ggml-cpu: Optimized x86 and generic cpu q1_0 dot (follow up) (#21636)

* Implemented optimized q1_0 dot for x86 and generic

* Removed redundant helper definition

* Removed two redundant instructions from AVX q1_0 dot

* Fixed inconsistency with fp16 conversion for generic q1_0 dot and deduplicated generic fallback

* Style cleanup around AVX q1_0 dot

* Replaced explicitly unrolled blocks with inner for loop for q1_0

* Replaced scalar ARM q1_0 impl with new generic one

.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-cross.yml

@@ -246,6 +246,7 @@ jobs:
           apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \
+                  spirv-headers \
                   gcc-14-loongarch64-linux-gnu \
                   g++-14-loongarch64-linux-gnu \
                   libvulkan-dev:loong64

.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/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

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"
@@ -291,7 +292,7 @@ static bool common_params_handle_remote_preset(common_params & params, llama_exa
         hf_tag = "default";
     }
 
-    std::string model_endpoint = get_model_endpoint();
+    std::string model_endpoint = common_get_model_endpoint();
     auto preset_url = model_endpoint + hf_repo + "/resolve/main/preset.ini";
 
     // prepare local path for caching
@@ -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);
         }
     ));
@@ -1315,13 +1316,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
-        {"--clear-idle"},
-        {"--no-clear-idle"},
+        {"--cache-idle-slots"},
+        {"--no-cache-idle-slots"},
         "save and clear idle slots on new task (default: enabled, requires unified KV and cache-ram)",
         [](common_params & params, bool value) {
-            params.clear_idle = value;
+            params.cache_idle_slots = value;
         }
-    ).set_env("LLAMA_ARG_CLEAR_IDLE").set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_env("LLAMA_ARG_CACHE_IDLE_SLOTS").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},

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

@@ -198,10 +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()) {
@@ -434,14 +443,14 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     if (!format.per_call_start.empty()) {
         auto wrapped_call = format.per_call_start + p.space() + tool_choice + p.space() + format.per_call_end;
         if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call) + p.space());
         } else {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call);
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.space());
         }
         if (!format.section_start.empty()) {
             tool_calls = p.trigger_rule("tool-calls",
                                         p.literal(format.section_start) + p.space() + tool_calls + p.space() +
-                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
+                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end) + p.space()));
         }
     } else {
         std::string separator = ", ";  // Default

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

@@ -2334,7 +2334,7 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
         ? input
         : params.generation_prompt + input;
 
-    LOG_DBG("Parsing PEG input with format %s: %s\n", common_chat_format_name(params.format), effective_input.c_str());
+    //LOG_DBG("Parsing PEG input with format %s: %s\n", common_chat_format_name(params.format), effective_input.c_str());
 
     common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_LENIENT;
     if (params.debug) {

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) {
@@ -1381,7 +1382,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
 common_init_result::~common_init_result() = default;
 
-std::string get_model_endpoint() {
+std::string common_get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
     // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
     const char * hf_endpoint_env = getenv("HF_ENDPOINT");
@@ -1396,6 +1397,42 @@ std::string get_model_endpoint() {
     return model_endpoint;
 }
 
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
+    auto * mem = llama_get_memory(ctx);
+    if (mem == nullptr) {
+        return COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+    }
+
+    common_context_seq_rm_type res = COMMON_CONTEXT_SEQ_RM_TYPE_PART;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx);
+
+    return res;
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;

common/common.h

@@ -2,15 +2,15 @@
 
 #pragma once
 
+#include "llama-cpp.h"
+
 #include "ggml-opt.h"
 #include "ggml.h"
-#include "llama-cpp.h"
 
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
-#include <variant>
 #include <vector>
 #include <map>
 
@@ -27,11 +27,6 @@
 #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 +48,6 @@ 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;
 
 //
@@ -315,15 +302,15 @@ struct common_params_speculative {
     // general-purpose speculative decoding parameters
 
     int32_t n_max   = 16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min   = 0; // minimum number of draft tokens to use for speculative decoding
+    int32_t n_min   = 0;  // minimum number of draft tokens to use for speculative decoding
     float   p_split = 0.1f; // speculative decoding split probability
     float   p_min   = 0.75f; // minimum speculative decoding probability (greedy)
 
     // ngram-based speculative decoding
 
-    uint16_t ngram_size_n     = 12; // ngram size for lookup
-    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
+    uint16_t ngram_size_n   = 12; // ngram size for lookup
+    uint16_t ngram_size_m   = 48; // mgram size for speculative tokens
+    uint16_t ngram_min_hits = 1; // minimum hits at ngram/mgram lookup for mgram to be proposed
 
     std::shared_ptr<common_ngram_mod> ngram_mod;
 
@@ -579,7 +566,7 @@ struct common_params {
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
     bool    cache_prompt        = true;  // whether to enable prompt caching
-    bool    clear_idle          = true;  // save and clear idle slots upon starting a new task
+    bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
     int32_t checkpoint_every_nt = 8192;  // make a checkpoint every n tokens during prefill
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
@@ -859,7 +846,23 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
-std::string                   get_model_endpoint();
+// model endpoint from env
+std::string common_get_model_endpoint();
+
+//
+// Context utils
+//
+
+enum common_context_seq_rm_type {
+    COMMON_CONTEXT_SEQ_RM_TYPE_NO   = 0, // seq_rm not supported (e.g. no memory module)
+    COMMON_CONTEXT_SEQ_RM_TYPE_PART = 1, // can seq_rm partial sequences
+    COMMON_CONTEXT_SEQ_RM_TYPE_FULL = 2, // can seq_rm full sequences only
+};
+
+// check if the llama_context can remove sequences
+// note: clears the memory of the context
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
+
 
 //
 // Batch utils

common/download.cpp

@@ -1,5 +1,6 @@
 #include "arg.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "download.h"
@@ -303,7 +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);
@@ -441,7 +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/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"}
     };
 
@@ -229,7 +230,7 @@ static nl::json api_get(const std::string & url,
 static std::string get_repo_commit(const std::string & repo_id,
                                    const std::string & token) {
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/refs", token);
 
         if (!json.is_object() ||
@@ -307,7 +308,7 @@ hf_files get_repo_files(const std::string & repo_id,
     hf_files files;
 
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/tree/" + commit + "?recursive=true", token);
 
         if (!json.is_array()) {

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/ngram-map.cpp

@@ -208,7 +208,7 @@ void common_ngram_map_begin(
                 count_keys, count_keys_del, count_values_del, count_map_entries_upd);
     }
 
-    map.idx_last_check = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
+    map.idx_last_check = size_begin;
     map.size_last_begin = size_begin;
 }
 
@@ -231,7 +231,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
     }
 
-    if (map.idx_last_check  > cur_len) {
+    if (map.idx_last_check > cur_len) {
         // Should not happen because of common_ngram_map_begin().
         GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
     }
@@ -386,7 +386,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         LOG_DBG("%s: key_idx = %zu, key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
                 curr_key.key_idx, key_offset, curr_key.key_num, draft.size());
 
-        map.last_draft_created   = false;
+        map.last_draft_created   = true;
         map.last_draft_key_idx   = key_offset;
         map.last_draft_value_idx = 0; // value 0 is used for simple mode
         return;
@@ -524,7 +524,7 @@ void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) {
     struct common_ngram_map_value & curr_value = curr_key.values[val_idx]; // value used for draft generation.
 
