RM accumulators + 1 B vector = RM + 1 <= 8  =>  RM <= 7
+// Microkernel: C[RM x vl] += A[RM x K] * B[K x N]
+template <int RM>
+static inline void rvv_simd_gemm_ukernel(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N, size_t vl)
+{
+    static_assert(RM >= 1 && RM <= 7, "RM must be 1..7 for LMUL=4");
+
+    vfloat32m4_t acc_0 = __riscv_vle32_v_f32m4(C + 0 * N, vl);
+    vfloat32m4_t acc_1, acc_2, acc_3, acc_4, acc_5, acc_6;
+    if constexpr (RM > 1) acc_1 = __riscv_vle32_v_f32m4(C + 1 * N, vl);
+    if constexpr (RM > 2) acc_2 = __riscv_vle32_v_f32m4(C + 2 * N, vl);
+    if constexpr (RM > 3) acc_3 = __riscv_vle32_v_f32m4(C + 3 * N, vl);
+    if constexpr (RM > 4) acc_4 = __riscv_vle32_v_f32m4(C + 4 * N, vl);
+    if constexpr (RM > 5) acc_5 = __riscv_vle32_v_f32m4(C + 5 * N, vl);
+    if constexpr (RM > 6) acc_6 = __riscv_vle32_v_f32m4(C + 6 * N, vl);
+
+    for (int kk = 0; kk < K; kk++) {
+        vfloat32m4_t b_0 = __riscv_vle32_v_f32m4(B + kk * N, vl);
+
+                              acc_0 = __riscv_vfmacc_vf_f32m4(acc_0, A[0 * K + kk], b_0, vl);
+        if constexpr (RM > 1) acc_1 = __riscv_vfmacc_vf_f32m4(acc_1, A[1 * K + kk], b_0, vl);
+        if constexpr (RM > 2) acc_2 = __riscv_vfmacc_vf_f32m4(acc_2, A[2 * K + kk], b_0, vl);
+        if constexpr (RM > 3) acc_3 = __riscv_vfmacc_vf_f32m4(acc_3, A[3 * K + kk], b_0, vl);
+        if constexpr (RM > 4) acc_4 = __riscv_vfmacc_vf_f32m4(acc_4, A[4 * K + kk], b_0, vl);
+        if constexpr (RM > 5) acc_5 = __riscv_vfmacc_vf_f32m4(acc_5, A[5 * K + kk], b_0, vl);
+        if constexpr (RM > 6) acc_6 = __riscv_vfmacc_vf_f32m4(acc_6, A[6 * K + kk], b_0, vl);
+    }
+
+                          __riscv_vse32_v_f32m4(C + 0 * N, acc_0, vl);
+    if constexpr (RM > 1) __riscv_vse32_v_f32m4(C + 1 * N, acc_1, vl);
+    if constexpr (RM > 2) __riscv_vse32_v_f32m4(C + 2 * N, acc_2, vl);
+    if constexpr (RM > 3) __riscv_vse32_v_f32m4(C + 3 * N, acc_3, vl);
+    if constexpr (RM > 4) __riscv_vse32_v_f32m4(C + 4 * N, acc_4, vl);
+    if constexpr (RM > 5) __riscv_vse32_v_f32m4(C + 5 * N, acc_5, vl);
+    if constexpr (RM > 6) __riscv_vse32_v_f32m4(C + 6 * N, acc_6, vl);
+}
+
+template <int RM>
+static inline void rvv_simd_gemm_dispatch_tail(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N, int KN, int remaining_rows)
+{
+    if constexpr (RM > 0) {
+        if (remaining_rows == RM) {
+            int64_t jj = 0;
+            for (; jj + KN <= N; jj += KN) {
+                rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, KN);
+            }
+            if (jj < N) {
+                rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, N - jj);
+            }
+        } else {
+            rvv_simd_gemm_dispatch_tail<RM - 1>(C, A, B, K, N, KN, remaining_rows);
+        }
+    }
+}
+
+static constexpr int GEMM_RM = 7;
+
+// C[M x N] += A[M x K] * B[K x N]
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    const int KN = (int)__riscv_vlenb();
+    int64_t ii = 0;
+    for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
+        int64_t jj = 0;
+        for (; jj + KN <= N; jj += KN) {
+            rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, KN);
+        }
+        if (jj < N) {
+            rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, N - jj);
+        }
+        A += GEMM_RM * K;
+        C += GEMM_RM * N;
+    }
+
+    int remaining_rows = M - ii;
+    rvv_simd_gemm_dispatch_tail<GEMM_RM - 1>(C, A, B, K, N, KN, remaining_rows);
+}
 
 #if defined(__GNUC__) && !defined(__clang__)
 #pragma GCC diagnostic pop

ggml/src/ggml-cuda/argsort.cu

@@ -58,26 +58,48 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 
     size_t temp_storage_bytes = 0;
 
+    bool is_capturing = false;
+#ifdef USE_CUDA_GRAPH
+    // Currently (confirmed for CCCL <= 3.2) DeviceSegmentedSort does not support stream capture, while DeviceSegmentedRadixSort does.
+    // See https://github.com/NVIDIA/cccl/issues/5661#issuecomment-3229037149
+    // TODO: constrain this to the CCCL versions that have this issue once it's resolved in a future CCCL release.
+    cudaStreamCaptureStatus capture_status;
+    CUDA_CHECK(cudaStreamIsCapturing(stream, &capture_status));
+    is_capturing = (capture_status != cudaStreamCaptureStatusNone);
+#endif  // USE_CUDA_GRAPH
+
     if (order == GGML_SORT_ORDER_ASC) {
         if (nrows == 1) {
             CUDA_CHECK(DeviceRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
-                                       temp_indices, dst,                                  // values (indices)
-                                       ncols, 0, sizeof(float) * 8, stream));
+                                                  temp_indices, dst,  // values (indices)
+                                                  ncols, 0, sizeof(float) * 8, stream));
+        } else if (is_capturing) {
+            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(
+                nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+                temp_indices, dst,                                  // values (indices)
+                ncols * nrows, nrows,                               // num items, num segments
+                offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
         } else {
-            CUDA_CHECK(DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
-                                           temp_indices, dst,                                  // values (indices)
-                                           ncols * nrows, nrows,  // num items, num segments
-                                           offset_iterator, offset_iterator + 1, stream));
+            CUDA_CHECK(DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys,
+                                                      temp_keys,             // keys (in-place)
+                                                      temp_indices, dst,     // values (indices)
+                                                      ncols * nrows, nrows,  // num items, num segments
+                                                      offset_iterator, offset_iterator + 1, stream));
         }
     } else {
         if (nrows == 1) {
-            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
-                                                 temp_indices, dst,                                  // values (indices)
-                                                 ncols, 0, sizeof(float) * 8, stream));
+            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys,
+                                                            temp_keys,          // keys (in-place)
+                                                            temp_indices, dst,  // values (indices)
+                                                            ncols, 0, sizeof(float) * 8, stream));
+        } else if (is_capturing) {
+            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairsDescending(
+                nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst, ncols * nrows, nrows,
+                offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
         } else {
-            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
-                                                     dst, ncols * nrows, nrows, offset_iterator, offset_iterator + 1,
-                                                     stream));
+            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys,
+                                                                temp_indices, dst, ncols * nrows, nrows,
+                                                                offset_iterator, offset_iterator + 1, stream));
         }
     }
 
@@ -86,22 +108,33 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 
     if (order == GGML_SORT_ORDER_ASC) {
         if (nrows == 1) {
-            CUDA_CHECK(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
-                                       temp_indices, dst,  // values (indices)
-                                       ncols, 0, sizeof(float) * 8, stream));
+            CUDA_CHECK(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys,
+                                                  temp_keys,          // keys (in-place)
+                                                  temp_indices, dst,  // values (indices)
+                                                  ncols, 0, sizeof(float) * 8, stream));
+        } else if (is_capturing) {
+            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+                                                           temp_indices, dst, ncols * nrows, nrows, offset_iterator,
+                                                           offset_iterator + 1, 0, sizeof(float) * 8, stream));
         } else {
-            CUDA_CHECK(DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
-                                           ncols * nrows, nrows, offset_iterator, offset_iterator + 1, stream));
+            CUDA_CHECK(DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+                                                      temp_indices, dst, ncols * nrows, nrows, offset_iterator,
+                                                      offset_iterator + 1, stream));
         }
     } else {
         if (nrows == 1) {
-            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
-                                                 temp_indices, dst,                                  // values (indices)
-                                                 ncols, 0, sizeof(float) * 8, stream));
+            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys,
+                                                            temp_keys,          // keys (in-place)
+                                                            temp_indices, dst,  // values (indices)
+                                                            ncols, 0, sizeof(float) * 8, stream));
+        } else if (is_capturing) {
+            CUDA_CHECK(DeviceSegmentedRadixSort::SortPairsDescending(
+                d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst, ncols * nrows, nrows,
+                offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
         } else {
-            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
-                                                     temp_indices, dst, ncols * nrows, nrows, offset_iterator,
-                                                     offset_iterator + 1, stream));
+            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys,
+                                                                temp_keys, temp_indices, dst, ncols * nrows, nrows,
+                                                                offset_iterator, offset_iterator + 1, stream));
         }
     }
 }

ggml/src/ggml-cuda/common.cuh

@@ -924,6 +924,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_F16> {
     static constexpr int qr = 1;
 };
 
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q1_0> {
+    static constexpr int qk = QK1_0;
+    static constexpr int qr = QR1_0;
+    static constexpr int qi = QI1_0;
+};
+
 template<>
 struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
     static constexpr int qk = QK4_0;
@@ -1092,10 +1099,6 @@ struct ggml_cuda_device_info {
     cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
 
     std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
-
-#ifdef GGML_USE_NCCL
-    ncclComm_t comms[GGML_CUDA_MAX_DEVICES];
-#endif // GGML_USE_NCCL
 };
 
 const ggml_cuda_device_info & ggml_cuda_info();
@@ -1183,9 +1186,13 @@ struct ggml_cuda_graph {
     std::vector<cudaGraphNode_t> nodes;
     bool disable_due_to_gpu_arch = false;
     bool warmup_complete = false;
+    uint64_t uid = 0;
+    int64_t last_used_time = 0;
     struct node_properties {
         ggml_tensor node;
-        void * node_src_data_ptrs[GGML_MAX_SRC];
+        void *   node_src_data_ptrs[GGML_MAX_SRC];
+        int64_t  node_src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
+        size_t   node_src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
     };
     std::vector<node_properties> node_props;
 
@@ -1362,12 +1369,28 @@ struct ggml_backend_cuda_context {
     // when the computation is split across CPU/GPU (e.g., with --n-cpu-moe)
     std::unordered_map<const void *, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
 
+    int64_t last_graph_eviction_sweep = 0;
+
     ggml_cuda_graph * cuda_graph(const void * first_node_ptr) {
+        const int64_t time_now = ggml_time_us();
+
+        // sweep every 5s, evicting cuda graphs unused for >=10s
+        if (time_now - last_graph_eviction_sweep >= 5'000'000) {
+            last_graph_eviction_sweep = time_now;
+            for (auto it = cuda_graphs.begin(); it != cuda_graphs.end(); ) {
+                if (time_now - it->second->last_used_time >= 10'000'000) {
+                    it = cuda_graphs.erase(it);
+                } else {
+                    ++it;
+                }
+            }
+        }
+
         auto it = cuda_graphs.find(first_node_ptr);
         if (it == cuda_graphs.end()) {
-            cuda_graphs[first_node_ptr] = std::make_unique<ggml_cuda_graph>();
-            return cuda_graphs[first_node_ptr].get();
+            it = cuda_graphs.emplace(first_node_ptr, std::make_unique<ggml_cuda_graph>()).first;
         }
+        it->second->last_used_time = time_now;
         return it->second.get();
     }
 

ggml/src/ggml-cuda/convert.cu

@@ -711,6 +711,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
 
 to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -767,6 +769,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
 
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -822,6 +826,8 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -843,6 +849,8 @@ to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -864,6 +872,8 @@ to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F16:
             return convert_unary_cuda<half, float>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:

ggml/src/ggml-cuda/dequantize.cuh

@@ -1,5 +1,27 @@
 #include "common.cuh"
 
+static __device__ __forceinline__ void dequantize_q1_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
+    const block_q1_0 * x = (const block_q1_0 *) vx;
+
+    const float d = x[ib].d;
+
+    const int bit_index_0 = iqs;
+    const int bit_index_1 = iqs + 1;
+
+    const int byte_index_0 = bit_index_0 / 8;
+    const int bit_offset_0 = bit_index_0 % 8;
+
+    const int byte_index_1 = bit_index_1 / 8;
+    const int bit_offset_1 = bit_index_1 % 8;
+
+    // Extract bits: 1 = +d, 0 = -d (branchless)
+    const int bit_0 = (x[ib].qs[byte_index_0] >> bit_offset_0) & 1;
+    const int bit_1 = (x[ib].qs[byte_index_1] >> bit_offset_1) & 1;
+
+    v.x = (2*bit_0 - 1) * d;
+    v.y = (2*bit_1 - 1) * d;
+}
+
 static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 

ggml/src/ggml-cuda/fattn.cu

@@ -75,13 +75,17 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
             return;
         }
 
-        if (use_gqa_opt && gqa_ratio % 2 == 0) {
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
-            return;
-        }
+        if constexpr (DKQ <= 256) {
+            if (use_gqa_opt && gqa_ratio % 2 == 0) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
+                return;
+            }
 
-        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
-        return;
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+            return;
+        } else {
+            GGML_ABORT("fatal error");
+        }
     }
 
     if (use_gqa_opt && gqa_ratio > 4) {
@@ -94,12 +98,16 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio > 1) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
-        return;
-    }
+    if constexpr (DKQ <= 256) {
+        if (use_gqa_opt && gqa_ratio > 1) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
+            return;
+        }
 
-    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+    } else {
+        GGML_ABORT("fatal error");
+    }
 }
 
 static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/getrows.cu

@@ -179,6 +179,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
             get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_Q1_0:
+            get_rows_cuda_q<QK1_0, QR1_0, dequantize_q1_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_Q4_0:
             get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);

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

@@ -324,28 +324,22 @@ static ggml_cuda_device_info ggml_cuda_init() {
     // configure logging to stdout
     // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
 
-    for (int id = 0; id < info.device_count; ++id) {
-        ggml_cuda_set_device(id);
-        for (int id_other = 0; id_other < info.device_count; ++id_other) {
-            if (id == id_other) {
-                continue;
-            }
-            int can_access_peer;
-            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
-            if (can_access_peer) {
-                CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+    if (getenv("GGML_CUDA_P2P") != nullptr) {
+        for (int id = 0; id < info.device_count; ++id) {
+            ggml_cuda_set_device(id);
+            for (int id_other = 0; id_other < info.device_count; ++id_other) {
+                if (id == id_other) {
+                    continue;
+                }
+                int can_access_peer;
+                CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
+                if (can_access_peer) {
+                    CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+                }
             }
         }
     }
 