     // update the value statistics
-    LOG_INF("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
+    LOG_DBG("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
             n_accepted, curr_value.n_accepted);
     curr_value.n_accepted = n_accepted;
 }

common/speculative.cpp

@@ -13,6 +13,7 @@
 #include <cstring>
 #include <iomanip>
 #include <map>
+#include <cinttypes>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -144,10 +145,28 @@ struct common_speculative_state {
     virtual void accept(uint16_t n_accepted) = 0;
 };
 
+struct common_speculative_checkpoint {
+    llama_pos pos_min  = 0;
+    llama_pos pos_max  = 0;
+
+    int64_t   n_tokens = 0;
+
+    std::vector<uint8_t> data;
+
+    size_t size() const {
+        return data.size();
+    }
+
+    size_t ckpt_size   = 0;
+};
+
 struct common_speculative_state_draft : public common_speculative_state {
     llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
     llama_context * ctx_dft;
 
+    bool use_ckpt = false;
+    struct common_speculative_checkpoint ckpt;
+
     common_sampler * smpl;
 
     llama_batch  batch;
@@ -160,10 +179,12 @@ struct common_speculative_state_draft : public common_speculative_state {
             enum common_speculative_type type,
             llama_context * ctx_tgt,
             llama_context * ctx_dft,
-            const std::vector<std::pair<std::string, std::string>> & replacements)
+            const std::vector<std::pair<std::string, std::string>> & replacements,
+            bool use_ckpt)
         : common_speculative_state(type)
         , ctx_tgt(ctx_tgt)
         , ctx_dft(ctx_dft)
+        , use_ckpt(use_ckpt)
     {
         batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
         smpl = nullptr;
@@ -218,7 +239,48 @@ struct common_speculative_state_draft : public common_speculative_state {
     }
 
     void begin(const llama_tokens & prompt) override {
-        GGML_UNUSED(prompt);
+        if (use_ckpt && ckpt.size() > 0) {
+            // delete checkpoint
+            LOG_DBG("%s: delete checkpoint, prompt.size=%zu, pos_min=%d, pos_max=%d, n_tokens=%" PRId64 ", size=%.3f MiB\n",
+                    __func__, prompt.size(), ckpt.pos_min, ckpt.pos_max, ckpt.n_tokens, (float) ckpt.data.size() / 1024 / 1024);
+            ckpt.pos_min   = 0;
+            ckpt.pos_max   = 0;
+            ckpt.n_tokens  = 0;
+            ckpt.ckpt_size = 0;
+            ckpt.data.clear();
+        }
+    }
+
+    size_t draft_create_checkpoint(int n_tokens_prompt, int n_tokens_batch) {
+        int slot_id = 0;
+        const size_t checkpoint_size = llama_state_seq_get_size_ext(ctx_dft, slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+
+        ckpt.pos_min  = llama_memory_seq_pos_min(llama_get_memory(ctx_dft), slot_id);
+        ckpt.pos_max  = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), slot_id);
+        ckpt.n_tokens = n_tokens_prompt - n_tokens_batch;
+        ckpt.data.resize(checkpoint_size);
+
+        const size_t n = llama_state_seq_get_data_ext(ctx_dft, ckpt.data.data(), checkpoint_size, slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        if (n != checkpoint_size) {
+            GGML_ABORT("checkpoint size mismatch: expected %zu, got %zu\n", checkpoint_size, n);
+        }
+
+        LOG_DBG("%s: pos_min = %d, pos_max = %d, size = %.3f MiB\n", __func__,
+                ckpt.pos_min, ckpt.pos_max, (float) ckpt.data.size() / 1024 / 1024);
+        return n;
+    }
+
+    size_t draft_restore_checkpoint(size_t ckpt_size_part_expected) {
+        int slot_id = 0;
+        LOG_DBG("%s: pos_min = %d, pos_max = %d\n", __func__, ckpt.pos_min, ckpt.pos_max);
+        const size_t n = llama_state_seq_set_data_ext(ctx_dft, ckpt.data.data(), ckpt.size(), slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        if (n != ckpt_size_part_expected) {
+            GGML_ABORT("%s: failed to restore context checkpoint (pos_min=%d, pos_max=%d, size=%zu, get_data_ext->%zu, set_data_ext->%zu",
+                        __func__, ckpt.pos_min, ckpt.pos_max, ckpt.size(), ckpt_size_part_expected, n);
+        }
+        llama_memory_seq_rm(llama_get_memory(ctx_dft), slot_id, ckpt.pos_max + 1, -1);
+
+        return n;
     }
 
     void draft(
@@ -236,8 +298,8 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         auto * mem_dft = llama_get_memory(ctx_dft);
 
-        int reuse_i = 0;
-        int reuse_n = 0;
+        int reuse_i = 0; // index of part to be reused in prompt_dft
+        int reuse_n = 0; // length of part to be reused in prompt_dft
 
         const int n_ctx = llama_n_ctx(ctx_dft) - params.n_max;
 
@@ -287,18 +349,26 @@ struct common_speculative_state_draft : public common_speculative_state {
             }
         }
 
-        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, #prompt_dft = %zu, #prompt_cur = %zu\n",
+                __func__, reuse_i, reuse_n, prompt_dft.size(), prompt_cur.size());
+        if (use_ckpt && ckpt.ckpt_size == 0 && reuse_n > 0) {
+            LOG_DBG("%s: no checkpoint available, no reuse, (reuse_i=%d, reuse_n=%d) -> (0, 0)\n",
+                    __func__, reuse_i, reuse_n);
+            reuse_i = 0;
+            reuse_n = 0;
+        }
 
         result.clear();
         result.reserve(params.n_max);
 
-        if (reuse_n == 0) {
+        bool needs_ckpt = use_ckpt && prompt_dft.size() > 0;
+        if (reuse_n == 0 || (use_ckpt && reuse_i > 0)) {
             llama_memory_clear(mem_dft, false);
             prompt_dft.clear();
         } else {
             // this happens when a previous draft has been discarded (for example, due to being too small), but the
             // target model agreed with it. in this case, we simply pass back the previous results to save compute
-            if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
+            if (reuse_i + reuse_n < (int64_t) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
                 for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
                     result.push_back(prompt_dft[i]);
 
@@ -310,19 +380,50 @@ struct common_speculative_state_draft : public common_speculative_state {
                 return;
             }
 
+            bool do_restore = false;
+            if (prompt_dft.size() > prompt_cur.size() && reuse_i + reuse_n < (int64_t) prompt_dft.size()) {
+                // This can happen after a partial acceptance (speculative decoding with checkpoints)
+                LOG_DBG("%s: #prompt_dft=%zu, #prompt_cur=%zu, shorten draft\n",
+                        __func__, prompt_dft.size(), prompt_cur.size());
+                prompt_dft.resize(prompt_cur.size());
+                do_restore = true;
+            }
+
             if (reuse_i > 0) {
-                llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                bool is_removed = llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                if (!is_removed) {
+                    LOG_ERR("%s: llama_memory_seq_rm failed, reuse_i=%d\n", __func__, reuse_i);
+                }
                 llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
 
                 prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
             }
 
-            if (reuse_n < (int) prompt_dft.size()) {
-                llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
-                prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+            if (reuse_n < (int) prompt_dft.size() || do_restore) {
+                if (use_ckpt) {
+                    if (ckpt.n_tokens > (int64_t) prompt_dft.size()) {
+                        LOG_INF("%s: checkpoint is too large, prompt_tgt.size=%zu, ckpt.n_tokens=%" PRId64 ", reuse_n=%d, prompt_dft.size=%zu\n",
+                                __func__, prompt_tgt.size(), ckpt.n_tokens, reuse_n, prompt_dft.size());
+                    }
+                    draft_restore_checkpoint(ckpt.ckpt_size);
+                    reuse_n = ckpt.n_tokens;
+                    prompt_dft.resize(reuse_n);
+                    needs_ckpt = false;
+                } else {
+                    bool is_removed = llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
+                    if (!is_removed) {
+                        LOG_ERR("%s: llama_memory_seq_rm failed, reuse_n=%d, prompt_dft.size=%zu\n",
+                                __func__, reuse_n, prompt_dft.size());
+                    }
+                    prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+                }
             }
         }
 
+        if (needs_ckpt) {
+            ckpt.ckpt_size = draft_create_checkpoint(prompt_dft.size(), batch.n_tokens);
+        }
+
         // prepare a batch to evaluate any new tokens in the prompt
         common_batch_clear(batch);
 
@@ -337,7 +438,11 @@ struct common_speculative_state_draft : public common_speculative_state {
         if (batch.n_tokens > 0) {
             //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
 
-            llama_decode(ctx_dft, batch);
+            int ret = llama_decode(ctx_dft, batch);
+            if (ret != 0 && ret != 1) {
+                LOG_WRN("%s: llama_decode returned %d, prompt_cur.size=%zu\n",
+                        __func__, ret, prompt_cur.size());
+            }
         }
 
         const llama_pos n_past = prompt_dft.size();
@@ -351,7 +456,11 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
 
-        llama_decode(ctx_dft, batch);
+        int ret = llama_decode(ctx_dft, batch);
+        if (ret != 0 && ret != 1) {
+            LOG_WRN("%s: llama_decode returned %d, prompt_cur.size=%zu, prompt_dft.size=%zu\n",
+                    __func__, ret, prompt_cur.size(), prompt_dft.size());
+        }
 
         common_sampler_reset(smpl);
 
@@ -387,7 +496,11 @@ struct common_speculative_state_draft : public common_speculative_state {
             common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
 
             // evaluate the drafted tokens on the draft model
-            llama_decode(ctx_dft, batch);
+            ret = llama_decode(ctx_dft, batch);
+            if (ret != 0) {
+                LOG_WRN("%s: llama_decode[%d] returned %d, prompt_cur.size=%zu, prompt_dft.size=%zu\n",
+                        __func__, i, ret, prompt_cur.size(), prompt_dft.size());
+            }
 
             prompt_dft.push_back(id);
         }
@@ -739,6 +852,7 @@ struct common_speculative_state_ngram_cache : public common_speculative_state {
 
 struct common_speculative {
     std::vector<std::unique_ptr<common_speculative_state>> impls; // list of implementations to use and their states
+
     common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats)
 };
 
@@ -798,42 +912,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
     return it->second;
 }
 
-bool common_speculative_is_compat(llama_context * ctx_tgt) {
-    auto * mem = llama_get_memory(ctx_tgt);
-    if (mem == nullptr) {
-        return false;
-    }
-
-    bool res = true;
-
-    llama_memory_clear(mem, true);
-
-    // eval 2 tokens to check if the context is compatible
-    std::vector<llama_token> tmp;
-    tmp.push_back(0);
-    tmp.push_back(0);
-
-    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
-    if (ret != 0) {
-        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
-        res = false;
-        goto done;
-    }
-
-    // try to remove the last tokens
-    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
-        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
-        res = false;
-        goto done;
-    }
-
-done:
-    llama_memory_clear(mem, true);
-    llama_synchronize(ctx_tgt);
-
-    return res;
-}
-
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
@@ -908,10 +986,13 @@ common_speculative * common_speculative_init(
             case COMMON_SPECULATIVE_TYPE_NONE:
                 break;
             case COMMON_SPECULATIVE_TYPE_DRAFT: {
+                const bool use_ckpt = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+
                 impls.push_back(std::make_unique<common_speculative_state_draft>(config.type,
                     /* .ctx_tgt      = */ ctx_tgt,
                     /* .ctx_dft      = */ ctx_dft,
-                    /* .replacements = */ params.replacements
+                    /* .replacements = */ params.replacements,
+                    /* .use_ckpt     = */ use_ckpt
                 ));
                 break;
             }
@@ -966,7 +1047,8 @@ common_speculative * common_speculative_init(
     }
 
     auto * result = new common_speculative {
-        /* .impls = */ std::move(impls)
+        /* .impls     = */ std::move(impls),
+        /* .curr_impl = */ nullptr,
     };
 
     return result;

common/speculative.h

@@ -14,10 +14,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);
 
-// check if the llama_context is compatible for speculative decoding
-// note: clears the memory of the context
-bool common_speculative_is_compat(llama_context * ctx_tgt);
-
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt);
@@ -39,3 +35,9 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted);
 
 // print statistics about the speculative decoding
 void common_speculative_print_stats(const common_speculative * spec);
+
+struct common_speculative_deleter {
+    void operator()(common_speculative * s) { common_speculative_free(s); }
+};
+
+typedef std::unique_ptr<common_speculative, common_speculative_deleter> common_speculative_ptr;

convert_hf_to_gguf.py

@@ -1850,20 +1850,28 @@ class TextModel(ModelBase):
             with open(module_path, encoding="utf-8") as f:
                 modules = json.load(f)
             for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
+                if mod["type"].endswith("Pooling"):
                     pooling_path = mod["path"]
                     break
 
+        mode_mapping = {
+            "mean": gguf.PoolingType.MEAN,
+            "cls": gguf.PoolingType.CLS,
+            "lasttoken": gguf.PoolingType.LAST,
+        }
+
         # get pooling type
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
+            if pooling.get("pooling_mode_mean_tokens"):
                 pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
+            elif pooling.get("pooling_mode_cls_token"):
                 pooling_type = gguf.PoolingType.CLS
-            elif pooling["pooling_mode_lasttoken"]:
+            elif pooling.get("pooling_mode_lasttoken"):
                 pooling_type = gguf.PoolingType.LAST
+            elif (pooling_mode := pooling.get("pooling_mode")) in mode_mapping:
+                pooling_type = mode_mapping[pooling_mode]
             else:
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
@@ -7180,7 +7188,7 @@ class EmbeddingGemma(Gemma3Model):
                 with open(modules_file, encoding="utf-8") as modules_json_file:
                     mods = json.load(modules_json_file)
                 for mod in mods:
-                    if mod["type"] == "sentence_transformers.models.Dense":
+                    if mod["type"].endswith("Dense"):
                         mod_path = mod["path"]
                         # check if model.safetensors file for Dense layer exists
                         model_tensors_file = self.dir_model / mod_path / "model.safetensors"
@@ -10893,7 +10901,64 @@ 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

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/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

@@ -130,6 +130,23 @@ Note:
 - 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/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

@@ -1,11 +1,5 @@
 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
@@ -254,6 +248,7 @@ 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:
@@ -1421,23 +1456,51 @@ struct ggml_backend_meta_context {
     int                         max_nnodes    = 0;
     size_t                      max_tmp_size  = 0;
     size_t                      max_subgraphs = 0;
+    size_t                      n_subgraphs   = 0;
+    uint64_t                    uid           = 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);
         }
@@ -1555,6 +1618,9 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     const size_t n_backends = ggml_backend_meta_n_backends(backend);
     ggml_backend_meta_context * backend_ctx = (ggml_backend_meta_context *) backend->context;
 
+    // If the previous cgraph had a defined UID it can be used to skip rebuilding the subgraphs per simple backend.
+    const bool needs_rebuild = (cgraph->uid == 0) || (cgraph->uid != backend_ctx->uid);
+
     bool max_nnodes_raised = false;
     if (cgraph->n_nodes > backend_ctx->max_nnodes) {
         for (size_t j = 0; j < n_backends; j++) {
@@ -1564,173 +1630,181 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         backend_ctx->max_nnodes = cgraph->n_nodes;
         max_nnodes_raised = true;
+        assert(needs_rebuild);
     }
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
 
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
-                // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
-                bcj.nodes[i] = node;
-                continue;
+    if (needs_rebuild) {
+        size_t n_subgraphs  = 0;
+        size_t max_tmp_size = 0;
+
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
+                    // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
+                    bcj.nodes[i] = node;
+                    continue;
+                }
+                bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
+                GGML_ASSERT(bcj.nodes[i]);
             }
-            bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
-            GGML_ASSERT(bcj.nodes[i]);
         }
-    }
 
-    size_t n_subgraphs  = 0;
-    size_t max_tmp_size = 0;
-    {
-        // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
-        auto get_i_delayed = [&](const int i) -> int {
-            int id = i; // i_delayed
-            int idr = i; // i_delayed return, last safe return value
+        {
+            // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
+            auto get_i_delayed = [&](const int i) -> int {
+                int id = i; // i_delayed
+                int idr = i; // i_delayed return, last safe return value
 
-            ggml_tensor * node = cgraph->nodes[id];
-            int32_t n_used = ggml_node_get_use_count(cgraph, id);
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
-                        ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                ggml_tensor * node = cgraph->nodes[id];
+                int32_t n_used = ggml_node_get_use_count(cgraph, id);
+                if (id + 1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
+                            ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    }
+                }
+                if (id + 1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_MUL && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    }
+                }
+
+                if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                for (int32_t k = 0; k < n_used; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
+                            next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
+                            ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_MUL && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-
-            if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                for (int32_t k = 0; k < n_used - 2; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
+                    id++;
+                }
+                idr = id;
                 return idr;
-            }
-            for (int32_t k = 0; k < n_used; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
-                        next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
-                        ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            for (int32_t k = 0; k < n_used - 2; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            idr = id;
-            return idr;
-        };
+            };
 
-        int i_start = 0;
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                continue;
-            }
-            const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
-            if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
-                max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
-            }
-            const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
-            if (!new_subgraph) {
-                continue;
-            }
+            int i_start = 0;
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    continue;
+                }
+                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
+                if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                    max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
+                }
+                const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
+                if (!new_subgraph) {
+                    continue;
+                }
 
-            i = get_i_delayed(i);
+                i = get_i_delayed(i);
 