-#ifdef GGML_USE_NCCL
-    int dev_ids[GGML_CUDA_MAX_DEVICES];
-    for (int id = 0; id < info.device_count; ++id) {
-        dev_ids[id] = id;
-    }
-    NCCL_CHECK(ncclCommInitAll(info.comms, info.device_count, dev_ids));
-#endif // GGML_USE_NCCL
-
     return info;
 }
 
@@ -1125,66 +1119,51 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_inte
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
-bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends) {
 #ifdef GGML_USE_NCCL
-    const int64_t ne = ggml_nelements(tensors[0]);
-    // FIXME the input of llm_graph_context::build_in_out_ids can produce a tensor with 0 elements if n_outputs == 0
-    // This then causes a crash in this function
-    if (ne == 0) {
-        return true;
-    }
-    for (size_t i = 0; i < n_backends; ++i) {
-        GGML_ASSERT(tensors[i] != nullptr);
-        GGML_ASSERT(ggml_nelements(tensors[i]) == ne);
-        GGML_ASSERT(ggml_is_contiguously_allocated(tensors[i]));
-    }
+struct ggml_backend_cuda_comm_context {
+    std::vector<ggml_backend_t> backends;
+    std::vector<ncclComm_t> comms;
 
-    const ggml_cuda_device_info info = ggml_cuda_info();
-
-    // For small tensors, simply reduce them as FP32.
-    // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
-    if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
-        NCCL_CHECK(ncclGroupStart());
-        for (size_t i = 0; i < n_backends; ++i) {
-            ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-            NCCL_CHECK(ncclAllReduce(tensors[i]->data, tensors[i]->data, ne, ncclFloat, ncclSum, info.comms[cuda_ctx->device], cuda_ctx->stream()));
+    ~ggml_backend_cuda_comm_context() {
+        for (ncclComm_t comm : comms) {
+            NCCL_CHECK(ncclCommDestroy(comm));
         }
-        NCCL_CHECK(ncclGroupEnd());
-
-        return true;
     }
+};
+#endif // GGML_USE_NCCL
 
-    // For large tensors it's faster to compress them to BF16 for the reduction:
-    to_bf16_cuda_t to_bf16 = ggml_get_to_bf16_cuda(GGML_TYPE_F32);
-    to_fp32_cuda_t to_fp32 = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
+static void ggml_backend_cuda_comm_free(void * comm_ctx_v) {
+#ifdef GGML_USE_NCCL
+    if (comm_ctx_v == nullptr) {
+        return;
+    }
+    ggml_backend_cuda_comm_context * comm_ctx = (ggml_backend_cuda_comm_context *) comm_ctx_v;
+    delete comm_ctx;
+#else
+    GGML_UNUSED(comm_ctx_v);
+#endif // GGML_USE_NCCL
+}
 
-    ggml_cuda_pool_alloc<nv_bfloat16> tmp[GGML_CUDA_MAX_DEVICES];
-    for (size_t i = 0; i < n_backends; ++i) {
+static void * ggml_backend_cuda_comm_init(ggml_backend_t * backends, size_t n_backends) {
+#ifdef GGML_USE_NCCL
+    for (size_t i = 0; i < n_backends; i++) {
+        if (!ggml_backend_is_cuda(backends[i])) {
+            return nullptr;
+        }
+    }
+    ggml_backend_cuda_comm_context * ret = new ggml_backend_cuda_comm_context;
+    std::vector<int> dev_ids;
+    ret->backends.reserve(n_backends);
+    dev_ids.reserve(n_backends);
+    for (size_t i = 0; i < n_backends; i++) {
+        ret->backends.push_back(backends[i]);
         ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-        tmp[i].pool = &cuda_ctx->pool();
-        tmp[i].alloc(ne);
-
-        ggml_cuda_set_device(i);
-        to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
-        CUDA_CHECK(cudaGetLastError());
+        dev_ids.push_back(cuda_ctx->device);
     }
 
-    NCCL_CHECK(ncclGroupStart());
-    for (size_t i = 0; i < n_backends; ++i) {
-        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-        NCCL_CHECK(ncclAllReduce(tmp[i].get(), tmp[i].get(), ne, ncclBfloat16, ncclSum, info.comms[cuda_ctx->device], cuda_ctx->stream()));
-    }
-    NCCL_CHECK(ncclGroupEnd());
-
-    for (size_t i = 0; i < n_backends; ++i) {
-        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-
-        ggml_cuda_set_device(i);
-        to_fp32(tmp[i].get(), (float *) tensors[i]->data, ne, cuda_ctx->stream());
-        CUDA_CHECK(cudaGetLastError());
-    }
-
-    return true;
+    ret->comms.resize(n_backends);
+    NCCL_CHECK(ncclCommInitAll(ret->comms.data(), n_backends, dev_ids.data()));
+    return ret;
 #else
     // If NCCL is installed it is used by default for optimal performance.
     // However, NVIDIA does not distribute NCCL with CUDA so users may be unwittingly missing this package.
@@ -1197,7 +1176,76 @@ bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_t
         warning_printed = true;
     }
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    GGML_UNUSED_VARS(backends, tensors, n_backends);
+    GGML_UNUSED_VARS(backends, n_backends);
+    return nullptr;
+#endif // GGML_USE_NCCL
+}
+
+static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct ggml_tensor ** tensors) {
+#ifdef GGML_USE_NCCL
+    const int64_t ne = ggml_nelements(tensors[0]);
+    // FIXME the input of llm_graph_context::build_in_out_ids can produce a tensor with 0 elements if n_outputs == 0
+    // This then causes a crash in this function
+    if (ne == 0) {
+        return true;
+    }
+
+    GGML_ASSERT(comm_ctx_v != nullptr);
+    ggml_backend_cuda_comm_context * comm_ctx = (ggml_backend_cuda_comm_context *) comm_ctx_v;
+    const size_t n_backends = comm_ctx->backends.size();
+
+    for (size_t i = 0; i < n_backends; ++i) {
+        GGML_ASSERT(tensors[i] != nullptr);
+        GGML_ASSERT(ggml_nelements(tensors[i]) == ne);
+        GGML_ASSERT(ggml_is_contiguously_allocated(tensors[i]));
+    }
+
+    // For small tensors, simply reduce them as FP32.
+    // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
+    if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
+        NCCL_CHECK(ncclGroupStart());
+        for (size_t i = 0; i < n_backends; ++i) {
+            ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+            NCCL_CHECK(ncclAllReduce(tensors[i]->data, tensors[i]->data, ne, ncclFloat, ncclSum, comm_ctx->comms[i], cuda_ctx->stream()));
+        }
+        NCCL_CHECK(ncclGroupEnd());
+
+        return true;
+    }
+
+    // For large tensors it's faster to compress them to BF16 for the reduction:
+    to_bf16_cuda_t to_bf16 = ggml_get_to_bf16_cuda(GGML_TYPE_F32);
+    to_fp32_cuda_t to_fp32 = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
+
+    ggml_cuda_pool_alloc<nv_bfloat16> tmp[GGML_CUDA_MAX_DEVICES];
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+        tmp[i].pool = &cuda_ctx->pool();
+        tmp[i].alloc(ne);
+
+        ggml_cuda_set_device(cuda_ctx->device);
+        to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    NCCL_CHECK(ncclGroupStart());
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+        NCCL_CHECK(ncclAllReduce(tmp[i].get(), tmp[i].get(), ne, ncclBfloat16, ncclSum, comm_ctx->comms[i], cuda_ctx->stream()));
+    }
+    NCCL_CHECK(ncclGroupEnd());
+
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+
+        ggml_cuda_set_device(cuda_ctx->device);
+        to_fp32(tmp[i].get(), (float *) tensors[i]->data, ne, cuda_ctx->stream());
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    return true;
+#else
+    GGML_UNUSED_VARS(comm_ctx_v, tensors);
     return false;
 #endif // GGML_USE_NCCL
 }
@@ -3060,6 +3108,15 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
     const void * graph_key = ggml_cuda_graph_get_key(cgraph);
     ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
+    if (cgraph->uid != 0 &&
+        cgraph->uid == graph->uid) {
+        GGML_LOG_DEBUG("CUDA Graph id %zu reused\n", cgraph->uid);
+        GGML_ASSERT((int)graph->node_props.size() == cgraph->n_nodes);
+        return false;
+    }
+
+    graph->uid = cgraph->uid;
+
     // Check if the graph size has changed
     if ((int)graph->node_props.size() != cgraph->n_nodes) {
         res = true;
@@ -3070,16 +3127,18 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
         ggml_cuda_graph::node_properties prop = {};
         memcpy(&prop.node, cgraph->nodes[i], sizeof(ggml_tensor));
 
-        // if the backend scheduler is making copies of CPU tensors, the src pointers can be the same but with different data, see:
-        // https://github.com/ggml-org/llama.cpp/pull/21472#discussion_r3052235188
         for (int j = 0; j < GGML_MAX_SRC; ++j) {
-            prop.node_src_data_ptrs[j] = cgraph->nodes[i]->src[j] ? cgraph->nodes[i]->src[j]->data : nullptr;
+            if (cgraph->nodes[i]->src[j]) {
+                prop.node_src_data_ptrs[j] = cgraph->nodes[i]->src[j]->data;
+                memcpy(prop.node_src_ne[j], cgraph->nodes[i]->src[j]->ne, sizeof(prop.node_src_ne[j]));
+                memcpy(prop.node_src_nb[j], cgraph->nodes[i]->src[j]->nb, sizeof(prop.node_src_nb[j]));
+            }
         }
 
-        if (!res && memcmp(&graph->node_props[i], &prop, sizeof(prop)) != 0) {
+        if (res || memcmp(&graph->node_props[i], &prop, sizeof(prop)) != 0) {
+            graph->node_props[i] = prop;
             res = true;
         }
-        graph->node_props[i] = prop;
     }
 
     return res;
@@ -4781,6 +4840,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 switch (a->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_Q1_0:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -4818,6 +4878,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_F32:
                     case GGML_TYPE_BF16:
                     case GGML_TYPE_I32:
+                    case GGML_TYPE_Q1_0:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -5218,8 +5279,14 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
 
 static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     GGML_UNUSED(reg);
-    if (strcmp(name, "ggml_backend_allreduce_tensor") == 0) {
-        return (void *)ggml_backend_cuda_allreduce_tensor;
+    if (strcmp(name, "ggml_backend_comm_init") == 0) {
+        return (void *)ggml_backend_cuda_comm_init;
+    }
+    if (strcmp(name, "ggml_backend_comm_free") == 0) {
+        return (void *)ggml_backend_cuda_comm_free;
+    }
+    if (strcmp(name, "ggml_backend_comm_allreduce_tensor") == 0) {
+        return (void *)ggml_backend_cuda_comm_allreduce_tensor;
     }
     if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
         return (void *)ggml_backend_cuda_split_buffer_type;

ggml/src/ggml-cuda/mmq.cu

@@ -5,6 +5,9 @@
 
 static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     switch (args.type_x) {
+        case GGML_TYPE_Q1_0:
+            mul_mat_q_case<GGML_TYPE_Q1_0>(ctx, args, stream);
+            break;
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -270,6 +273,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     bool mmq_supported;
 
     switch (type) {
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:

ggml/src/ggml-cuda/mmq.cuh

@@ -57,6 +57,8 @@ static_assert(sizeof(block_fp4_mmq)  == sizeof(block_q8_1_mmq),    "Unexpected b
 
 static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
     switch (type_x) {
+        case GGML_TYPE_Q1_0:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
             return MMQ_Q8_1_DS_LAYOUT_DS4;
@@ -185,6 +187,7 @@ static constexpr __device__ int get_mmq_y_device() {
 
 static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_Q4_0:    return MMQ_DP4A_TXS_Q4_0;
         case GGML_TYPE_Q4_1:    return MMQ_DP4A_TXS_Q4_1;
         case GGML_TYPE_Q5_0:    return MMQ_DP4A_TXS_Q8_0;
@@ -229,6 +232,7 @@ static_assert(MMQ_MMA_TILE_X_K_NVFP4 % 8 == 4, "Wrong padding.");
 
 static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_Q4_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_Q4_1:    return MMQ_MMA_TILE_X_K_Q8_1;
         case GGML_TYPE_Q5_0:    return MMQ_MMA_TILE_X_K_Q8_0;
@@ -302,6 +306,87 @@ static constexpr __device__ int mmq_get_nwarps_device() {
 
 // ------------------------------------------------------------
 
+template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_q1_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    constexpr int nwarps = mmq_get_nwarps_device();
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+
+    constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0;
+    constexpr int threads_per_row = blocks_per_iter * QI1_0;
+    constexpr int nrows = warp_size / threads_per_row;
+    constexpr int scale_entries_per_block = QK1_0 / QK8_1;
+    constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block;
+
+    const int txi  = threadIdx.x % threads_per_row;
+    const int kbx  = txi / QI1_0;
+    const int kqsx = txi % QI1_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) {
+        int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx;
+        const int qs_offset = 4*kqsx;
+        const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) |
+                        (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24);
+
+        int unpacked_bytes[8];
+#pragma unroll
+        for (int j = 0; j < 8; ++j) {
+            const int shift = j * 4;
+            const int bits4 = (qs0 >> shift) & 0x0F;
+            const int b0 = (bits4 & 0x01) ? 1 : -1;
+            const int b1 = (bits4 & 0x02) ? 1 : -1;
+            const int b2 = (bits4 & 0x04) ? 1 : -1;
+            const int b3 = (bits4 & 0x08) ? 1 : -1;
+            unpacked_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+        }
+
+        const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0;
+#pragma unroll
+        for (int j = 0; j < 8; ++j) {
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + dst_offset + j] = unpacked_bytes[j];
+#else
+            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j];
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        }
+    }
+
+    const int ksx = threadIdx.x % scale_entries_per_row;
+    const int scale_block = ksx / scale_entries_per_block;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block;
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + ksx] = bxi->d;
+#else
+        x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d;
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    }
+}
+
 template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
     const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
     constexpr int nwarps = mmq_get_nwarps_device();
@@ -3290,6 +3375,14 @@ static __device__ __forceinline__ void mmq_write_back_mma(
 template <int mmq_x, int mmq_y, bool need_check, ggml_type type>
 struct mmq_type_traits;
 