+                for (size_t j = 0; j < n_backends; j++) {
+                    auto & bcj = backend_ctx->backend_configs[j];
+                    bcj.cgraphs[n_subgraphs].offset = i_start;
+                }
+                n_subgraphs++;
+                i_start = i + 1;
+            }
+            GGML_ASSERT(i_start == cgraph->n_nodes);
+        }
+
+        backend_ctx->uid         = cgraph->uid;
+        backend_ctx->n_subgraphs = n_subgraphs;
+
+        if (max_tmp_size > backend_ctx->max_tmp_size) {
             for (size_t j = 0; j < n_backends; j++) {
                 auto & bcj = backend_ctx->backend_configs[j];
-                bcj.cgraphs[n_subgraphs].offset = i_start;
+                bcj.buf.reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
             }
-            n_subgraphs++;
-            i_start = i + 1;
+            backend_ctx->max_tmp_size = max_tmp_size;
         }
-        GGML_ASSERT(i_start == cgraph->n_nodes);
-    }
 
-    if (max_tmp_size > backend_ctx->max_tmp_size) {
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            bcj.buf.reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
-        }
-        backend_ctx->max_tmp_size = max_tmp_size;
-    }
-
-
-    if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
-        backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
-        const size_t n_reduce_steps = backend_ctx->n_reduce_steps();
-        const size_t n_nodes_per_device = 2 * n_reduce_steps; // tmp + ADD per step
-        const size_t n_cgraphs_per_device = n_reduce_steps;    // 1 ADD graph per step
-        const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
-        const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
-        const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
-        ggml_init_params params = {
-            /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        backend_ctx->ctx.reset(ggml_init(params));
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            for (size_t i = 0; i < n_subgraphs; i++) {
-                bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+        if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
+            backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
+            const size_t n_reduce_steps = backend_ctx->n_reduce_steps();
+            const size_t n_nodes_per_device = 2 * n_reduce_steps; // tmp + ADD per step
+            const size_t n_cgraphs_per_device = n_reduce_steps;    // 1 ADD graph per step
+            const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
+            const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
+            const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
+            ggml_init_params params = {
+                /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
+                /*.mem_buffer =*/ nullptr,
+                /*.no_alloc   =*/ true,
+            };
+            backend_ctx->ctx.reset(ggml_init(params));
+            for (size_t j = 0; j < n_backends; j++) {
+                auto & bcj = backend_ctx->backend_configs[j];
+                for (size_t i = 0; i < n_subgraphs; i++) {
+                    bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+                }
+            }
+            backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
+                backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
+            }
+            backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
+                backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
             }
         }
-        backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
-            backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
-        }
-        backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
-            backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
-        }
-    }
 
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
-        for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
-            ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
-            const size_t i_node_start = bcj.cgraphs[i_graph].offset;
-            const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
-            cgraph_ij->n_nodes = i_node_stop - i_node_start;
-            ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
-            for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
-                ggml_tensor * node_ij = bcj.nodes[i_node];
-                cgraph_ij->nodes[i_node - i_node_start] = node_ij;
-                const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
-                const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
-                cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
+                ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
+                const size_t i_node_start = bcj.cgraphs[i_graph].offset;
+                const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
+                cgraph_ij->n_nodes = i_node_stop - i_node_start;
+                ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
+                for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
+                    ggml_tensor * node_ij = bcj.nodes[i_node];
+                    cgraph_ij->nodes[i_node - i_node_start] = node_ij;
+                    const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
+                    const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
+                    cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+                }
+                cgraph_ij->uid = ggml_graph_next_uid();
             }
         }
     }
@@ -1837,7 +1911,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     };
 
 
-    for (size_t i = 0; i < n_subgraphs; i++) {
+    for (size_t i = 0; i < backend_ctx->n_subgraphs; i++) {
         for (size_t j = 0; j < n_backends; j++) {
             auto & bcj = backend_ctx->backend_configs[j];
             const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, bcj.cgraphs[i].cgraph_main);
@@ -1846,22 +1920,17 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
             }
         }
 
-        if (n_backends > 1 && i < n_subgraphs - 1) {
+        if (n_backends > 1 && i < backend_ctx->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-fallback.h

@@ -83,7 +83,6 @@
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
-#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4

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

@@ -151,8 +151,6 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q1_0 * GGML_RESTRICT x = vx;
     const block_q8_0 * GGML_RESTRICT y = vy;
 
-    float sumf = 0.0f;
-
 #if defined(__ARM_NEON)
     float32x4_t sumv = vdupq_n_f32(0.0f);
 
@@ -212,31 +210,13 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
     }
 
-    sumf = vaddvq_f32(sumv);
+    *s = vaddvq_f32(sumv);
 #else
-    // Scalar fallback
-    for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-
-        // Process 4 Q8_0 blocks
-        for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
-            int sumi = 0;
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
-
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi += xi * y[i*4 + k].qs[j];
-            }
-            sumf += d0 * d1 * sumi;
-        }
-    }
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
-    *s = sumf;
 }
 
 
@@ -783,6 +763,7 @@ 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 +775,40 @@ 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 +819,7 @@ 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/arch/x86/quants.c

@@ -274,6 +274,18 @@ static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const
 }
 #endif
 #elif defined(__SSSE3__)
+static inline __m128i bytes_from_bits_16(const uint8_t * x) {
+    uint16_t x16;
+    memcpy(&x16, x, sizeof(uint16_t));
+
+    const __m128i shuf_mask = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    __m128i bytes = _mm_shuffle_epi8(_mm_set1_epi16((short) x16), shuf_mask);
+    const __m128i bit_mask = _mm_set_epi64x(0x7fbfdfeff7fbfdfe, 0x7fbfdfeff7fbfdfe);
+    bytes = _mm_or_si128(bytes, bit_mask);
+
+    return _mm_cmpeq_epi8(bytes, _mm_set1_epi64x(-1));
+}
+
 // horizontally add 4x4 floats
 static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
     __m128 res_0 =_mm_hadd_ps(a, b);
@@ -540,6 +552,152 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK1_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q1_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+    const __m256i ones_8 = _mm256_set1_epi8(1);
+    const __m256i ones_16 = _mm256_set1_epi16(1);
+    const __m256i byte_shuf = _mm256_setr_epi8(
+            0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+            2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3);
+    const __m256i bit_masks = _mm256_setr_epi8(
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128,
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128);
+    const __m256i zero = _mm256_setzero_si256();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const uint32_t * GGML_RESTRICT qs32 = (const uint32_t *) x[ib].qs;
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+        __m256 acc_block;
+        {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[0].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[0]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), _mm256_cvtepi32_ps(s32));
+        }
+        for (int K = 1; K < 4; ++K) {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[K].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[K]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[K].d)), _mm256_cvtepi32_ps(s32), acc_block);
+        }
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d0), acc_block, acc);
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__AVX__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+        __m256 acc_block;
+        {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[0]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), q);
+        }
+        for(int K = 1; K < 4; ++K) {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[(K) * 4]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_add_ps(acc_block, _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(K)].d)), q));
+        }
+#undef Q1_AVX_BLOCK
+
+        acc = _mm256_add_ps(acc, _mm256_mul_ps(_mm256_set1_ps(d0), acc_block));
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__SSSE3__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const __m128 d0 = _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+#define Q1_SSSE3_BLOCK(QS_OFF, Y_IDX, ACC) \
+        { \
+            const __m128i bit_mask_0 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 0]); \
+            const __m128i bit_mask_1 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 2]); \
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[0]); \
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[16]); \
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero); \
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero); \
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0); \
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1); \
+            const __m128i sum_0 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_0), ones_16); \
+            const __m128i sum_1 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_1), ones_16); \
+            const __m128 q = _mm_cvtepi32_ps(_mm_add_epi32(sum_0, sum_1)); \
+            (ACC) = _mm_add_ps((ACC), _mm_mul_ps(_mm_mul_ps(d0, _mm_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(Y_IDX)].d))), q)); \
+        }
+        Q1_SSSE3_BLOCK(0,  0, acc_0)
+        Q1_SSSE3_BLOCK(4,  1, acc_1)
+        Q1_SSSE3_BLOCK(8,  2, acc_2)
+        Q1_SSSE3_BLOCK(12, 3, acc_3)
+#undef Q1_SSSE3_BLOCK
+    }
+
+    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

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/quants.c

@@ -137,22 +137,28 @@ void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     float sumf = 0.0;
 
     for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         float sumi = 0.0f;
 
         for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 4 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
             int sumi_block = 0;
 