+template <int mmq_x, int mmq_y, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q1_0> {
+    static constexpr int              vdr          = VDR_Q1_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q1_0<mmq_y, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y>;
+};
+
 template <int mmq_x, int mmq_y, bool need_check>
 struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q4_0> {
     static constexpr int              vdr          = VDR_Q4_0_Q8_1_MMQ;

ggml/src/ggml-cuda/mmvq.cu

@@ -9,6 +9,7 @@ typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_
 
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return vec_dot_q1_0_q8_1;
         case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
         case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
         case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
@@ -36,6 +37,7 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
 
 static constexpr __host__ __device__ int get_vdr_mmvq(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return VDR_Q1_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
         case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
@@ -886,6 +888,12 @@ static void mul_mat_vec_q_switch_type(
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const int ids_stride, cudaStream_t stream) {
     switch (type_x) {
+        case GGML_TYPE_Q1_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q1_0>
+                (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_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,

ggml/src/ggml-cuda/ssm-conv.cu

@@ -134,8 +134,9 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
     switch (nc) {
         case 3: launch_kernel(std::integral_constant<int, 3>{}); break;
         case 4: launch_kernel(std::integral_constant<int, 4>{}); break;
+        case 5: launch_kernel(std::integral_constant<int, 5>{}); break;
         case 9: launch_kernel(std::integral_constant<int, 9>{}); break;
-        default: GGML_ABORT("Only support kernel sizes 3, 4, 9 right now.");
+        default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 9 right now.");
     }
 }
 

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

@@ -32,6 +32,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f
 SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size_kq}, {head_size_v}, {ncols1}, {ncols2});\n"
 
 TYPES_MMQ = [
+    "GGML_TYPE_Q1_0",
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
     "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K",
     "GGML_TYPE_IQ2_XXS", "GGML_TYPE_IQ2_XS", "GGML_TYPE_IQ2_S", "GGML_TYPE_IQ3_XXS", "GGML_TYPE_IQ3_S",

ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q1_0);

ggml/src/ggml-cuda/vecdotq.cuh

@@ -106,6 +106,9 @@ static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
+#define VDR_Q1_0_Q8_1_MMVQ 1  // Process one 32-element chunk at a time for parallelism
+#define VDR_Q1_0_Q8_1_MMQ  4  // Q1_0 has 128 bits (4 ints) per block
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -669,6 +672,51 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
     return d6 * sumf_d;
 }
 
+static __device__ __forceinline__ float vec_dot_q1_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q1_0 * bq1_0 = (const block_q1_0 *) vbq + kbx;
+
+    // Q1_0: 128 elements with ONE scale
+    // Q8_1: 32 elements per block with individual scales
+    // iqs selects which of the 4 chunks of 32 elements to process (0-3)
+
+    const float d1 = bq1_0->d;
+
+    // Process only the chunk specified by iqs
+    const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
+
+    // Load 32 bits (4 bytes) for this chunk from Q1_0
+    const int offset = iqs * 4;
+    const int v = bq1_0->qs[offset + 0] | (bq1_0->qs[offset + 1] << 8) |
+                  (bq1_0->qs[offset + 2] << 16) | (bq1_0->qs[offset + 3] << 24);
+
+    // Unpack 32 bits into 32 signed values (-1 or +1)
+    int vi_bytes[8];
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int shift = j * 4;
+        const int bits4 = (v >> shift) & 0x0F;
+        const int b0 = (bits4 & 0x01) ? 1 : -1;
+        const int b1 = (bits4 & 0x02) ? 1 : -1;
+        const int b2 = (bits4 & 0x04) ? 1 : -1;
+        const int b3 = (bits4 & 0x08) ? 1 : -1;
+        vi_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+    }
+
+    // Compute dot product for this 32-element chunk
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int u = get_int_b4(bq8_1_chunk->qs, j);
+        sumi = ggml_cuda_dp4a(vi_bytes[j], u, sumi);
+    }
+
+    // Apply Q1_0's single scale and this chunk's Q8_1 scale
+    const float d8 = __low2float(bq8_1_chunk->ds);
+    return d1 * d8 * sumi;
+}
+
 static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
 

ggml/src/ggml-ext.h

@@ -1,56 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-// This is a "staging" header for new ggml API
-// It is not publicly available and it should not be used by 3rd party projects
-//
-// When the API matures enough, it will be moved to the official public API
-
-//
-// Meta backend
-//
-
-#define GGML_BACKEND_META_MAX_DEVICES 16
-
-enum ggml_backend_meta_split_axis {
-    // tensor split by tensor dimensions:
-    GGML_BACKEND_SPLIT_AXIS_0   =  0,
-    GGML_BACKEND_SPLIT_AXIS_1   =  1,
-    GGML_BACKEND_SPLIT_AXIS_2   =  2,
-    GGML_BACKEND_SPLIT_AXIS_3   =  3,
-
-    GGML_BACKEND_SPLIT_AXIS_MIRRORED = 10, // all values on all backends
-    GGML_BACKEND_SPLIT_AXIS_PARTIAL  = 11, // each backend has a partial sum
-
-    // for internal bookkeeping only:
-    GGML_BACKEND_SPLIT_AXIS_NONE     = 98,
-    GGML_BACKEND_SPLIT_AXIS_UNKNOWN  = 99,
-};
-GGML_API const char * ggml_backend_meta_split_axis_name(enum ggml_backend_meta_split_axis split_axis);
-
-struct ggml_backend_meta_split_state {
-    enum ggml_backend_meta_split_axis axis;
-
-    // for tensors with axis >= 0 && axis < GGML_MAX_DIMS:
-    //   - each device has a slice of the tensor along the split axis
-    //   - most tensors have n_segments == 1 and a contiguous slice of the tensor data
-    //   - some tensors have an inhomogenenous data layout along the split axis,
-    //     those tensors are divided into segments which are each individually split across devices
-    //   - ne has one entry per segment and device that add up to ggml_tensor::ne for that axis,
-    //     the outer/inner loops are over segments/devices like [seg0_dev0, seg0_dev1, seg1_dev0, seg1_dev1],
-    //   - for example, a transformer may have a fused QKV matrix rather than 3 matrices, those would be 3 separate segments
-    //     that each need to be split individually across devices so that each device gets a slice of Q, K, and V
-    int64_t  ne[16*GGML_BACKEND_META_MAX_DEVICES];
-    uint32_t n_segments;
-};
-
-// function to assign split states for statically allocated tensors, compute tensor split states will be assigned to be compatible:
-typedef struct ggml_backend_meta_split_state(*ggml_backend_meta_get_split_state_t)(const struct ggml_tensor * tensor, void * userdata);
-
-// create a new meta device from "simple" devices, meta buffer type/buffer/backend is then derived from this:
-// TODO: this looks a bit strange - a backend API creates a device. I think we should try
-//       express this as a backend registry functionality instead
-GGML_API ggml_backend_dev_t ggml_backend_meta_device(
-    ggml_backend_dev_t * devs, size_t n_devs, ggml_backend_meta_get_split_state_t get_split_state, void * get_split_state_ud);

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -47,6 +47,7 @@ list(FIND HTP_HMX_VERSIONS ${DSP_VERSION} _hmx_idx)
 
 if (_hmx_idx GREATER_EQUAL 0)
     target_sources(${HTP_LIB} PRIVATE
+        hmx-queue.c
         hmx-matmul-ops.c
     )
 

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

@@ -14,59 +14,42 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 #include "htp-ops.h"
 
-#define htp_act_preamble3              \
-    const uint32_t ne00 = src0->ne[0]; \
-    const uint32_t ne01 = src0->ne[1]; \
-    const uint32_t ne02 = src0->ne[2]; \
-    const uint32_t ne03 = src0->ne[3]; \
-                                       \
-    const uint32_t ne10 = src1->ne[0]; \
-    const uint32_t ne11 = src1->ne[1]; \
-    const uint32_t ne12 = src1->ne[2]; \
-    const uint32_t ne13 = src1->ne[3]; \
-                                       \
-    const uint32_t ne0 = dst->ne[0];   \
-    const uint32_t ne1 = dst->ne[1];   \
-    const uint32_t ne2 = dst->ne[2];   \
-    const uint32_t ne3 = dst->ne[3];   \
-                                       \
-    const uint32_t nb00 = src0->nb[0]; \
-    const uint32_t nb01 = src0->nb[1]; \
-    const uint32_t nb02 = src0->nb[2]; \
-    const uint32_t nb03 = src0->nb[3]; \
-                                       \
-    const uint32_t nb10 = src1->nb[0]; \
-    const uint32_t nb11 = src1->nb[1]; \
-    const uint32_t nb12 = src1->nb[2]; \
-    const uint32_t nb13 = src1->nb[3]; \
-                                       \
-    const uint32_t nb0 = dst->nb[0];   \
-    const uint32_t nb1 = dst->nb[1];   \
-    const uint32_t nb2 = dst->nb[2];   \
-    const uint32_t nb3 = dst->nb[3];
-
-#define htp_act_preamble2              \
-    const uint32_t ne00 = src0->ne[0]; \
-    const uint32_t ne01 = src0->ne[1]; \
-    const uint32_t ne02 = src0->ne[2]; \
-    const uint32_t ne03 = src0->ne[3]; \
-                                       \
-    const uint32_t ne0 = dst->ne[0];   \
-    const uint32_t ne1 = dst->ne[1];   \
-    const uint32_t ne2 = dst->ne[2];   \
-    const uint32_t ne3 = dst->ne[3];   \
-                                       \
-    const uint32_t nb00 = src0->nb[0]; \
-    const uint32_t nb01 = src0->nb[1]; \
-    const uint32_t nb02 = src0->nb[2]; \
-    const uint32_t nb03 = src0->nb[3]; \
-                                       \
-    const uint32_t nb0 = dst->nb[0];   \
-    const uint32_t nb1 = dst->nb[1];   \
-    const uint32_t nb2 = dst->nb[2];   \
+#define htp_act_preamble                                 \
+    const struct htp_tensor * src0 = actx->octx->src[0]; \
+    const struct htp_tensor * src1 = actx->octx->src[1]; \
+    const struct htp_tensor * dst  = actx->octx->dst;    \
+                                                         \
+    const uint32_t ne00 = src0->ne[0];                   \
+    const uint32_t ne01 = src0->ne[1];                   \
+    const uint32_t ne02 = src0->ne[2];                   \
+    const uint32_t ne03 = src0->ne[3];                   \
+                                                         \
+    const uint32_t nb00 = src0->nb[0];                   \
+    const uint32_t nb01 = src0->nb[1];                   \
+    const uint32_t nb02 = src0->nb[2];                   \
+    const uint32_t nb03 = src0->nb[3];                   \
+                                                         \
+    const uint32_t ne10 = src1 ? src1->ne[0] : 0;        \
+    const uint32_t ne11 = src1 ? src1->ne[1] : 0;        \
+    const uint32_t ne12 = src1 ? src1->ne[2] : 0;        \
+    const uint32_t ne13 = src1 ? src1->ne[3] : 0;        \
+                                                         \
+    const uint32_t nb10 = src1 ? src1->nb[0] : 0;        \
+    const uint32_t nb11 = src1 ? src1->nb[1] : 0;        \
+    const uint32_t nb12 = src1 ? src1->nb[2] : 0;        \
+    const uint32_t nb13 = src1 ? src1->nb[3] : 0;        \
+                                                         \
+    const uint32_t ne0 = dst->ne[0];                     \
+    const uint32_t ne1 = dst->ne[1];                     \
+    const uint32_t ne2 = dst->ne[2];                     \
+    const uint32_t ne3 = dst->ne[3];                     \
+                                                         \
+    const uint32_t nb0 = dst->nb[0];                     \
+    const uint32_t nb1 = dst->nb[1];                     \
+    const uint32_t nb2 = dst->nb[2];                     \
     const uint32_t nb3 = dst->nb[3];
 
 struct htp_act_context {
@@ -97,10 +80,7 @@ struct htp_act_context {
 
 static void glu_swiglu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
     struct htp_act_context * actx = (struct htp_act_context *) data;
-    const struct htp_tensor * src0 = &actx->octx->src0;
-    const struct htp_tensor * src1 = &actx->octx->src1;
-    const struct htp_tensor * dst  = &actx->octx->dst;
-    htp_act_preamble3;
+    htp_act_preamble;
 
     size_t src0_row_size = actx->src0_row_size;
     size_t src1_row_size = actx->src1_row_size;
@@ -207,10 +187,7 @@ static void glu_swiglu_f32_per_thread(unsigned int nth, unsigned int ith, void *
 
 static void glu_swiglu_oai_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
     struct htp_act_context * actx = (struct htp_act_context *) data;
-    const struct htp_tensor * src0 = &actx->octx->src0;
-    const struct htp_tensor * src1 = &actx->octx->src1;
-    const struct htp_tensor * dst  = &actx->octx->dst;
-    htp_act_preamble3;
+    htp_act_preamble;
 
     uint64_t t1, t2;
     t1 = HAP_perf_get_qtimer_count();
@@ -332,9 +309,7 @@ static void glu_swiglu_oai_f32_per_thread(unsigned int nth, unsigned int ith, vo
 
 static void unary_gelu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
     struct htp_act_context * actx = (struct htp_act_context *) data;
-    const struct htp_tensor * src0 = &actx->octx->src0;
-    const struct htp_tensor * dst  = &actx->octx->dst;
-    htp_act_preamble2;
+    htp_act_preamble;
 
     uint64_t t1, t2;
     t1 = HAP_perf_get_qtimer_count();
@@ -433,9 +408,7 @@ static void unary_gelu_f32_per_thread(unsigned int nth, unsigned int ith, void *
 
 static void unary_silu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
     struct htp_act_context * actx = (struct htp_act_context *) data;
-    const struct htp_tensor * src0 = &actx->octx->src0;
-    const struct htp_tensor * dst  = &actx->octx->dst;
-    htp_act_preamble2;
+    htp_act_preamble;
 
     uint64_t t1, t2;
     t1 = HAP_perf_get_qtimer_count();
@@ -533,10 +506,7 @@ static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
 
 static void glu_geglu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
     struct htp_act_context * actx = (struct htp_act_context *) data;
-    const struct htp_tensor * src0 = &actx->octx->src0;
-    const struct htp_tensor * src1 = &actx->octx->src1;
-    const struct htp_tensor * dst  = &actx->octx->dst;
-    htp_act_preamble3;
+    htp_act_preamble;
 
     size_t src0_row_size = actx->src0_row_size;
     size_t src1_row_size = actx->src1_row_size;
@@ -652,9 +622,9 @@ static void glu_geglu_f32_per_thread(unsigned int nth, unsigned int ith, void *
 }
 
 static int execute_op_activations_f32(struct htp_ops_context * octx) {
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * src1 = &octx->src1;
-    struct htp_tensor *       dst  = &octx->dst;
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
 
     if (((src0->ne[0] * SIZEOF_FP32) != src0->nb[1]) || ((dst->ne[0] * SIZEOF_FP32) != dst->nb[1])) {
         FARF(ERROR, "Non-contiguous tensors are not supported at this time \n");
@@ -697,25 +667,20 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     size_t src0_row_size = src0->nb[1];
-    size_t src1_row_size = src1->nb[1]; // zero bytes if src1 is not used
+    size_t src1_row_size = src1 ? src1->nb[1] : src0->nb[1];
     size_t dst_row_size  = dst->nb[1];
 
-    const bool src1_valid = src1->ne[0];
-    if (!src1_valid) {
-        src1_row_size = src0_row_size;
-    }
-
     const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
     const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
     const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+
     // VTCM scratchpads for all tensors
     // N rows per thread, padded to HVX vector size
-
     size_t spad_size_per_row   = (src0_row_size_aligned + src1_row_size_aligned) + dst_row_size_aligned;
     size_t vtcm_row_per_thread = (octx->ctx->vtcm_size)/ (n_threads* spad_size_per_row);
 
     // Make sure the reserved vtcm size is sufficient
-    if(vtcm_row_per_thread ==0){
+    if (vtcm_row_per_thread == 0) {
         FARF(ERROR, "act-%s : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n", op_type, octx->ctx->vtcm_size,
              spad_size_per_row * n_threads);
         return HTP_STATUS_VTCM_TOO_SMALL;
@@ -733,7 +698,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
     octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
 
-    if (src1->ne[0]) {
+    octx->src0_spad.src = NULL;
+    octx->src1_spad.src = NULL;
+    octx->dst_spad.src  = NULL;
+
+    if (src1) {
         FARF(HIGH, "%s: %ux%ux%ux%u x %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
              op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
              src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
@@ -773,9 +742,9 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
 
     // Pointers and GLU logic
     const uint8_t * data_src0 = (const uint8_t *) src0->data;
-    const uint8_t * data_src1 = (const uint8_t *) src1->data;
+    const uint8_t * data_src1 = src1 ? (const uint8_t *) src1->data : NULL;
 
-    if (!src1_valid && (octx->op == HTP_OP_GLU_SWIGLU || octx->op == HTP_OP_GLU_SWIGLU_OAI || octx->op == HTP_OP_GLU_GEGLU)) {
+    if (!src1 && (octx->op == HTP_OP_GLU_SWIGLU || octx->op == HTP_OP_GLU_SWIGLU_OAI || octx->op == HTP_OP_GLU_GEGLU)) {
          const int32_t swapped = octx->op_params[1];
          data_src1 = data_src0;
          actx.src1_row_size = actx.src0_row_size;
@@ -799,7 +768,7 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
 int op_activations(struct htp_ops_context * octx) {
     int err = HTP_STATUS_OK;
 
-    switch (octx->src0.type) {
+    switch (octx->src[0]->type) {
         case HTP_TYPE_F32:
             err = execute_op_activations_f32(octx);
             break;

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

@@ -12,7 +12,7 @@
 #include "hex-dma.h"
 
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 #include "htp-ops.h"
 
 #ifndef MIN
@@ -175,8 +175,8 @@ static void htp_argsort_f32(unsigned int n, unsigned int i, void * data) {
     struct htp_ops_context * octx = actx->octx;
 
     // Unpack context
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * dst = octx->dst;
 
     // Scratchpad memory
     uint8_t * spad = octx->src0_spad.data + octx->src0_spad.size_per_thread * i;
@@ -249,16 +249,16 @@ static void htp_argsort_f32(unsigned int n, unsigned int i, void * data) {
 
 int op_argsort(struct htp_ops_context * octx) {
     // Check supported types
-    if (octx->src0.type != HTP_TYPE_F32) {
+    if (octx->src[0]->type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
+    const uint32_t total_rows = octx->src[0]->ne[1] * octx->src[0]->ne[2] * octx->src[0]->ne[3];
     const uint32_t n_threads = MIN(total_rows, octx->n_threads);
 
     // Allocate scratchpad
     // We need 1 row of float + 1 row of int32 per thread.
-    uint32_t ne00 = octx->src0.ne[0];
+    uint32_t ne00 = octx->src[0]->ne[0];
     size_t values_size  = hex_round_up(ne00 * sizeof(float), 128);
     size_t indices_size = hex_round_up(ne00 * sizeof(int32_t), 128);
     size_t spad_per_thread = values_size + indices_size;
@@ -278,9 +278,9 @@ int op_argsort(struct htp_ops_context * octx) {
     octx->src0_spad.size_per_thread = spad_per_thread;
 
     FARF(HIGH, "argsort: %ux%ux%ux%u -> %ux%ux%ux%u (0x%x, 0x%x)",
-         octx->src0.ne[0], octx->src0.ne[1], octx->src0.ne[2], octx->src0.ne[3],
-         octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
-         octx->src0.data, octx->dst.data);
+         octx->src[0]->ne[0], octx->src[0]->ne[1], octx->src[0]->ne[2], octx->src[0]->ne[3],
+         octx->dst->ne[0], octx->dst->ne[1], octx->dst->ne[2], octx->dst->ne[3],
+         octx->src[0]->data, octx->dst->data);
 
     struct htp_argsort_context actx;
     actx.octx = octx;

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

@@ -14,7 +14,7 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 #include "htp-ops.h"
 
 #ifndef MIN
@@ -43,10 +43,10 @@ struct htp_binary_context {
     bool split_at_ne02;
 };
 
-#define htp_binary_preamble                       \
-    const struct htp_tensor * src0 = &octx->src0; \
-    const struct htp_tensor * src1 = &octx->src1; \
-    struct htp_tensor *       dst  = &octx->dst;  \
+#define htp_binary_preamble                        \
+    const struct htp_tensor * src0 = octx->src[0]; \
+    const struct htp_tensor * src1 = octx->src[1]; \
+    const struct htp_tensor * dst  = octx->dst;    \
                                        \
     const uint32_t ne00 = src0->ne[0]; \
     const uint32_t ne01 = src0->ne[1]; \
@@ -181,7 +181,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type = octx->src[0]->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;
@@ -274,7 +274,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type = octx->src[0]->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;
@@ -374,7 +374,7 @@ 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->src[0]->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;
@@ -455,7 +455,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type = octx->src[0]->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;
@@ -540,7 +540,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type = octx->src[0]->type;
     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;
@@ -629,10 +629,10 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
     struct htp_binary_context * bctx = (struct htp_binary_context *) data;
     struct htp_ops_context * octx = bctx->octx;
 
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * src1 = &octx->src1;
-    const struct htp_tensor * src2 = &octx->src2;
-    struct htp_tensor *       dst  = &octx->dst;
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * src2 = octx->src[2];
+    const struct htp_tensor * dst  = octx->dst;
 
     const uint32_t ne00 = src0->ne[0];
     const uint32_t ne01 = src0->ne[1];
@@ -723,15 +723,15 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
 }
 
 static int execute_op_binary(struct htp_ops_context * octx) {
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * src1 = &octx->src1;
-    struct htp_tensor *       dst  = &octx->dst;
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
 
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
     const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     // Use packed row sizes for VTCM allocation
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type = octx->src[0]->type;
     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;
@@ -799,9 +799,9 @@ static int execute_op_binary(struct htp_ops_context * octx) {
         return HTP_STATUS_VTCM_TOO_SMALL;
     }
 
-    octx->src0_spad.data = octx->ctx->vtcm_base;
-    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
-    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+    octx->src0_spad.data = octx->ctx->vtcm_base;                        octx->src0_spad.src = NULL;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size; octx->src1_spad.src = NULL;
+    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size; octx->dst_spad.src  = NULL;
 
     if ((octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
         return HTP_STATUS_OK;
@@ -857,12 +857,12 @@ static int execute_op_binary(struct htp_ops_context * octx) {
 int op_binary(struct htp_ops_context * octx) {
 
     // Does not support permutations of src1
-    const struct htp_tensor * src1 = &octx->src1;
+    const struct htp_tensor * src1 = octx->src[1];
     if (src1->nb[1] < src1->nb[0]) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t src0_type = octx->src0.type;
+    const uint32_t src0_type = octx->src[0]->type;
     if ((src0_type == HTP_TYPE_F32) || (src0_type == HTP_TYPE_F16)) {
         return execute_op_binary(octx);
     }

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

@@ -11,7 +11,7 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 #include "htp-ops.h"
 #include "hvx-utils.h"
 
@@ -32,10 +32,10 @@ struct htp_copy_context {
     void (*copy)(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith);
 };
 
-#define cpy_preamble                       \
-    struct htp_tensor *src0 = &octx->src0; \
-    struct htp_tensor *dst  = &octx->dst;  \
-                                           \
+#define cpy_preamble                              \
+    const struct htp_tensor *src0 = octx->src[0]; \
+    const struct htp_tensor *dst  = octx->dst;    \
+                                                  \
     const uint32_t ne00 = src0->ne[0];     \
     const uint32_t ne01 = src0->ne[1];     \
     const uint32_t ne02 = src0->ne[2];     \

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

@@ -13,9 +13,9 @@
 #include "hvx-utils.h"
 #include "hex-dma.h"
 
-#define htp_cumsum_tensors_preamble                  \
-    struct htp_tensor * restrict src0 = &octx->src0; \
-    struct htp_tensor * restrict dst  = &octx->dst;  \
+#define htp_cumsum_tensors_preamble                         \
+    const struct htp_tensor * restrict src0 = octx->src[0]; \
+    const struct htp_tensor * restrict dst  = octx->dst;    \
                                                      \
     const uint32_t ne00 = src0->ne[0];               \
     const uint32_t ne01 = src0->ne[1];               \
@@ -206,8 +206,8 @@ static void cumsum_thread_f32(unsigned int nth, unsigned int ith, void * data) {
 }
 
 int op_cumsum_f32(struct htp_ops_context * octx) {
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * dst  = &octx->dst;
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * dst  = octx->dst;
 
     if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
         return HTP_STATUS_OK;
@@ -226,10 +226,12 @@ int op_cumsum_f32(struct htp_ops_context * octx) {
 
     octx->src0_spad.size_per_thread = src_row_size_aligned * 2;
     octx->dst_spad.size_per_thread  = dst_row_size_aligned * 2;
-    octx->src0_spad.size            = n_threads * octx->src0_spad.size_per_thread;
-    octx->dst_spad.size             = n_threads * octx->dst_spad.size_per_thread;
-    octx->src0_spad.data            = octx->ctx->vtcm_base;
-    octx->dst_spad.data             = octx->src0_spad.data + octx->src0_spad.size;
+
+    octx->src0_spad.size  = n_threads * octx->src0_spad.size_per_thread;
+    octx->dst_spad.size   = n_threads * octx->dst_spad.size_per_thread;
+
+    octx->src0_spad.data  = octx->ctx->vtcm_base;                        octx->src0_spad.src = NULL;
+    octx->dst_spad.data   = octx->src0_spad.data + octx->src0_spad.size; octx->dst_spad.src  = NULL;
 
     struct htp_cumsum_context cctx = {
         .octx                 = octx,
@@ -251,8 +253,9 @@ int op_cumsum_f32(struct htp_ops_context * octx) {
 }
 
 int op_cumsum(struct htp_ops_context * octx) {
-    int                 err = HTP_STATUS_OK;
-    struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * dst = octx->dst;
+
+    int err = HTP_STATUS_OK;
 
     switch (dst->type) {
         case HTP_TYPE_F32:

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -15,7 +15,7 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 #include "htp-ops.h"
 
 // Must be multiple of 32
@@ -278,12 +278,12 @@ static inline void hvx_scale_vec_f32_aa(uint8_t * restrict dst, const uint8_t *
 static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void * data) {
     struct htp_fa_context * factx = (struct htp_fa_context *) data;
     const struct htp_ops_context * octx = factx->octx;
-    const struct htp_tensor * q = &octx->src0;
-    const struct htp_tensor * k = &octx->src1;
-    const struct htp_tensor * v = &octx->src2;
-    const struct htp_tensor * mask  = (octx->src3.data) ? &octx->src3 : NULL;
-    const struct htp_tensor * sinks = (octx->src4.data) ? &octx->src4 : NULL;
-    const struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * q     = octx->src[0];
+    const struct htp_tensor * k     = octx->src[1];
+    const struct htp_tensor * v     = octx->src[2];
+    const struct htp_tensor * mask  = octx->src[3];
+    const struct htp_tensor * sinks = octx->src[4];
+    const struct htp_tensor * dst   = octx->dst;
 
     const uint32_t neq0 = q->ne[0];
     const uint32_t neq1 = q->ne[1];
@@ -610,11 +610,11 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
 }
 
 int op_flash_attn_ext(struct htp_ops_context * octx) {
-    const struct htp_tensor * q = &octx->src0;
-    const struct htp_tensor * k = &octx->src1;
-    const struct htp_tensor * v = &octx->src2;
-    const struct htp_tensor * mask = (octx->src3.data) ? &octx->src3 : NULL;
-    const struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * q    = octx->src[0];
+    const struct htp_tensor * k    = octx->src[1];
+    const struct htp_tensor * v    = octx->src[2];
+    const struct htp_tensor * mask = octx->src[3];
+    const struct htp_tensor * dst  = octx->dst;
 
     // Check support
     if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) || k->type != HTP_TYPE_F16 || v->type != HTP_TYPE_F16) {
@@ -701,13 +701,11 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
         return HTP_STATUS_VTCM_TOO_SMALL;
     }
 
-    octx->src0_spad.data = octx->ctx->vtcm_base;
-    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
-    octx->src2_spad.data = octx->src1_spad.data + octx->src1_spad.size;
-    octx->src3_spad.data = octx->src2_spad.data + octx->src2_spad.size;
-    octx->dst_spad.data  = octx->src3_spad.data + octx->src3_spad.size;
-
-    // FARF(ERROR, "fa: qrows-per-thread %u", factx.qrows_per_thread);
+    octx->src0_spad.data = octx->ctx->vtcm_base;                        octx->src0_spad.src = NULL;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size; octx->src1_spad.src = NULL;
+    octx->src2_spad.data = octx->src1_spad.data + octx->src1_spad.size; octx->src2_spad.src = NULL;
+    octx->src3_spad.data = octx->src2_spad.data + octx->src2_spad.size; octx->src3_spad.src = NULL;
+    octx->dst_spad.data  = octx->src3_spad.data + octx->src3_spad.size; octx->dst_spad.src  = NULL;
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
         worker_pool_run_func(octx->ctx->worker_pool, flash_attn_ext_f16_thread, &factx, octx->n_threads);

ggml/src/ggml-hexagon/htp/get-rows-ops.c

@@ -11,7 +11,7 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 #include "htp-ops.h"
 #include "hvx-utils.h"
 
@@ -23,27 +23,33 @@ struct get_rows_context {
 };
 