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
+            const uint8_t * GGML_RESTRICT bits = &x[i].qs[k * 4];
+            const int8_t  * GGML_RESTRICT qy   = yb->qs;
 
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi_block += xi * y[i*4 + k].qs[j];
+            for (int b = 0; b < 4; ++b, qy += 8) {
+                const unsigned mask = bits[b];
+                sumi_block += ((mask & 0x01) ? qy[0] : -qy[0])
+                           +  ((mask & 0x02) ? qy[1] : -qy[1])
+                           +  ((mask & 0x04) ? qy[2] : -qy[2])
+                           +  ((mask & 0x08) ? qy[3] : -qy[3])
+                           +  ((mask & 0x10) ? qy[4] : -qy[4])
+                           +  ((mask & 0x20) ? qy[5] : -qy[5])
+                           +  ((mask & 0x40) ? qy[6] : -qy[6])
+                           +  ((mask & 0x80) ? qy[7] : -qy[7]);
             }
 
             sumi += d1 * sumi_block;

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/common.cuh

@@ -269,10 +269,6 @@ static const char * cu_get_error_str(CUresult err) {
 #define FLASH_ATTN_AVAILABLE
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
-#if defined(TURING_MMA_AVAILABLE)
-#define LDMATRIX_TRANS_AVAILABLE
-#endif // defined(TURING_MMA_AVAILABLE)
-
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
@@ -924,6 +920,13 @@ 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 +1095,6 @@ 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();
@@ -1183,6 +1182,8 @@ 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];
@@ -1364,12 +1365,28 @@ 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-mma-f16.cuh

@@ -305,12 +305,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     // K/V data is loaded with decreasing granularity for D for better memory bandwidth.
-    // The minimum granularity with cp.async is 16 bytes, with synchronous data loading it's 4 bytes.
+    // The minimum granularity is 16 bytes.
+    constexpr int h2_per_chunk = 16/sizeof(half2);
+    const int chunks_per_row = D2 / h2_per_chunk;
     if constexpr (use_cp_async) {
+        static_assert(warp_size == 32, "bad warp_size");
         static_assert(!oob_check, "OOB check not compatible with cp_async");
         constexpr int preload = 64;
-        constexpr int h2_per_chunk = 16/sizeof(half2);
-        const int chunks_per_row = D2 / h2_per_chunk;
 
         const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
 
@@ -348,11 +349,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         // 6: max  1*16= 16 bytes,   8 half
         ggml_cuda_unroll<6>{}(load);
     } else {
-        // TODO use ggml_cuda_memcpy_1
+        const half2 zero[4] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
         auto load = [&] __device__ (const int n) {
-            const int stride_k = warp_size >> n;
-            const int k0_start = stride_k == warp_size ? 0 : D2 - D2 % (2*stride_k);
-            const int k0_stop  =                             D2 - D2 % (1*stride_k);
+            const int stride_k = 32 >> n;
+            const int k0_start = stride_k == 32 ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
+            const int k0_stop  =                      chunks_per_row - chunks_per_row % (1*stride_k);
             const int stride_i = warp_size / stride_k;
 
             if (k0_start == k0_stop) {
@@ -371,15 +372,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
-                    tile_KV[i*stride_tile + k] = !oob_check || i < i_sup ? KV[i*stride_KV + k] : make_half2(0.0f, 0.0f);
+                    ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
+                        !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
                 }
             }
         };
-        // 1: max 32* 4=128 bytes,  64 half
-        // 2: max 16* 4= 64 bytes,  32 half
-        // 3: max  8* 4= 32 bytes,  16 half
-        // 4: max  4* 4= 16 bytes,   8 half
-        ggml_cuda_unroll<4>{}(load);
+        // 1: max 32*16=512 bytes, 256 half
+        // 2: max 16*16=256 bytes, 128 half
+        // 3: max  8*16=128 bytes,  64 half
+        // 4: max  4*16= 64 bytes,  32 half
+        // 5: max  2*16= 32 bytes,  16 half
+        // 6: max  1*16= 16 bytes,   8 half
+        ggml_cuda_unroll<6>{}(load);
     }
 }
 
@@ -862,11 +866,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 
-#if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-    T_A_VKQ A_identity;
-    make_identity_mat(A_identity);
-#endif // defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
     for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
@@ -897,29 +896,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
 
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
-#if defined(LDMATRIX_TRANS_AVAILABLE)
                 load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-#elif defined(AMD_MFMA_AVAILABLE)
-                // MFMA A register layout: A_mat[i=lane%16][k=4*(lane/16)+reg].
-                // Normal load gives A_mat[seq][dv] but we need A_mat[dv][seq] = V^T.
-                // Load with transposed addressing: 4 strided half loads.
-                {
-                    const half2 * xs0 = tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2;
-                    const half * xs0_h = (const half *) xs0;
-                    const int stride_h = stride_tile_V * 2; // stride in half units
-                    half * A_h = (half *) A.x;
-#pragma unroll
-                    for (int l = 0; l < 4; ++l) {
-                        A_h[l] = xs0_h[(4*(threadIdx.x / 16) + l) * stride_h + threadIdx.x % 16];
-                    }
-                }
-#else
-                // TODO: Try to transpose tile_V when loading gmem to smem.
-                // Use mma to transpose T_A_VKQ for RDNA.
-                T_A_VKQ A_trans;
-                load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-                mma(A, A_trans, A_identity);
-#endif // defined(LDMATRIX_TRANS_AVAILABLE)
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {

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;
@@ -4783,6 +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:
@@ -4820,6 +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:
@@ -5220,8 +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/mma.cuh

@@ -86,17 +86,12 @@ namespace ggml_cuda_mma {
     //   - (I_MAJOR, I_MAJOR_MIRRORED) -> I_MAJOR
     //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
 
-    static constexpr bool is_i_major(const data_layout dl) {
-        return dl == DATA_LAYOUT_I_MAJOR ||
-               dl == DATA_LAYOUT_I_MAJOR_MIRRORED;
-    }
-
     static constexpr __device__ data_layout get_input_data_layout() {
-#if defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
         return DATA_LAYOUT_I_MAJOR_MIRRORED;
 #else
         return DATA_LAYOUT_I_MAJOR;
-#endif // defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
@@ -113,7 +108,6 @@ namespace ggml_cuda_mma {
         T x[ne] = {0};
 
         static constexpr __device__ bool supported() {
-            if (I == 64 && J ==  2) return true;
             if (I == 16 && J ==  8) return true;
             if (I == 32 && J ==  4) return true;
             if (I == 16 && J == 16) return true;
@@ -122,7 +116,7 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+            if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 16;
             } else if constexpr (I == 16 && J == 8) {
                 return threadIdx.x % 16;
@@ -139,8 +133,8 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-                return (2 * ((threadIdx.x / 16) % 2) + l);
+            if constexpr (I == 16 && J == 4) {
+                return threadIdx.x / 16;
             } else if constexpr (I == 16 && J == 8) {
                 return 2 * (threadIdx.x / 16) + l;
             } else if constexpr (I == 32 && J == 4) {
@@ -154,7 +148,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#elif defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
@@ -283,7 +277,7 @@ namespace ggml_cuda_mma {
         static constexpr int         J  = J_;
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
@@ -407,7 +401,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     };
 
     template <int I_, int J_>
@@ -701,57 +695,12 @@ namespace ggml_cuda_mma {
     }
 #endif // defined(TURING_MMA_AVAILABLE)
 