 #define get_rows_preamble \
-    const uint32_t ne00 = octx->src0.ne[0]; \
-    const uint32_t ne01 = octx->src0.ne[1]; \
-    const uint32_t ne02 = octx->src0.ne[2]; \
-    const uint32_t ne03 = octx->src0.ne[3]; \
-                                            \
-    const uint32_t ne10 = octx->src1.ne[0]; \
-    const uint32_t ne11 = octx->src1.ne[1]; \
-    const uint32_t ne12 = octx->src1.ne[2]; \
-                                            \
-    const uint32_t nb01 = octx->src0.nb[1]; \
-    const uint32_t nb02 = octx->src0.nb[2]; \
-    const uint32_t nb03 = octx->src0.nb[3]; \
-                                            \
-    const uint32_t nb10 = octx->src1.nb[0]; \
-    const uint32_t nb11 = octx->src1.nb[1]; \
-    const uint32_t nb12 = octx->src1.nb[2]; \
-                                            \
-    const uint32_t nb1 = octx->dst.nb[1];   \
-    const uint32_t nb2 = octx->dst.nb[2];   \
-    const uint32_t nb3 = octx->dst.nb[3];   \
-                                            \
+    const uint32_t ne00 = octx->src[0]->ne[0]; \
+    const uint32_t ne01 = octx->src[0]->ne[1]; \
+    const uint32_t ne02 = octx->src[0]->ne[2]; \
+    const uint32_t ne03 = octx->src[0]->ne[3]; \
+                                               \
+    const uint32_t ne10 = octx->src[1]->ne[0]; \
+    const uint32_t ne11 = octx->src[1]->ne[1]; \
+    const uint32_t ne12 = octx->src[1]->ne[2]; \
+    const uint32_t ne13 = octx->src[1]->ne[3]; \
+                                               \
+    const uint32_t ne0 = octx->dst->ne[0];     \
+    const uint32_t ne1 = octx->dst->ne[1];     \
+    const uint32_t ne2 = octx->dst->ne[2];     \
+    const uint32_t ne3 = octx->dst->ne[3];     \
+                                               \
+    const uint32_t nb01 = octx->src[0]->nb[1]; \
+    const uint32_t nb02 = octx->src[0]->nb[2]; \
+    const uint32_t nb03 = octx->src[0]->nb[3]; \
+                                               \
+    const uint32_t nb10 = octx->src[1]->nb[0]; \
+    const uint32_t nb11 = octx->src[1]->nb[1]; \
+    const uint32_t nb12 = octx->src[1]->nb[2]; \
+                                               \
+    const uint32_t nb1 = octx->dst->nb[1];     \
+    const uint32_t nb2 = octx->dst->nb[2];     \
+    const uint32_t nb3 = octx->dst->nb[3];     \
+                                               \
     const uint32_t nr = ne10 * ne11 * ne12;
 
 static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
@@ -51,12 +57,14 @@ static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
     struct htp_ops_context * octx = grctx->octx;
     get_rows_preamble;
 
+    uint64_t qt = HAP_perf_get_qtimer_count();
+
     // parallelize by src1 elements (which correspond to dst rows)
     const uint32_t dr  = grctx->src1_nrows_per_thread;
     const uint32_t ir0 = dr * ith;
     const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;
 
-    const bool is_i32 = (octx->src1.type == HTP_TYPE_I32);
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
 
     for (uint32_t i = ir0; i < ir1; ++i) {
         const uint32_t i12 = fastdiv(i, &grctx->get_rows_div_ne10_ne11);
@@ -64,7 +72,7 @@ static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
         const uint32_t i11 = fastdiv(rem, &grctx->get_rows_div_ne10);
         const uint32_t i10 = rem - i11 * ne10;
 
-        const uintptr_t src1_addr = octx->src1.data + i10*nb10 + i11*nb11 + i12*nb12;
+        const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
 
         uint32_t i01 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
 
@@ -73,10 +81,14 @@ static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
             continue;
         }
 
-        const uintptr_t src0_ptr = octx->src0.data + i01*nb01 + i11*nb02 + i12*nb03;
-        const uintptr_t dst_ptr  = octx->dst.data  + i10*nb1  + i11*nb2  + i12*nb3;
+        const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i11*nb02 + i12*nb03;
+        const uintptr_t dst_ptr  = octx->dst->data    + i10*nb1  + i11*nb2  + i12*nb3;
         hvx_copy_f32_uu((uint8_t *)dst_ptr, (const uint8_t *)src0_ptr, ne00);
     }
+
+    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
+    FARF(HIGH, "get-rows-f32-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
 }
 
 int op_get_rows(struct htp_ops_context * octx) {
@@ -84,15 +96,15 @@ int op_get_rows(struct htp_ops_context * octx) {
 
     const uint32_t n_threads = MIN(nr, octx->n_threads);
 
-    if (octx->src0.type != HTP_TYPE_F32) {
+    if (octx->src[0]->type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    if (octx->dst.type != HTP_TYPE_F32) {
+    if (octx->dst->type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    if (octx->src1.type != HTP_TYPE_I32 && octx->src1.type != HTP_TYPE_I64) {
+    if (octx->src[1]->type != HTP_TYPE_I32 && octx->src[1]->type != HTP_TYPE_I64) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
@@ -102,8 +114,8 @@ int op_get_rows(struct htp_ops_context * octx) {
 
     struct get_rows_context grctx;
     grctx.octx = octx;
-    grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src1.ne[0]);
-    grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src1.ne[0] * octx->src1.ne[1]);
+    grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src[1]->ne[0]);
+    grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src[1]->ne[0] * octx->src[1]->ne[1]);
 
     grctx.src1_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 

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

@@ -3,8 +3,10 @@
 
 #include <stdbool.h>
 #include <stdint.h>
+#include <qurt_memory.h>
 
 #include "hexagon_types.h"
+#include "hexagon_protos.h"
 
 #include "hex-fastdiv.h"
 #include "hex-dump.h"
@@ -29,6 +31,14 @@ static inline uint64_t hex_get_pktcnt() {
     return pktcnt;
 }
 
+static inline uint32_t hex_ceil_pow2(uint32_t x) {
+    if (x <= 1) { return 1; }
+    int p = 2;
+    x--;
+    while (x >>= 1) { p <<= 1; }
+    return p;
+}
+
 static inline size_t hmx_ceil_div(size_t num, size_t den) {
     return (num + den - 1) / den;
 }
@@ -68,4 +78,26 @@ static inline void hex_l2fetch(const void * p, uint32_t width, uint32_t stride,
     Q6_l2fetch_AP((void *) p, control);
 }
 
+#define HEX_L2_LINE_SIZE  64
+#define HEX_L2_FLUSH_SIZE (128 * 1024)
+
+static inline void hex_l2flush(void * addr, size_t size) {
+    if (size > HEX_L2_FLUSH_SIZE) {
+        qurt_mem_cache_clean((qurt_addr_t) 0, 0, QURT_MEM_CACHE_FLUSH_INVALIDATE_ALL, QURT_MEM_DCACHE);
+    } else {
+        const uint32_t s = (uint32_t) addr;
+        const uint32_t e = s + size;
+        for (uint32_t i = s; i < e; i += HEX_L2_LINE_SIZE * 4) {
+            Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 0);
+            Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 1);
+            Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 2);
+            Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 3);
+        }
+    }
+}
+
+static inline void hex_pause() {
+    asm volatile(" pause(#255)\n");
+}
+
 #endif /* HEX_UTILS_H */

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

@@ -16,14 +16,16 @@
 #include "ggml-common.h"
 
 #include "hex-dma.h"
+#include "worker-pool.h"
+
 #include "hvx-utils.h"
 #include "hvx-dump.h"
-#include "worker-pool.h"
 #include "htp-ctx.h"
-#include "htp-msg.h"
+#include "htp-ops.h"
 
-#include "hmx-utils.h"
 #include "hmx-ops.h"
+#include "hmx-utils.h"
+#include "hmx-queue.h"
 #include "hmx-profile.h"
 
 static const __fp16 q4_0_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
@@ -47,7 +49,8 @@ static const __fp16 iq4_nl_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
 static const int32_t weight_transpose_scatter_offsets[32] __attribute__((aligned(VLEN))) = {
     0*128,  1*128,  2*128,  3*128,  4*128,  5*128,  6*128,  7*128,
     8*128,  9*128, 10*128, 11*128, 12*128, 13*128, 14*128, 15*128,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    16*128, 17*128, 18*128, 19*128, 20*128, 21*128, 22*128, 23*128,
+    24*128, 25*128, 26*128, 27*128, 28*128, 29*128, 30*128, 31*128
 };
 
 // Scales per x4x2 logical block: 8 × sizeof(__fp16) = 16 bytes
@@ -109,36 +112,45 @@ static inline bool hmx_add_overflow(size_t a, size_t b, size_t *out) {
     return false;
 }
 
-// Search for optimal (mc, nc) chunk sizes that maximize mc * nc within VTCM budget.
+// Search for optimal (mc, nc) chunk sizes within VTCM budget.
 //
-// Cost model: total = nc * per_n_cost + mc * per_m_cost + mc * nc * per_mn_cost + overhead
-//   per_n_cost:  bytes per nc column (weight + scratch buffers)
-//   per_m_cost:  bytes per mc row (activation)
-//   per_mn_cost: bytes per mc*nc element (output)
-//   overhead:    fixed bytes (scales 256B, eye_tile 2048B, etc.)
+// VTCM model: nc * per_n_cost + mc * per_m_cost + mc * nc * per_mn_cost + overhead
+//
+// Minimize ceil(m/mc) * m_block_cost + ceil(n/nc) * n_block_cost.
+// All matmul paths repeat weight processing per M-block and activation loading
+// per N-block, so discrete block counts drive total overhead.
+// Tie-break: when cost is equal, prefer larger mc * nc.
+//
+// Caller-provided coefficients:
+//   m_block_cost: penalty per extra M-block (weight redundancy, scales with n).
+//   n_block_cost: penalty per extra N-block (activation redundancy, scales with m).
 //
 // Algorithm: nc sweeps from n_max down by 32, analytically solving for mc_max.
 // Returns 0 on success, -1 if VTCM is insufficient.
-static int hmx_compute_chunks(
-    size_t vtcm_total, size_t overhead,
-    size_t per_n_cost, size_t per_m_cost, size_t per_mn_cost,
-    int m, int n,
-    size_t *m_chunk_out, size_t *n_chunk_out,
-    size_t *total_out)
-{
+static int hmx_compute_chunks(size_t   vtcm_total,
+                              size_t   overhead,
+                              size_t   per_n_cost,
+                              size_t   per_m_cost,
+                              size_t   per_mn_cost,
+                              int      m,
+                              int      n,
+                              size_t   m_block_cost,
+                              size_t   n_block_cost,
+                              size_t * m_chunk_out,
+                              size_t * n_chunk_out,
+                              size_t * total_out) {
     if (m <= 0 || n <= 0) return -1;
     if (vtcm_total <= overhead) return -1;
     if (per_n_cost == 0 || per_m_cost == 0 || per_mn_cost == 0) return -1;
 
     const size_t usable = vtcm_total - overhead;
-    size_t best_mn = 0, best_m = 0, best_n = 0;
+
+    size_t best_cost = SIZE_MAX;
+    size_t best_mn   = 0;
+    size_t best_m = 0, best_n = 0;
 
     const size_t n_max = hex_align_down((size_t)n, HMX_FP16_TILE_N_COLS);
     for (size_t nc = n_max; nc >= HMX_FP16_TILE_N_COLS; nc -= HMX_FP16_TILE_N_COLS) {
-        // Early exit: if nc * m_max cannot beat best, smaller nc won't either
-        if (nc * hex_align_down((size_t)m, HMX_FP16_TILE_N_ROWS) <= best_mn)
-            break;
-
         size_t n_fixed = 0, ncmn = 0, mc_denom = 0;
         if (hmx_mul_overflow(nc, per_n_cost, &n_fixed)) continue;
         if (n_fixed >= usable) goto next_nc;
@@ -152,10 +164,19 @@ static int hmx_compute_chunks(
             mc = hex_align_down(mc, HMX_FP16_TILE_N_ROWS);
             mc = hex_smin(mc, (size_t)m);
 
-            if (mc > 0 && mc * nc > best_mn) {
-                best_mn = mc * nc;
-                best_m  = mc;
-                best_n  = nc;
+            if (mc == 0) {
+                goto next_nc;
+            }
+
+            size_t mblocks = ((size_t) m + mc - 1) / mc;
+            size_t nblocks = ((size_t) n + nc - 1) / nc;
+            size_t cost    = mblocks * m_block_cost + nblocks * n_block_cost;
+            size_t mn      = mc * nc;
+            if (cost < best_cost || (cost == best_cost && mn > best_mn)) {
+                best_cost = cost;
+                best_mn   = mn;
+                best_m    = mc;
+                best_n    = nc;
             }
         }
 
@@ -233,7 +254,7 @@ static inline HVX_Vector dequantize_x4x2_q4_0_group_hvx(
     const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
     HVX_Vector v_scales = hvx_vec_splat_f16(*scale);
     // q4x4x2 stores two int4 values per byte. Keep only the selected nibble.
-    HVX_Vector v_quants = upper_nibbles ? Q6_Vub_vlsr_VubR(vq, 4) : vq;
+    HVX_Vector v_quants =  Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
     v_quants = Q6_V_vand_VV(v_quants, mask_h4);
     // Shuffle before LUT
     v_quants = Q6_Vb_vshuff_Vb(v_quants);
@@ -257,7 +278,7 @@ static inline void dequantize_x4x2_q4_0_x4groups_hvx(
     // Load all 128 packed bytes (4 contiguous 32-byte groups)
     HVX_Vector vq = hvx_vmemu(packed_128);
     const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
-    HVX_Vector v_quants = upper_nibbles ? Q6_Vub_vlsr_VubR(vq, 4) : vq;
+    HVX_Vector v_quants = Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
     v_quants = Q6_V_vand_VV(v_quants, mask_h4);
 
     // Shuffle before LUT
@@ -277,10 +298,8 @@ static inline void dequantize_x4x2_q4_0_x4groups_hvx(
     v_hi = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hi, v_sc23));
 
     // Extract individual groups: scatter uses q_mask64 so only first 64 bytes matter
-    out[0] = v_lo;                      // group0 already in [0:63]
-    out[1] = Q6_V_vror_VR(v_lo, 64);    // group1 rotated to [0:63]
-    out[2] = v_hi;                      // group2 already in [0:63]
-    out[3] = Q6_V_vror_VR(v_hi, 64);    // group3 rotated to [0:63]
+    out[0] = v_lo; // group0 already in [0:63]
+    out[1] = v_hi; // group2 already in [0:63]
 }
 
 // Dequantize one x4x2 Q8_0 group (32 int8 quants) -> 32 FP16 in first 64 bytes.
@@ -384,8 +403,9 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
         size_t row_stride, int weight_type,
         int start_tile, int end_tile) {
 
-    const int n_k_tiles = k_block / HMX_FP16_TILE_N_COLS;
-    const int qrow_size = (weight_type == HTP_TYPE_Q8_0) ? k_block : (k_block / 2);
+    const int n_k_tiles = (unsigned)k_block / HMX_FP16_TILE_N_COLS;
+    const bool is_q4 = (weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL);
+    const int qrow_size = is_q4 ? ((unsigned)k_block / 2) : k_block;
 
     const HVX_Vector vlut_cvt = (weight_type == HTP_TYPE_IQ4_NL) ? hvx_vmem(iq4_nl_to_fp16_lut) :
                                 (weight_type == HTP_TYPE_MXFP4)  ? hvx_vmem(mxfp4_to_fp16_lut) :
@@ -398,47 +418,46 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
     const HVX_Vector v_scat_step = Q6_V_vsplat_R(4);  // 4 bytes = 1 column step
     const HVX_VectorPred q_mask64 = Q6_Q_vsetq_R(64);  // first 16 words (64 bytes)
 
-    for (int t = start_tile; t < end_tile; ) {
-        int ct = t / n_k_tiles;  // column tile index
-        int kt = t % n_k_tiles;  // K tile index
+    unsigned ct = (unsigned)start_tile / n_k_tiles;  // column tile index
+    unsigned kt = (unsigned)start_tile % n_k_tiles;  // K tile index
+    for (unsigned t = start_tile; t < end_tile; ) {
+        if (kt >= n_k_tiles) { kt = 0; ct++; }
 
-        // --- Batch-4 fast path for Q4_0/IQ4_NL: process 4 contiguous K-tiles with one vlut16 per row ---
-        if ((weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL) && (kt % 4 == 0) && (t + 4 <= end_tile) &&
-            ((t + 3) / n_k_tiles == ct)) {
-            int blk_idx      = (kt * 32) / QK_Q4_0x4x2;
-            int sub_blk_base = ((kt * 32) % QK_Q4_0x4x2) / 32;  // 0 or 4
-            bool upper       = (sub_blk_base >= 4);
-            int packed_off   = blk_idx * (QK_Q4_0x4x2 / 2);     // 128 contiguous packed bytes
-            int scale_off    = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE
-                             + sub_blk_base * (int)sizeof(__fp16);   // 4 consecutive scales
+        // --- Batch-4 fast path for Q4: process 4 contiguous K-tiles with one vlut16 per row ---
+        if (is_q4 && (kt % 4 == 0) && (t + 4 <= end_tile) && ((t + 3) / n_k_tiles == ct)) {
+            unsigned blk_idx      = (kt * 32) / QK_Q4_0x4x2;
+            unsigned sub_blk_base = ((kt * 32) % QK_Q4_0x4x2) / 32;  // 0 or 4
+            bool upper            = (sub_blk_base >= 4);
+            unsigned packed_off   = blk_idx * (QK_Q4_0x4x2 / 2);     // 128 contiguous packed bytes
+            unsigned scale_off    = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE
+                                  + sub_blk_base * (int)sizeof(__fp16);   // 4 consecutive scales
 
             __fp16 *tile_bases[4];
-            for (int g = 0; g < 4; g++) { tile_bases[g] = vtcm_dst + (t + g) * HMX_FP16_TILE_N_ELMS; }
+            for (unsigned g = 0; g < 4; g++) { tile_bases[g] = vtcm_dst + (t + g) * HMX_FP16_TILE_N_ELMS; }
 
             HVX_Vector v_off = v_scat_base;
-            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {
-                int row0 = ct * HMX_FP16_TILE_N_COLS + r;
-                int row1 = row0 + 1;
-                const uint8_t *r0 = vtcm_src + row0 * row_stride;
-                const uint8_t *r1 = vtcm_src + row1 * row_stride;
 
-                HVX_Vector v0[4], v1[4];
+            unsigned row_offset = ct * HMX_FP16_TILE_N_COLS * row_stride;
+            unsigned row1 = ct * HMX_FP16_TILE_N_COLS + 1;
+
+            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2, row1 += 2) {
+                HVX_Vector v0[2];
+                const uint8_t *r0 = vtcm_src + row_offset; row_offset += row_stride;
                 dequantize_x4x2_q4_0_x4groups_hvx(r0 + packed_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt, v0);
-                if (row1 < n_cols) {
-                    dequantize_x4x2_q4_0_x4groups_hvx(r1 + packed_off, upper, (const __fp16 *)(r1 + scale_off), vlut_cvt, v1);
-                } else {
-                    v1[0] = v1[1] = v1[2] = v1[3] = Q6_V_vzero();
-                }
-
-                for (int g = 0; g < 4; g++) { Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_bases[g], HMX_FP16_TILE_SIZE - 1, v_off, v0[g]); }
+                Q6_vscatter_RMVwV((size_t)tile_bases[0], 2 * HMX_FP16_TILE_SIZE - 1, v_off, v0[0]);
+                Q6_vscatter_RMVwV((size_t)tile_bases[2], 2 * HMX_FP16_TILE_SIZE - 1, v_off, v0[1]);
                 v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
-                for (int g = 0; g < 4; g++) { Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_bases[g], HMX_FP16_TILE_SIZE - 1, v_off, v1[g]); }
+
+
+                r0 = vtcm_src + row_offset; row_offset += row_stride;
+                dequantize_x4x2_q4_0_x4groups_hvx(r0 + packed_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt, v0);
+                Q6_vscatter_RMVwV((size_t)tile_bases[0], 2 * HMX_FP16_TILE_SIZE - 1, v_off, v0[0]);
+                Q6_vscatter_RMVwV((size_t)tile_bases[2], 2 * HMX_FP16_TILE_SIZE - 1, v_off, v0[1]);
                 v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
             }
 
             for (int g = 0; g < 4; g++) { (void) *(volatile HVX_Vector *)(tile_bases[g]); }
-
-            t += 4;
+            t += 4; kt += 4;
             continue;
         }
 
@@ -495,20 +514,19 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
         // --- Single-tile fallback ---
         __fp16 *tile_base = vtcm_dst + t * HMX_FP16_TILE_N_ELMS;
 
-        if (weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL) {
-            int blk_idx  = (kt * 32) / QK_Q4_0x4x2;
-            int sub_blk  = ((kt * 32) % QK_Q4_0x4x2) / 32;
-            bool upper   = (sub_blk >= 4);
-            int byte_off  = blk_idx * (QK_Q4_0x4x2 / 2) + (upper ? (sub_blk - 4) : sub_blk) * 32;
-            int scale_off = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE + sub_blk * (int)sizeof(__fp16);
+        if (is_q4) {
+            unsigned blk_idx   = (kt * 32) / QK_Q4_0x4x2;
+            unsigned sub_blk   = ((kt * 32) % QK_Q4_0x4x2) / 32;
+            bool upper         = (sub_blk >= 4);
+            unsigned byte_off  = blk_idx * (QK_Q4_0x4x2 / 2) + (upper ? (sub_blk - 4) : sub_blk) * 32;
+            unsigned scale_off = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE + sub_blk * (int)sizeof(__fp16);
 
             HVX_Vector v_off = v_scat_base;  // reset to column 0
-            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {
-                int row0 = ct * HMX_FP16_TILE_N_COLS + r;
-                int row1 = row0 + 1;
-
-                const uint8_t *r0 = vtcm_src + row0 * row_stride;
-                const uint8_t *r1 = vtcm_src + row1 * row_stride;
+            unsigned row_offset = ct * HMX_FP16_TILE_N_COLS * row_stride;
+            unsigned row1 = ct * HMX_FP16_TILE_N_COLS + 1;
+            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2, row1 += 2) {
+                const uint8_t *r0 = vtcm_src + row_offset; row_offset += row_stride;
+                const uint8_t *r1 = vtcm_src + row_offset; row_offset += row_stride;
 
                 HVX_Vector v0 = dequantize_x4x2_q4_0_group_hvx(
                     r0 + byte_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt);
@@ -585,7 +603,7 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
             }
             (void) *(volatile HVX_Vector *)(tile_base);
         }
-        ++t;
+        ++t; ++kt;
     }
 
     // Drain HVX scatter write buffer: a vmem load on the same HW thread retires
@@ -630,9 +648,9 @@ static void dequantize_x4x2_weight_chunk_to_fp16_tiles(
     assert(n_cols  % HMX_FP16_TILE_N_COLS == 0);
     assert(k_block % HMX_FP16_TILE_N_COLS == 0);
 
-    int n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
-    int n_k_tiles   = k_block / HMX_FP16_TILE_N_COLS;
-    int n_tot_tiles = n_col_tiles * n_k_tiles;
+    size_t n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
+    size_t n_k_tiles   = k_block / HMX_FP16_TILE_N_COLS;
+    size_t n_tot_tiles = n_col_tiles * n_k_tiles;
 
     size_t n_tiles_per_task = hmx_ceil_div(n_tot_tiles, ctx->n_threads);
 
@@ -653,49 +671,91 @@ static void dequantize_x4x2_weight_chunk_to_fp16_tiles(
 // --- End x4x2 dequantizers ---
 
 // requires external HMX lock
-static void core_dot_chunk_fp16(__fp16 *output, const __fp16 *activation, const __fp16 *weight, const __fp16 *scales,
+static void core_dot_chunk_fp16(__fp16 *restrict output, const __fp16 *restrict activation, const __fp16 *restrict weight, const __fp16 *restrict scales,
                                 int n_row_tiles, int n_col_tiles, int n_dot_tiles) {
-    hmx_set_output_scales(scales);
+    __builtin_assume(n_row_tiles > 0);
+    __builtin_assume(n_col_tiles > 0);
+    __builtin_assume(n_dot_tiles > 0);
 
+    Q6_bias_mxmem2_A((void *)scales);
     for (int r = 0; r < n_row_tiles; ++r) {
-        for (int c = 0; c < n_col_tiles; ++c) {
+        for (size_t c = 0; c < n_col_tiles; ++c) {
             Q6_mxclracc_hf();
 
             const __fp16 *row_tiles = activation + r * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
             const __fp16 *col_tiles = weight + c * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
 
             for (int k = 0; k < n_dot_tiles; ++k) {
-                int offset = k * HMX_FP16_TILE_N_ELMS;
-                hmx_load_tile_pair_fp16(row_tiles + offset, col_tiles + offset);
+                Q6_activation_hf_mxmem_RR((unsigned int)row_tiles, 2047);
+                Q6_weight_hf_mxmem_RR((unsigned int)col_tiles, 2047);
+                row_tiles += HMX_FP16_TILE_N_ELMS;
+                col_tiles += HMX_FP16_TILE_N_ELMS;
             }
 
             __fp16 *out_tile = output + (r * n_col_tiles + c) * HMX_FP16_TILE_N_ELMS;
-            hmx_consume_accumulator_fp16(out_tile);
+            Q6_mxmem_AR_after_hf(out_tile, 0);
         }
     }
 }
 
+// --- Async HMX matmul job (for pipeline overlap) ---
+
+typedef struct {
+    __fp16 *       output;
+    const __fp16 * activation;
+    const __fp16 * weight;
+    const __fp16 * scales;
+    uint32_t       n_row_tiles;
+    uint32_t       n_col_tiles;
+    uint32_t       n_dot_tiles;
+} hmx_matmul_job_t;
+
+static void hmx_matmul_worker_fn(void * data) {
+    hmx_matmul_job_t * job = (hmx_matmul_job_t *) data;
+    FARF(HIGH, "hmx-mm-job: n_row_tiles %u n_col_tiles %u n_dot_tiles %u", job->n_row_tiles, job->n_col_tiles, job->n_dot_tiles);
+    core_dot_chunk_fp16(job->output, job->activation, job->weight, job->scales, job->n_row_tiles, job->n_col_tiles, job->n_dot_tiles);
+}
+
+static inline void hmx_matmul_job_init(hmx_matmul_job_t * job,
+                                       __fp16 *           output,
+                                       const __fp16 *     activation,
+                                       const __fp16 *     weight,
+                                       const __fp16 *     scales,
+                                       int                n_row_tiles,
+                                       int                n_col_tiles,
+                                       int                n_dot_tiles) {
+    job->output      = output;
+    job->activation  = activation;
+    job->weight      = weight;
+    job->scales      = scales;
+    job->n_row_tiles = n_row_tiles;
+    job->n_col_tiles = n_col_tiles;
+    job->n_dot_tiles = n_dot_tiles;
+}
+
+// --- End async HMX matmul job ---
+
 static void transfer_output_chunk_fp16_to_fp32(float *restrict dst, const __fp16 *restrict vtcm_src, int n_rows, int n_cols, int n) {
     assert(n_cols % HMX_FP16_TILE_N_COLS == 0);
-    const int n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
+    const size_t tile_row_stride = (n_cols / HMX_FP16_TILE_N_COLS) * HMX_FP16_TILE_N_ELMS;
 
     const HVX_Vector one = hvx_vec_splat_f16(1.0);
 
-    for (int r = 0; r < n_rows; r += 2) {
-        int r0 = r / HMX_FP16_TILE_N_ROWS;
-        int r1 = r % HMX_FP16_TILE_N_ROWS;
+    for (size_t r = 0; r < n_rows; r += 2) {
+        const size_t r0 = r / HMX_FP16_TILE_N_ROWS;
+        const size_t r1 = (r % HMX_FP16_TILE_N_ROWS) / 2;  // index of the row pair within the tile
+        const __fp16 *row_base = vtcm_src + r0 * tile_row_stride;
+        float *output_row_base = dst + r * n;  // global memory row base for row r (and r+1)
 
         #pragma unroll(4)
-        for (int c = 0; c < n_cols; c += HMX_FP16_TILE_N_COLS) {
-            int c0 = c / HMX_FP16_TILE_N_COLS;
-
-            const __fp16 *tile = vtcm_src + (r0 * n_col_tiles + c0) * HMX_FP16_TILE_N_ELMS;
-
-            HVX_Vector v = ((const HVX_Vector *) tile)[r1 / 2];
+        for (size_t c = 0; c < n_cols; c += HMX_FP16_TILE_N_COLS) {
+            const size_t c0 = c / HMX_FP16_TILE_N_COLS;
+            const __fp16 *tile = row_base + c0 * HMX_FP16_TILE_N_ELMS;
+            HVX_Vector v = ((const HVX_Vector *) tile)[r1];
             HVX_VectorPair vp = Q6_Wqf32_vmpy_VhfVhf(v, one);
 
-            volatile HVX_Vector *pv_out0 = (volatile HVX_Vector *) (dst + (r * n + c + 0));
-            volatile HVX_Vector *pv_out1 = (volatile HVX_Vector *) (dst + (r * n + c + n));  // next row in global memory
+            volatile HVX_Vector *pv_out0 = (volatile HVX_Vector *) (output_row_base + c + 0);
+            volatile HVX_Vector *pv_out1 = (volatile HVX_Vector *) (output_row_base + c + n);  // next row in global memory
 
             *pv_out0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(vp));
             if (r + 1 < n_rows) {
@@ -733,7 +793,7 @@ static void transfer_output_chunk_threaded(struct htp_context *ctx, float *dst,
     assert(n_cols % HMX_FP16_TILE_N_COLS == 0);
 
     size_t n_tot_chunks      = n_rows;
-    size_t n_chunks_per_task = 32;  // must be multiple of HMX_FP16_TILE_N_ROWS (32)
+    size_t n_chunks_per_task = HMX_FP16_TILE_N_ROWS;  // must be multiple of HMX_FP16_TILE_N_ROWS (32)
 
     output_transfer_task_state_t state;
     state.n_tasks           = (n_tot_chunks + n_chunks_per_task - 1) / n_chunks_per_task;
@@ -821,7 +881,7 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
     // and each q_head is computed individually to avoid tile-major packing
     // issues.  m_chunk_n_rows is always a multiple of 32 (from
     // hmx_compute_chunks), so per-head tile arrays don't overlap.
-    const size_t vtcm_budget  = ctx->vtcm_scratch_size;
+    const size_t vtcm_budget  = ctx->vtcm_size;
     const size_t vec_dot_size = params->k * sizeof(__fp16);
 