-    static __device__ __forceinline__ void make_identity_mat(tile<16, 8, half2> & t) {
-#if defined(RDNA4)
-        const int row = t.get_i(0);
-        const int left_right = t.get_j(0) / 4;
-        const int up_down = row / 8;
-        const int idx = row % 8;
-        reinterpret_cast<half*>(t.x)[idx] = left_right == up_down ? 1.0f : 0.0f;
-#else
-        GGML_UNUSED_VARS(t);
-        NO_DEVICE_CODE;
-#endif // defined(RDNA4)
-    }
-
     template <int I, int J, typename T, data_layout dl>
     static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
-#if defined(AMD_MFMA_AVAILABLE)
-        if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        } else {
-            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-        }
-#elif defined(AMD_WMMA_AVAILABLE)
-        // All wmma layout has contiguous data when i-major.
-        if constexpr (is_i_major(dl)) {
-            // the data must be aligned to 16 bytes when bigger than ggml_cuda_get_max_cpy_bytes()
-            constexpr int aligned_copy_bytes = ggml_cuda_get_max_cpy_bytes();
-            if constexpr (sizeof(t.x) > aligned_copy_bytes) {
-                static_assert(sizeof(t.x) % aligned_copy_bytes == 0, "bad type size");
-                constexpr int aligned_copy_count = sizeof(t.x)/aligned_copy_bytes;
-#pragma unroll
-                for (int i = 0; i < aligned_copy_count; ++i) {
-                    ggml_cuda_memcpy_1<aligned_copy_bytes>(t.x + t.ne/aligned_copy_count*i, xs0 + t.get_i(0) * stride + t.get_j(t.ne/aligned_copy_count*i));
-                }
-            } else {
-                ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-            }
-        } else {
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        }
-#else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
             t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
         }
-#endif // defined(AMD_MFMA_AVAILABLE)
     }
 
     template <typename T>
@@ -764,26 +713,37 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(t, xs0, stride);
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
-    template <typename T>
+    template <typename T, data_layout dl>
     static __device__ __forceinline__ void load_ldmatrix(
-            tile<16, 4, T> & t, const T * __restrict__ xs0, const int stride) {
+            tile<16, 4, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
         int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride;
         asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<8>(t.x + 2, xs0 + t.get_i(0)*stride + 2);
+#else
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 4, "bad ne");
+        ggml_cuda_memcpy_1<4>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -796,19 +756,26 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
-#else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-#if 1
-        // TODO: more generic handling
-        static_assert(sizeof(T) == 4, "bad type size");
+#elif defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 32, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<16>(t.x + 4, xs0 + t.get_i(0)*stride + 4);
 #else
-        load_generic(t, xs0, stride);
-#endif // 1
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -827,23 +794,30 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void load_ldmatrix(
             tile<32, 4, half2> & t, const half2 * __restrict__ xs0, const int stride) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
-        int * xi = (int * ) t.x;
+        int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
             : "l"(xs));
+#elif defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        half * xh = (half *) t.x;
+#pragma unroll
+        for (int l = 0; l < t.ne; ++l) {
+            xh[2*l + 0] = ((const half *) xs0)[(2*t.get_j(l) + 0)*(2*stride) + t.get_i(l)];
+            xh[2*l + 1] = ((const half *) xs0)[(2*t.get_j(l) + 1)*(2*stride) + t.get_i(l)];
+        }
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
@@ -1218,73 +1192,27 @@ namespace ggml_cuda_mma {
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * acc = (int32x4_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0],
-                                                      B.x[0],
-                                                      acc[0],
-                                                      0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1],
-                                                      B.x[1],
-                                                      acc[0],
-                                                      0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
-
 #elif defined(AMD_WMMA_AVAILABLE)
-
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
-
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
-
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[1], true, b_vec[1], acc[0], true);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[1], true, b_vec[1], acc[0], true);
 #endif // RDNA4
-
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -1297,19 +1225,10 @@ namespace ggml_cuda_mma {
         using int32x16_t = __attribute__((__vector_size__(16 * sizeof(int)))) int;
         int32x16_t * acc = (int32x16_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0],
-                                                     B.x[0],
-                                                     acc[0],
-                                                     0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1],
-                                                     B.x[1],
-                                                     acc[0],
-                                                     0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
 
 #else
@@ -1329,7 +1248,7 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void mma(
             tile<32, 8, float> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1344,12 +1263,12 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     static __device__ __forceinline__ void mma(
             tile<32, 4, half2> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1364,41 +1283,35 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <data_layout dl_d, data_layout dl_ab>
     static __device__ __forceinline__ void mma(
             tile<16, 16, int, dl_d> & D, const tile<16, 4, int, dl_ab> & A, const tile<16, 4, int, dl_ab> & B) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * acc = (int32x4_t *) D.x;
+#if defined(CDNA4) || defined(CDNA3)
+        const int64_t xA = uint32_t(A.x[0]);
+        const int64_t xB = uint32_t(B.x[0]);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(xA, xB, acc[0], 0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA1)
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+#endif // defined(CDNA4) || defined(CDNA3)
+#elif defined(AMD_WMMA_AVAILABLE)
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], false);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], false);
 #endif // RDNA4
 #else
         GGML_UNUSED(D);

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;
@@ -102,7 +104,7 @@ struct tile_x_sizes {
 };
 
 static int get_mmq_x_max_host(const int cc) {
-    return (amd_mfma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
+    return (turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
         GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
             128                     : 64;
@@ -112,9 +114,9 @@ static int get_mmq_x_max_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_x_max_device() {
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     return 128;
-#else // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#else // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
 #if defined(GGML_USE_HIP)
     return 64;
@@ -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();
@@ -969,13 +1054,13 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             float dB;
             const int j = j0 + tile_C::get_j(0);
@@ -1210,13 +1295,13 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]);
@@ -1350,57 +1435,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * y_qs = (const int   *) y + 4;
-    const float * y_df = (const float *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1] / 2;
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C C;
-                mma(C, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -1425,13 +1460,13 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
@@ -1657,74 +1692,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * y_qs = (const int   *) y + 4;
-    const half2 * y_ds = (const half2 *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x/2 : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y/2;
-            const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0
-                                              : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y
-                                                             : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x);
-
-            tile_C Cm;
-            if (k01 >= MMQ_TILE_NE_K * 3/4) {
-                tile_A A1;
-                A1.x[0] = 0x01010101;
-                A1.x[1] = 0x01010101;
-                mma(Cm, A1, B[0]);
-            }
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C Cd;
-                mma(Cd, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]);
-                    float tmp = Cd.x[l]*dm.x;
-                    if (k01 >= MMQ_TILE_NE_K * 3/4) {
-                        tmp -= Cm.x[l]*dm.y;
-                    }
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB;
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -1749,13 +1717,13 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y;
@@ -2488,59 +2456,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * x_sc = (const int   *) x_df + MMQ_TILE_NE_K/QI6_K;
-    const int   * y_qs = (const int   *) y + 4;
-    const float * y_df = (const float *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1] / 2;
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C C;
-                mma(C, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16);
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -2566,13 +2482,13 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
@@ -3290,6 +3206,14 @@ 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/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-cuda/vendors/hip.h

@@ -33,7 +33,6 @@
 #define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
 #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
-#define NCCL_CHECK(fn) {ncclResult_t err = fn; if(err != ncclSuccess) { GGML_ABORT("RCCL Failure RCCL returned: %i\n", err); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
 #define __shfl_up_sync(mask, var, laneMask, width) __shfl_up(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)

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/hex-utils.h

@@ -31,6 +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;
 }
@@ -73,8 +81,7 @@ static inline void hex_l2fetch(const void * p, uint32_t width, uint32_t stride,
 #define HEX_L2_LINE_SIZE  64
 #define HEX_L2_FLUSH_SIZE (128 * 1024)
 
-static inline void hex_l2flush(void * addr, size_t size)
-{
+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 {
@@ -89,4 +96,8 @@ static inline void hex_l2flush(void * addr, size_t size)
     }
 }
 
+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-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;
@@ -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)
@@ -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);
@@ -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);
             {
@@ -1279,7 +1348,8 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
             }
 
             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);
                 {
@@ -1304,8 +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);
@@ -1318,20 +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);
 
@@ -1352,31 +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;
@@ -1386,36 +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);
 
@@ -1434,29 +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);
         }
     }
 }
@@ -1540,12 +1625,41 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
 
     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);
@@ -1554,7 +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;
     }
 
@@ -1568,8 +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)
     {
@@ -1583,7 +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);
@@ -1601,7 +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
@@ -1617,13 +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;
 
@@ -1663,7 +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);
                 }

ggml/src/ggml-hexagon/htp/hmx-queue.c

@@ -0,0 +1,158 @@
+#pragma clang diagnostic ignored "-Wunused-function"
+
+#include <stdbool.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <qurt_thread.h>
+#include <qurt_futex.h>
+
+#include <HAP_compute_res.h>
+
+#include "hmx-queue.h"
+
+#define QURT_LOWEST_PRIO (254)
+
+static inline void hmx_lock(struct hmx_queue *q)
+{
+    if (!q->hmx_locked) {
+        HAP_compute_res_hmx_lock(q->hap_rctx);
+        q->hmx_locked = true;
+    }
+}
+
+static inline void hmx_unlock(struct hmx_queue *q)
+{
+    if (q->hmx_locked) {
+        HAP_compute_res_hmx_unlock(q->hap_rctx);
+        q->hmx_locked = false;
+    }
+}
+
+static inline void hmx_queue_process(struct hmx_queue *q, bool* killed) {
+    unsigned int ir = atomic_load(&q->idx_read);
+
+    while (ir != atomic_load(&q->idx_write)) {
+        struct hmx_queue_desc *d = &q->desc[ir];
+        if (!d->done) {
+            FARF(HIGH, "hmx-queue-process: ir %u func %p data %p", ir, d->func, d->data);
+
+            enum hmx_queue_signal sig = (enum hmx_queue_signal) (unsigned int) d->func;
+            switch (sig) {
+                case HMX_QUEUE_NOOP:    /* noop */;     break;
+                case HMX_QUEUE_KILL:    *killed = true; break;
+                case HMX_QUEUE_SUSPEND: hmx_unlock(q);  break;
+                default:
+                    hmx_lock(q);
+                    d->func(d->data);
+                    break;
+            }
+
+            atomic_fetch_add(&d->done, 1);
+        }
+
+        ir = (ir + 1) & q->idx_mask;
+        atomic_store(&q->idx_read, ir);
+    }
+}
+
+static void hmx_queue_thread(void * arg) {
+    struct hmx_queue * q = (struct hmx_queue *) arg;
+
+    FARF(HIGH, "hmx-queue-thread: started");
+
+    bool killed = false;
+
+    unsigned int poll_cnt  = HMX_QUEUE_POLL_COUNT;
+    unsigned int prev_seqn = 0;
+    while (!killed) {
+        unsigned int seqn = atomic_load(&q->seqn);
+        if (seqn == prev_seqn) {
+            if (--poll_cnt) { hex_pause(); continue; }
+            FARF(HIGH, "hmx-queue-thread: sleeping");
+            qurt_futex_wait(&q->seqn, prev_seqn);
+            continue;
+        }
+        prev_seqn = seqn;
+        poll_cnt  = HMX_QUEUE_POLL_COUNT;
+
+        FARF(HIGH, "hmx-queue-thread: new work");
+
+        hmx_queue_process(q, &killed);
+    }
+
+    FARF(HIGH, "hmx-queue-thread: stopped");
+}
+
+struct hmx_queue * hmx_queue_create(size_t capacity, uint32_t hap_rctx) {
+    capacity = hex_ceil_pow2(capacity);
+
+    struct hmx_queue * q = (struct hmx_queue *) memalign(32, sizeof(struct hmx_queue));
+    if (q == NULL) {
+        FARF(ERROR, "%s: failed to allocate DMA queue\n", __FUNCTION__);
+        return NULL;
+    }
+    memset(q, 0, sizeof(struct hmx_queue));
+    q->capacity = capacity;
+    q->idx_mask = capacity - 1;
+    q->hap_rctx = hap_rctx;
+
+    q->desc = (struct hmx_queue_desc *) memalign(64, capacity * sizeof(struct hmx_queue_desc));
+    if (!q->desc) {
+        FARF(ERROR, "hmx-queue: failed to allocate HMX queue descriptors\n");
+        return NULL;
+    }
+    memset(q->desc, 0, capacity * sizeof(struct hmx_queue_desc));
+
+    const size_t stack_size = HMX_QUEUE_THREAD_STACK_SIZE;
+    q->stack = (unsigned char *) memalign(64, stack_size);
+    if (!q->stack) {
+        FARF(ERROR, "hmx-queue: thread stack allocation failed (%zu bytes)", stack_size);
+        return NULL;
+    }
+    memset(q->stack, 0, stack_size);
+
+    // Match caller thread priority (same pattern as worker-pool.c).
+    int prio = qurt_thread_get_priority(qurt_thread_get_id());
+    if (prio < 1) {
+        prio = 1;
+    }
+    if (prio > QURT_LOWEST_PRIO) {
+        prio = QURT_LOWEST_PRIO;
+    }
+
+    qurt_thread_attr_t attr;
+    qurt_thread_attr_init(&attr);
+    qurt_thread_attr_set_stack_addr(&attr, q->stack);
+    qurt_thread_attr_set_stack_size(&attr, stack_size);
+    qurt_thread_attr_set_priority(&attr, prio);
+    qurt_thread_attr_set_name(&attr, "hmx-queue");
+
+    int err = qurt_thread_create(&q->thread, &attr, hmx_queue_thread, q);
+    if (err) {
+        FARF(ERROR, "hmx-worker: thread create failed (%d)", err);
+        return NULL;
+    }
+
+    FARF(HIGH, "hmx-queue: capacity %u\n", capacity);
+
+    return q;
+}
+
+void hmx_queue_delete(struct hmx_queue * q) {
+    if (!q) {
+        return;
+    }
+
+    // Tell the worker to exit.
+    hmx_queue_flush(q);
+    hmx_queue_signal(q, HMX_QUEUE_KILL);
+    hmx_queue_flush(q);
+
+    int status;
+    qurt_thread_join(q->thread, &status);
+
+    free(q->desc);
+    free(q->stack);
+    free(q);
+}

ggml/src/ggml-hexagon/htp/hmx-queue.h

@@ -0,0 +1,134 @@
+#ifndef HMX_QUEUE_H
+#define HMX_QUEUE_H
+
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdatomic.h>
+
+#include <hexagon_types.h>
+#include <qurt_thread.h>
+#include <qurt_futex.h>
+#include <HAP_farf.h>
+
+#include "hex-utils.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define HMX_QUEUE_THREAD_STACK_SIZE (16 * 1024)
+#define HMX_QUEUE_POLL_COUNT        2000
+
+typedef void (*hmx_queue_func)(void *);
+
+// Dummy funcs used as signals
+enum hmx_queue_signal {
+    HMX_QUEUE_NOOP = 0, // aka NULL
+    HMX_QUEUE_SUSPEND,
+    HMX_QUEUE_KILL
+};
+
+struct hmx_queue_desc {
+    hmx_queue_func   func;
+    void *           data;
+    atomic_uint      done;
+};
+
+struct hmx_queue {
+    struct hmx_queue_desc * desc;
+    atomic_uint      idx_write; // updated by producer (push)
+    atomic_uint      idx_read;  // updated by consumer (process)
+    unsigned int     idx_pop;   // updated by producer (pop)
+    uint32_t         idx_mask;
+    uint32_t         capacity;
+
+    atomic_uint      seqn;      // incremented for all pushes, used with futex
+    qurt_thread_t    thread;
+    void *           stack;
+    uint32_t         hap_rctx;
+    bool             hmx_locked;
+};
+
+struct hmx_queue * hmx_queue_create(size_t capacity, uint32_t hap_rctx);
+void hmx_queue_delete(struct hmx_queue * q);
+
+static inline struct hmx_queue_desc hmx_queue_make_desc(hmx_queue_func func, void * data) {
+    struct hmx_queue_desc d = { func, data };
+    return d;
+}
+
+static inline bool hmx_queue_push(struct hmx_queue * q, struct hmx_queue_desc d) {
+    unsigned int ir = atomic_load(&q->idx_read);
+    unsigned int iw = q->idx_write;
+
+    if (((iw + 1) & q->idx_mask) == ir) {
+        FARF(HIGH, "hmx-queue-push: queue is full\n");
+        return false;
+    }
+
+    atomic_store(&d.done, 0);
+
+    FARF(HIGH, "hmx-queue-push: iw %u func %p data %p\n", iw, d.func, d.data);
+
+    q->desc[iw] = d;
+    atomic_store(&q->idx_write, (iw + 1) & q->idx_mask);
+    // wake up our thread
+    atomic_fetch_add(&q->seqn, 1);
+    qurt_futex_wake(&q->seqn, 1);
+
+    return true;
+}
+
+static inline bool hmx_queue_signal(struct hmx_queue *q, enum hmx_queue_signal sig) {
+    return hmx_queue_push(q, hmx_queue_make_desc((hmx_queue_func) sig, NULL));
+}
+
+static inline bool hmx_queue_empty(struct hmx_queue * q) {
+    return q->idx_pop == q->idx_write;
+}
+
+static inline uint32_t hmx_queue_depth(struct hmx_queue * q) {
+    return (q->idx_read - q->idx_read) & q->idx_mask;
+}
+
+static inline uint32_t hmx_queue_capacity(struct hmx_queue * q) {
+    return q->capacity;
+}
+
+static inline struct hmx_queue_desc hmx_queue_pop(struct hmx_queue * q) {
+    unsigned int ip = q->idx_pop;
+    unsigned int iw = q->idx_write;
+
+    struct hmx_queue_desc rd = { NULL, NULL };
+    if (ip == iw) {
+        return rd;
+    }
+
+    // Wait for desc to complete
+    struct hmx_queue_desc * d = &q->desc[ip];
+    while (!atomic_load(&d->done)) {
+        FARF(HIGH, "hmx-queue-pop: waiting for HMX queue : %u\n", ip);
+        hex_pause();
+    }
+
+    rd = *d;
+    q->idx_pop = (ip + 1) & q->idx_mask;
+
+    FARF(HIGH, "hmx-queue-pop: ip %u func %p data %p\n", ip, rd.func, rd.data);
+    return rd;
+}
+
+static inline void hmx_queue_flush(struct hmx_queue * q) {
+    while (hmx_queue_pop(q).func != NULL) ;
+}
+
+static inline void hmx_queue_suspend(struct hmx_queue *q) {
+    hmx_queue_signal(q, HMX_QUEUE_SUSPEND);
+    hmx_queue_flush(q);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* HMX_QUEUE_H */