     // When the activation has a large stride (e.g. permuted Q tensor with
@@ -832,12 +892,13 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
     const size_t f32_scratch_per_m = use_dma_activation ? (size_t) params->k * sizeof(float) : 0;
 
     size_t m_chunk_n_rows = 0, n_chunk_n_cols = 0, vtcm_used = 0;
+    // FP16 weight: interleave and activation load have similar per-element cost.
     if (hmx_compute_chunks(vtcm_budget, /*overhead=*/256,
-                             /*per_n=*/3 * vec_dot_size,
-                             /*per_m=*/group_size * vec_dot_size + f32_scratch_per_m,
-                             /*per_mn=*/sizeof(__fp16),
-                             params->m, params->n,
-                             &m_chunk_n_rows, &n_chunk_n_cols, &vtcm_used) != 0) {
+                           /*per_n=*/3 * vec_dot_size,
+                           /*per_m=*/group_size * vec_dot_size + f32_scratch_per_m,
+                           /*per_mn=*/sizeof(__fp16), params->m, params->n,
+                           /*m_block_cost=*/(size_t) params->n,
+                           /*n_block_cost=*/(size_t) params->m, &m_chunk_n_rows, &n_chunk_n_cols, &vtcm_used) != 0) {
         FARF(HIGH, "%s: grouped path does not fit VTCM, falling back to legacy batched loop", __func__);
         return hmx_mat_mul_permuted_w16a32_batched_legacy(ctx, params);
     }
@@ -864,7 +925,7 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
         return hmx_mat_mul_permuted_w16a32_batched_legacy(ctx, params);
     }
 
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     FARF(MEDIUM, "%s: grouped path m=%d k=%d n=%d group=%d streams=%d mc=%zu nc=%zu vtcm=%zu/%zu",
             __func__, params->m, params->k, params->n, group_size, params->ne13,
@@ -882,12 +943,15 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
     const size_t fp16_row_bytes   = (size_t) params->k * sizeof(__fp16);
     const size_t weight_row_bytes = (size_t) params->weight_stride * sizeof(__fp16);
 
+    HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
+
     for (int b3 = 0; b3 < params->ne13; ++b3) {
         for (int b2_base = 0; b2_base < params->ne12; b2_base += group_size) {
             const __fp16 *weight_group = hmx_matmul_weight_batch_ptr(params, b2_base, b3);
 
             for (size_t mr = 0; mr < (size_t) params->m; mr += m_chunk_n_rows) {
                 const size_t n_rows = hex_smin((size_t) params->m - mr, m_chunk_n_rows);
+                const size_t n_row_tiles = hmx_ceil_div((int) n_rows, HMX_FP16_TILE_N_ROWS);
 
                 // Pre-load activations for all heads in the group (once per m_chunk).
                 // When the source is strided (permuted Q), use 2D DMA to gather
@@ -925,10 +989,9 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                                       fp16_row_bytes, weight_row_bytes, fp16_row_bytes, n_cols_first);
                 }
 
-                HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
-
                 for (size_t nc = 0; nc < (size_t) params->n; nc += n_chunk_n_cols) {
                     const size_t n_cols = hex_smin((size_t) params->n - nc, n_chunk_n_cols);
+                    const size_t n_col_tiles = hmx_ceil_div((int) n_cols, HMX_FP16_TILE_N_COLS);
 
                     TIMER_START(weight_load);
                     {
@@ -952,11 +1015,9 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                     for (int g = 0; g < group_size; ++g) {
                         TIMER_START(hmx_core);
                         {
-                            const __fp16 *vtcm_act_g = vtcm_activation + (size_t) g * act_head_stride;
-                            const int n_row_tiles = hmx_ceil_div((int) n_rows, HMX_FP16_TILE_N_ROWS);
-                            const int n_col_tiles = hmx_ceil_div((int) n_cols, HMX_FP16_TILE_N_COLS);
-                            core_dot_chunk_fp16(vtcm_output, vtcm_act_g, vtcm_weight, vtcm_scales,
-                                                n_row_tiles, n_col_tiles, params->k / 32);
+                            const __fp16 * vtcm_act_g = vtcm_activation + (size_t) g * act_head_stride;
+                            core_dot_chunk_fp16(vtcm_output, vtcm_act_g, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles,
+                                                params->k / 32);
                         }
                         TIMER_STOP(hmx_core);
 
@@ -968,12 +1029,12 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                         TIMER_STOP(output_store);
                     }
                 }
-
-                HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
             }
         }
     }
 
+    HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
+
     TIMER_STOP(total);
 
 #if defined(ENABLE_PROFILE_TIMERS)
@@ -998,7 +1059,7 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
     }
 
     // --- Dynamic VTCM layout ---
-    const size_t vtcm_budget  = ctx->vtcm_scratch_size;
+    const size_t vtcm_budget  = ctx->vtcm_size;
     const size_t vec_dot_size = k * sizeof(__fp16);
 
     // DMA-based activation gather for strided tensors (see batched path comment).
@@ -1006,13 +1067,15 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
     const size_t f32_scratch_per_m = use_dma_activation ? (size_t) k * sizeof(float) : 0;
 
     size_t m_chunk_n_rows = 0, n_chunk_n_cols = 0, vtcm_used = 0;
+    // FP16 weight: interleave and activation load have similar per-element cost.
     if (hmx_compute_chunks(vtcm_budget,
-                              /*overhead=*/ 256,
-                              /*per_n=*/    3 * vec_dot_size,  // W + S0 + S1
-                              /*per_m=*/    vec_dot_size + f32_scratch_per_m,  // A + optional F32 scratch
-                              /*per_mn=*/   sizeof(__fp16),     // O
-                              m, n,
-                              &m_chunk_n_rows, &n_chunk_n_cols, &vtcm_used) != 0) {
+                           /*overhead=*/256,
+                           /*per_n=*/3 * vec_dot_size,                  // W + S0 + S1
+                           /*per_m=*/vec_dot_size + f32_scratch_per_m,  // A + optional F32 scratch
+                           /*per_mn=*/sizeof(__fp16),                   // O
+                           m, n,
+                           /*m_block_cost=*/(size_t) n,
+                           /*n_block_cost=*/(size_t) m, &m_chunk_n_rows, &n_chunk_n_cols, &vtcm_used) != 0) {
         FARF(HIGH, "%s: VTCM too small (m=%d k=%d n=%d budget=%zu)", __func__, m, k, n, vtcm_budget);
         return -1;
     }
@@ -1039,7 +1102,7 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
         return -1;
     }
 
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     FARF(MEDIUM, "%s: m=%d k=%d n=%d mc=%zu nc=%zu vtcm=%zu/%zu",
          __func__, m, k, n, m_chunk_n_rows, n_chunk_n_cols,
@@ -1057,7 +1120,8 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
 
     for (size_t mr = 0; mr < m; mr += m_chunk_n_rows) {
         // transfer activation matrix chunk into VTCM
-        size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+        const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+        const size_t n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
 
         TIMER_START(activation_load);
         {
@@ -1095,7 +1159,8 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
         }
 
         for (size_t nc = 0; nc < n; nc += n_chunk_n_cols) {
-            size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+            const size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+            const size_t n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
 
             TIMER_START(weight_load);
             {
@@ -1120,8 +1185,6 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
 
             TIMER_START(hmx_core);
             {
-                const int n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
-                const int n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
                 core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles, k / 32);
             }
             TIMER_STOP(hmx_core);
@@ -1157,6 +1220,8 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
 int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict out, const float *restrict x, const uint8_t *restrict w, int m,
                                        int k, int n, int w_type);
 
+#define FALLBACK_TO_STANDARD 1
+
 int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict dst, const float *restrict activation,
                                      const uint8_t *restrict permuted_weight, int m, int k, int n,
                                      int weight_type) {
@@ -1169,9 +1234,12 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
 
     // for large m, k (e.g. prefill FFN Down), use out-stationary version
     if (m >= 128 && k > n && n > 1024) {
-        FARF(MEDIUM, "hmx_matmul_qk: OUT-STATIONARY path m=%d k=%d n=%d type=%d (K_BLOCK=512, %d K-iters with fp16 intermediate)",
-             m, k, n, weight_type, (k + 511) / 512);
-        return mat_mul_qk_0_d16a32_out_stationary(ctx, dst, activation, permuted_weight, m, k, n, weight_type);
+        int rc = mat_mul_qk_0_d16a32_out_stationary(ctx, dst, activation, permuted_weight, m, k, n, weight_type);
+        if (rc != FALLBACK_TO_STANDARD) {
+            return rc;  // 0 success, -1 error
+        }
+        FARF(MEDIUM, "hmx_matmul_qk: out-stationary fallback to standard m=%d k=%d n=%d", m, k, n);
+        // fall through to standard path
     }
 
     size_t row_stride = get_x4x2_row_stride(weight_type, k);
@@ -1182,7 +1250,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
     FARF(MEDIUM, "hmx_matmul_qk: STANDARD path m=%d k=%d n=%d type=%d", m, k, n, weight_type);
 
     // --- Dynamic VTCM layout ---
-    const size_t vtcm_budget   = ctx->vtcm_scratch_size;
+    const size_t vtcm_budget   = ctx->vtcm_size;
     const size_t vec_dot_size  = k * sizeof(__fp16);
     const bool   use_pipeline  = (m >= 128) && (k <= n);
 
@@ -1197,9 +1265,10 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
     }
 
     size_t m_chunk_n_rows = 0, n_chunk_n_cols = 0, vtcm_used = 0;
-    if (hmx_compute_chunks(vtcm_budget, /*overhead=*/256,
-                              per_n_cost, /*per_m=*/vec_dot_size, per_mn_cost,
-                              m, n, &m_chunk_n_rows, &n_chunk_n_cols, &vtcm_used) != 0) {
+    // Quantized weight: dequant ~1.5x more expensive per element than activation load.
+    if (hmx_compute_chunks(vtcm_budget, /*overhead=*/256, per_n_cost, /*per_m=*/vec_dot_size, per_mn_cost, m, n,
+                           /*m_block_cost=*/(size_t) n * 3,
+                           /*n_block_cost=*/(size_t) m * 2, &m_chunk_n_rows, &n_chunk_n_cols, &vtcm_used) != 0) {
         FARF(HIGH, "%s: VTCM too small (m=%d k=%d n=%d pipe=%d budget=%zu)",
              __func__, m, k, n, use_pipeline, vtcm_budget);
         return -1;
@@ -1237,7 +1306,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
         return -1;
     }
 
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     FARF(MEDIUM, "%s: m=%d k=%d n=%d wtype=%d pipe=%d mc=%zu nc=%zu vtcm=%zu/%zu",
          __func__, m, k, n, weight_type, use_pipeline,
@@ -1256,12 +1325,12 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
          use_pipeline ? "PIPELINE" : "SEQUENTIAL", m_chunk_n_rows, n_chunk_n_cols,
          (size_t)(vtcm_ptr - (uint8_t *)ctx->vtcm_base), vtcm_budget);
 
-    HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
-
     if (!use_pipeline) {
+        HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
         for (size_t mr = 0; mr < m; mr += m_chunk_n_rows) {
             // transfer activation matrix chunk into VTCM
-            size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+            const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+            const size_t n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
 
             TIMER_START(activation_load);
             {
@@ -1273,16 +1342,14 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
             void *buf_curr = vtcm_scratch0;
             void *buf_next = vtcm_scratch1;
 
-            // issue async DDR data transfer for the first weight chunk
-            // NOTE: use 2D DMA (n_cols rows x row_stride bytes) instead of 1D
-            // 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(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) {
-                size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+                const size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+                const size_t n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
 
                 TIMER_START(weight_load);
                 {
@@ -1307,8 +1374,6 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
 
                 TIMER_START(hmx_core);
                 {
-                    const int n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
-                    const int n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
                     core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles, k / 32);
                 }
                 TIMER_STOP(hmx_core);
@@ -1321,20 +1386,22 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
                 TIMER_STOP(output_store);
             }
         }
+        HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
     } else {
         // 4-stage pipeline: DMA load (A), dequantize (B), HMX matmul (C), store (D)
-        // stage B and D (dequantize and store) are expected to be on the critical path
+        // HMX compute (C) runs on dedicated worker thread, overlapping with HVX stages (B, D).
 
         // A --> B: vtcm_qweight, 1 buffer
         // B --> C: vtcm_weight0/vtcm_weight1, 2 buffers
         // C --> D: vtcm_output0/vtcm_output1, 2 buffers
 
-        //
-        // LD ||A3|  | B3 ||
-        // MM ||    C2    ||
-        // ST || D1 |     ||
+        // Async timeline (C overlaps B+D):
+        //   main+HVX:   [A0][Act][B0][A1][sub C0][B1‖C0][A2][wait,sub C1][D0+B2‖C1][wait,sub C2][D1‖C2][wait][D2]
+        //   HMX queue:                   [████ C0 ████████][████ C1 ████████████][████ C2 ████████]
 
         int n_chunk_cnt = hmx_ceil_div(n, n_chunk_n_cols);
+        hmx_matmul_job_t job_slots[2];  // persistent double-buffered job descriptors
+
         for (size_t mr = 0; mr < m; mr += m_chunk_n_rows) {
             const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
 
@@ -1355,31 +1422,34 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
                 transfer_activation_chunk_threaded(ctx, vtcm_activation, activation_chunk, n_rows, k, k);
             }
 
-            // prologue: B0, A1, C0, B1
+            // prologue: B0, A1, submit C0 (async), B1 (overlaps C0)
             {
-                // B0
+                // B0: wait for DMA, dequant weight chunk 0
                 dma_queue_pop(ctx->dma[0]);
                 dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[0], vtcm_qweight, n_cols_A0, k, row_stride, weight_type);
 
-                // A1
+                // A1: issue DMA for weight chunk 1
                 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(ctx->dma[0], dma_make_ptr(vtcm_qweight, qweight_chunk_A1), row_stride, row_stride, row_stride, n_cols_A1);
                 }
 
-                // C0
-                core_dot_chunk_fp16((__fp16 *) vtcm_output_bufs[0], (__fp16 *) vtcm_activation, (__fp16 *) vtcm_weight_bufs[0], vtcm_scales,
-                         hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS), hmx_ceil_div(n_cols_A0, HMX_FP16_TILE_N_COLS), k / HMX_FP16_TILE_N_ROWS);
+                // submit C0 (non-blocking — HMX worker executes in parallel)
+                hmx_matmul_job_init(&job_slots[0], (__fp16 *) vtcm_output_bufs[0], (__fp16 *) vtcm_activation,
+                                    (__fp16 *) vtcm_weight_bufs[0], vtcm_scales,
+                                    hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS),
+                                    hmx_ceil_div(n_cols_A0, HMX_FP16_TILE_N_COLS), k / HMX_FP16_TILE_N_ROWS);
+                hmx_queue_push(ctx->hmx_queue, hmx_queue_make_desc(hmx_matmul_worker_fn, &job_slots[0]));
 
-                // B1
+                // B1: DMA pop + dequant (runs in parallel with C0 on HMX worker)
                 if (1 < n_chunk_cnt) {
                     dma_queue_pop(ctx->dma[0]);
                     dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[1], vtcm_qweight, n_cols_A1, k, row_stride, weight_type);
                 }
             }
 
-            // main loop
+            // main loop: wait C_i → submit C_{i+1} → D_i + B_{i+2} (parallel with C_{i+1})
             for (int i = 0; i < n_chunk_cnt; ++i) {
                 const size_t nc    = i * n_chunk_n_cols;
                 const size_t nc_p1 = nc + 1 * n_chunk_n_cols;
@@ -1389,36 +1459,41 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
                 const size_t n_cols_p1 = hex_smin(n - nc_p1, n_chunk_n_cols);
                 const size_t n_cols_p2 = hex_smin(n - nc_p2, n_chunk_n_cols);
 
-                // issue A_{i+2}
+                // issue A_{i+2}: DMA push (non-blocking)
                 if (i + 2 < n_chunk_cnt) {
                     const uint8_t *qweight_chunk_p2 = permuted_weight + nc_p2 * row_stride;
                     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
+                // wait C_i: block until prologue/previous C completes
+                hmx_queue_pop(ctx->hmx_queue);
 
-                // result of B_{i+1} (input of C_{i+1}) should be ready now
-
-                // issue C_{i+1}
+                // submit C_{i+1} (non-blocking, overlaps with D_i + B_{i+2} below)
+                // job_slots[(i+1)%2] is safe: C_i just completed, freeing slot i%2's
+                // counterpart — and (i+1)%2 was last used by C_{i-1} which completed
+                // before C_i was submitted.
                 if (i + 1 < n_chunk_cnt) {
-                    core_dot_chunk_fp16((__fp16 *) vtcm_output_bufs[(i + 1) % 2], (__fp16 *) vtcm_activation, (__fp16 *) vtcm_weight_bufs[(i + 1) % 2], vtcm_scales,
-                        hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS), hmx_ceil_div(n_cols_p1, HMX_FP16_TILE_N_COLS), k / HMX_FP16_TILE_N_ROWS);
+                    hmx_matmul_job_init(&job_slots[(i + 1) % 2], (__fp16 *) vtcm_output_bufs[(i + 1) % 2],
+                                        (__fp16 *) vtcm_activation, (__fp16 *) vtcm_weight_bufs[(i + 1) % 2],
+                                        vtcm_scales, hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS),
+                                        hmx_ceil_div(n_cols_p1, HMX_FP16_TILE_N_COLS), k / HMX_FP16_TILE_N_ROWS);
+                    hmx_queue_push(ctx->hmx_queue, hmx_queue_make_desc(hmx_matmul_worker_fn, &job_slots[(i + 1) % 2]));
                 }
 
-                // compute D_{i}
+                // D_i: store output (multi-thread HVX, parallel with C_{i+1})
                 float *output_chunk = dst + (mr * n + nc);
                 transfer_output_chunk_threaded(ctx, output_chunk, vtcm_output_bufs[i % 2], n_rows, n_cols, n);
 
-                // wait for DMA (A_{i+2}), compute B_{i+2}
+                // B_{i+2}: DMA pop + dequant (multi-thread HVX, parallel with C_{i+1})
                 if (i + 2 < n_chunk_cnt) {
                     dma_queue_pop(ctx->dma[0]);
                     dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[(i + 2) % 2], vtcm_qweight, n_cols_p2, k, row_stride, weight_type);
                 }
             }
         }
-    }
 
-    HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
+        hmx_queue_suspend(ctx->hmx_queue);
+    }
 
     TIMER_STOP(total);
 
@@ -1437,29 +1512,36 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
 }
 
 // C += AB
-void core_mma_chunk_fp16(__fp16 *c, const __fp16 *a, const __fp16 *b, const __fp16 *col_scales, const __fp16 *eye_tile,
+void core_mma_chunk_fp16(__fp16 *restrict c, const __fp16 *restrict a, const __fp16 *restrict b, const __fp16 *restrict col_scales, const __fp16 *restrict eye_tile,
                          int n_row_tiles, int n_col_tiles, int n_dot_tiles, bool zero_init) {
+    __builtin_assume(n_row_tiles > 0);
+    __builtin_assume(n_col_tiles > 0);
+    __builtin_assume(n_dot_tiles > 0);
 
-    hmx_set_output_scales(col_scales);
+    Q6_bias_mxmem2_A((void *)col_scales);
 
-    for (int i = 0; i < n_row_tiles; ++i) {
-        for (int j = 0; j < n_col_tiles; ++j) {
+    const size_t dot_tile_stride = n_dot_tiles * HMX_FP16_TILE_N_ELMS;
+    for (size_t i = 0; i < n_row_tiles; ++i) {
+        const __fp16 *row_base = a + i * dot_tile_stride;
+        __fp16 *res_base = c + i * n_col_tiles * HMX_FP16_TILE_N_ELMS;
+        for (size_t j = 0; j < n_col_tiles; ++j) {
             Q6_mxclracc_hf();
 
-            const __fp16 *row_tiles = a + i * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
-            const __fp16 *col_tiles = b + j * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
-
-            __fp16 *accum_tile = c + (i * n_col_tiles + j) * HMX_FP16_TILE_N_ELMS;
+            const __fp16 *col_tiles = b + j * dot_tile_stride;
+            const __fp16 *row_tiles = row_base;
+            __fp16 *accum_tile = res_base + j * HMX_FP16_TILE_N_ELMS;
             if (!zero_init) {
-                hmx_load_tile_pair_fp16(accum_tile, eye_tile);
+                Q6_activation_hf_mxmem_RR((unsigned int)accum_tile, 2047);
+                Q6_weight_hf_mxmem_RR((unsigned int)eye_tile, 2047);
             }
 
             for (int k = 0; k < n_dot_tiles; ++k) {
-                int offset = k * HMX_FP16_TILE_N_ELMS;
-                hmx_load_tile_pair_fp16(row_tiles + offset, col_tiles + offset);
+                Q6_activation_hf_mxmem_RR((unsigned int)row_tiles, 2047);
+                Q6_weight_hf_mxmem_RR((unsigned int)col_tiles, 2047);
+                row_tiles += HMX_FP16_TILE_N_ELMS;
+                col_tiles += HMX_FP16_TILE_N_ELMS;
             }
-
-            hmx_consume_accumulator_fp16(accum_tile);
+            Q6_mxmem_AR_after_hf(accum_tile, 0);
         }
     }
 }
@@ -1533,27 +1615,51 @@ void transfer_activation_chunk_threaded(struct htp_context *ctx, __fp16 *dst, co
     worker_pool_run_func(ctx->worker_pool, transfer_activation_chunk_worker_fn, &state, ctx->n_threads);
 }
 
-int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict out, const float *restrict x, const uint8_t *restrict w, int m,
-                                       int k, int n, int weight_type) {
-    // Runtime check -- k >= 16384 exceeds 2D DMA limit
-    if (k >= 16384) {
-        FARF(HIGH, "%s: k=%d exceeds 2D DMA limit", __func__, k);
-        return -1;
-    }
+int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict out, const float *restrict x, const uint8_t *restrict w,
+                                       int m, int k, int n, int weight_type) {
     // assume k % 32 == 0 && n % 32 == 0
     const size_t row_stride = get_x4x2_row_stride(weight_type, k);
     if (row_stride == 0) {
         return -1;
     }
 
-    const size_t vtcm_budget = ctx->vtcm_scratch_size;
+    const size_t vtcm_budget = ctx->vtcm_size;
 
-    const size_t M_BLOCK_SIZE = 512;
-    const size_t N_BLOCK_SIZE = 512;
-    const size_t K_BLOCK_SIZE = 512;
+    const size_t K_BLOCK_SIZE = 1024;
 
-    // Compute precise buffer sizes
+    // Fallback: if k doesn't need K-blocking, out-stationary has no advantage
+    const size_t k_iters_check = (k + K_BLOCK_SIZE - 1) / K_BLOCK_SIZE;
+    if (k_iters_check <= 1) {
+        FARF(MEDIUM, "%s: K_BLK=%zu >= k=%d, fallback to standard path", __func__, K_BLOCK_SIZE, k);
+        return FALLBACK_TO_STANDARD;
+    }
+
+    // Dynamic M,N search via hmx_compute_chunks
     const size_t sub_row_stride_alloc = get_x4x2_row_stride(weight_type, K_BLOCK_SIZE);
+    const size_t per_m                = K_BLOCK_SIZE * sizeof(float)  // scratch1: M×K×4 (act DMA staging F32)
+                         + K_BLOCK_SIZE * sizeof(__fp16);             // activation: M×K×2 (F16 tiles)
+    const size_t per_n = sub_row_stride_alloc                         // scratch0: N×sub_row(K) (packed quant)
+                         + K_BLOCK_SIZE * sizeof(__fp16);             // weight: N×K×2 (F16 tiles)
+    const size_t per_mn       = sizeof(__fp16);                       // output: M×N×2 (out-stationary)
+    // Alignment margin: hex_align_up can add up to 2047 bytes per buffer;
+    // scratch1 (mc×6144) is naturally 2048-aligned, remaining 4 buffers need margin
+    const size_t align_margin = 4 * HMX_FP16_TILE_SIZE;
+    const size_t overhead     = HMX_FP16_TILE_SIZE + 256 + align_margin;  // eye_tile + scales + alignment
+
+    size_t       M_BLOCK_SIZE, N_BLOCK_SIZE, vtcm_used;
+    // Cost-based search: minimize ceil(m/mc)*m_block_cost + ceil(n/nc)*n_block_cost.
+    // From profiling: wt_dequant per element ≈ 1.5× activation load per element.
+    // m_block_cost = n*3: each extra M-block re-dequants all N×K weight (expensive).
+    // n_block_cost = m*2: each extra N-block re-loads all M×K activation (cheaper).
+    const size_t m_block_cost = (size_t) n * 3;
+    const size_t n_block_cost = (size_t) m * 2;
+    if (hmx_compute_chunks(vtcm_budget, overhead, per_n, per_m, per_mn, m, n, m_block_cost, n_block_cost, &M_BLOCK_SIZE,
+                           &N_BLOCK_SIZE, &vtcm_used) != 0) {
+        FARF(HIGH, "%s: VTCM too small (m=%d k=%d n=%d budget=%zu)", __func__, m, k, n, vtcm_budget);
+        return -1;
+    }
+
+    // Compute precise buffer sizes from searched M,N and fixed K
     const size_t weight_size  = hex_align_up(N_BLOCK_SIZE * K_BLOCK_SIZE * sizeof(__fp16), HMX_FP16_TILE_SIZE);
     const size_t act_size     = hex_align_up(M_BLOCK_SIZE * K_BLOCK_SIZE * sizeof(__fp16), HMX_FP16_TILE_SIZE);
     const size_t out_size     = hex_align_up(M_BLOCK_SIZE * N_BLOCK_SIZE * sizeof(__fp16), HMX_FP16_TILE_SIZE);
@@ -1562,7 +1668,8 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
 
     const size_t total_vtcm = weight_size + act_size + out_size + scratch0_sz + scratch1_sz + HMX_FP16_TILE_SIZE + 256;
     if (total_vtcm > vtcm_budget) {
-        FARF(HIGH, "%s: VTCM too small: need %zu have %zu (m=%d k=%d n=%d)", __func__, total_vtcm, vtcm_budget, m, k, n);
+        FARF(HIGH, "%s: VTCM overflow after search: need %zu have %zu (M=%zu N=%zu K=%zu)", __func__, total_vtcm,
+             vtcm_budget, M_BLOCK_SIZE, N_BLOCK_SIZE, K_BLOCK_SIZE);
         return -1;
     }
 
@@ -1576,9 +1683,8 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
     __fp16  *vtcm_scales     = (__fp16 *) vtcm_seq_alloc(&vtcm_ptr, 256);
     assert((size_t)(vtcm_ptr - (uint8_t *)ctx->vtcm_base) <= vtcm_budget);
 
-    FARF(MEDIUM, "%s: m=%d k=%d n=%d wtype=%d vtcm=%zu/%zu",
-         __func__, m, k, n, weight_type,
-         (size_t)(vtcm_ptr - (uint8_t *)ctx->vtcm_base), vtcm_budget);
+    FARF(HIGH, "hmx-mm: m=%d k=%d n=%d wtype=%d block M=%zu N=%zu K=%zu vtcm=%zu/%zu", __func__, m, k, n, weight_type,
+         M_BLOCK_SIZE, N_BLOCK_SIZE, K_BLOCK_SIZE, (size_t) (vtcm_ptr - (uint8_t *) ctx->vtcm_base), vtcm_budget);
 
     // initialize eye tile (32x32 identity matrix)
     {
@@ -1592,7 +1698,7 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
             v = Q6_V_vror_VR(v, VLEN - 8);
         }
     }
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     TIMER_DEFINE(fetch);
     TIMER_DEFINE(act_load);
@@ -1610,7 +1716,7 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
             const int n_col_tiles = hmx_ceil_div(n_blk_sz, HMX_FP16_TILE_N_COLS);
 
             for (size_t kk = 0; kk < k; kk += K_BLOCK_SIZE) {
-                size_t k_blk_sz = hex_smin(k - kk, K_BLOCK_SIZE);
+                const size_t k_blk_sz = hex_smin(k - kk, K_BLOCK_SIZE);
 
                 TIMER_START(fetch);
                 // fetch activation block into VTCM
@@ -1626,13 +1732,13 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                 }
 
                 // fetch weight block into VTCM (x4x2 sub-block: quants + scales)
+                const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
                 {
                     qweight_fetch_task_state_t s;
 
                     const int blk_start = kk / QK_Q4_0x4x2;
                     const int nb_sub = (k_blk_sz + QK_Q4_0x4x2 - 1) / QK_Q4_0x4x2;
                     const int    full_qrow      = (weight_type == HTP_TYPE_Q8_0) ? k : (k / 2);
-                    const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
                     const int    scale_blk_size =
                         (weight_type == HTP_TYPE_MXFP4) ? HMX_X4X2_MXFP4_EBLK_SIZE : HMX_X4X2_DBLK_SIZE;
 
@@ -1672,7 +1778,6 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                     dma_queue_pop(ctx->dma[0]);
                     // vtcm_scratch0 is used to store the qweight chunk
                     // worker_pool_run_func already returned, so fetch is done
-                    const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
                     dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight, vtcm_scratch0,
                                                                 n_blk_sz, k_blk_sz, sub_row_stride, weight_type);
                 }