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

@@ -14,10 +14,6 @@
 
 #define HMX_INLINE_ALWAYS inline __attribute__((unused, always_inline))
 
-static HMX_INLINE_ALWAYS void hmx_set_output_scales(const void *scales) {
-    asm volatile("bias = mxmem2(%0)" :: "r"(scales));
-}
-
 // Initialise aligned 256-byte area with scale vector + zero padding.
 static HMX_INLINE_ALWAYS void hmx_init_column_scales(void *out_scales, HVX_Vector v_scale) {
     HVX_Vector *pv = (HVX_Vector *)out_scales;
@@ -25,58 +21,6 @@ static HMX_INLINE_ALWAYS void hmx_init_column_scales(void *out_scales, HVX_Vecto
     *pv   = Q6_V_vzero();
 }
 
-// Load multiple contiguous tiles with :deep streaming.
-// Rt = total region size - 1; the hardware streams through [Rs, Rs + Rt].
-// IMPORTANT: the tile region [Rs, Rs + Rt] must NOT cross a VTCM 4 MB bank
-// boundary, otherwise the mxmem instruction will raise a precise bus error.
-// Callers must ensure their VTCM layout satisfies this constraint.
-static HMX_INLINE_ALWAYS void hmx_load_tiles_fp16(const __fp16 *row_tiles,
-                                                   const __fp16 *col_tiles,
-                                                   size_t n_tiles) {
-    size_t limit = n_tiles * HMX_FP16_TILE_SIZE - 1;
-    asm volatile(
-        "{ activation.hf = mxmem(%0, %1):deep\n"
-        "weight.hf = mxmem(%2, %3) }\n"
-        :: "r"(row_tiles), "r"(limit), "r"(col_tiles), "r"(limit)
-        : "memory");
-}
-
-// Load a single activation+weight tile pair (no :deep streaming).
-// Rt defines the accessible region [Rs, Rs+Rt].  Following the reference formula
-// (limit = n_tiles * HMX_FP16_TILE_SIZE - 1), for a single tile Rt = 2047.
-// The original code used Rt=0x7FFF (32 KB region); when dynamic VTCM allocation
-// places a tile near a 4 MB bank boundary, the oversized region crosses it and
-// triggers a precise bus error (0x2601).  Rt=2047 confines accesses to exactly
-// one 2048-byte tile while covering all 16 HVX vectors (offsets 0..2047).
-static HMX_INLINE_ALWAYS void hmx_load_tile_pair_fp16(const __fp16 *act_tile,
-                                                       const __fp16 *wt_tile) {
-    asm volatile(
-        "{ activation.hf = mxmem(%0, %1)\n"
-        "weight.hf = mxmem(%2, %3) }\n"
-        :: "r"(act_tile), "r"(2047),
-           "r"(wt_tile),  "r"(2047)
-        : "memory");
-}
-
-static HMX_INLINE_ALWAYS void hmx_consume_accumulator_fp16(__fp16 *out) {
-    // Use the combined convert-and-store instruction (matches the reference
-    // Q6_mxmem_AR_after_hf intrinsic).  The previous two-instruction sequence
-    // "cvt.hf = acc(2); mxmem = cvt" used an undocumented Rs=2 parameter.
-    asm volatile(
-        "mxmem(%0, %1):after.hf = acc\n"
-        :: "r"(out), "r"(0)
-        : "memory");
-}
-
-// Compute inner product of two vectors of tiles and store result.
-static HMX_INLINE_ALWAYS void hmx_dot_fp16(__fp16 *out,
-                                            const __fp16 *row_tiles,
-                                            const __fp16 *col_tiles,
-                                            size_t n_tiles) {
-    hmx_load_tiles_fp16(row_tiles, col_tiles, n_tiles);
-    hmx_consume_accumulator_fp16(out);
-}
-
 // --- VTCM sequential allocator (from htp-ops-lib/include/dsp/vtcm_mgr.h) ---
 
 static inline uint8_t *vtcm_seq_alloc(uint8_t **vtcm_ptr, size_t size) {

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -2,6 +2,7 @@
 #define HTP_CTX_H
 
 #include "hex-dma.h"
+#include "hmx-queue.h"
 #include "htp-ops.h"
 #include "worker-pool.h"
 
@@ -30,6 +31,8 @@ struct htp_spad {
     uint32_t                  size_per_thread; // size per thread
 };
 
+struct htp_context;
+
 // Context while processing an Op
 // TODO: fold this into the main context
 struct htp_ops_context {
@@ -72,6 +75,10 @@ struct htp_context {
     atomic_bool            vtcm_needs_release;
 
     struct htp_ops_context octx;
+
+#ifdef HTP_HAS_HMX
+    struct hmx_queue *     hmx_queue; // Async HMX queue for pipeline overlap
+#endif
 };
 
 int op_matmul(struct htp_ops_context * octx);

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -91,7 +91,14 @@ enum htp_op_code {
 #define HTP_OP_MAX_BUFS    8
 #define HTP_OP_MAX_REQS    256
 #define HTP_OP_MAX_TENSORS (HTP_OP_MAX_REQS * HTP_OP_MAX_INPUTS + HTP_OP_MAX_REQS)
+
+#if __HVX_ARCH__ < 75
+#define HTP_OP_MAX_VMEM    (3167538380u)
+#else
 #define HTP_OP_MAX_VMEM    (3221225472u)
+#endif
+
+#define HTP_MMAP_MAX_VMEM  (2147483648u)
 
 enum htp_tensor_flags {
     HTP_TENSOR_COMPUTE = (1U << 0), // Tensor buffer temporal compute data (not weights)

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

@@ -116,9 +116,14 @@ static inline HVX_VectorPred hvx_vec_is_nan_f16(HVX_Vector v) {
 }
 
 static inline HVX_Vector hvx_vec_f32_to_f16_shuff(HVX_Vector v0, HVX_Vector v1) {
+#if __HVX_ARCH__ >= 81
+    HVX_Vector q0 = Q6_Vqf32_equals_Vsf(v0);
+    HVX_Vector q1 = Q6_Vqf32_equals_Vsf(v1);
+#else
     const HVX_Vector zero = Q6_V_vzero();
     HVX_Vector q0 = Q6_Vqf32_vadd_VsfVsf(v0, zero);
     HVX_Vector q1 = Q6_Vqf32_vadd_VsfVsf(v1, zero);
+#endif
     return Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(q1, q0));
 }
 

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

@@ -18,8 +18,9 @@
 #include <remote.h>
 #include <string.h>
 
-#include "hex-dma.h"
 #include "hex-utils.h"
+#include "hex-dma.h"
+#include "hmx-queue.h"
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
@@ -117,7 +118,11 @@ AEEResult htp_iface_close(remote_handle64 handle) {
     // release the mmaps (if any)
     for (uint32_t i=0; i<HTP_MAX_MMAPS; i++) {
         if (ctx->mmap[i].size) {
+#if __HVX_ARCH__ > 73
             HAP_munmap2((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#else
+            HAP_munmap((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#endif
             ctx->mmap[i].size = 0;
             ctx->mmap[i].base = NULL;
             ctx->mmap[i].fd   = -1;
@@ -172,8 +177,16 @@ AEEResult htp_iface_mmap(remote_handle64 handle, int fd, uint32_t size, uint32_t
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
             FARF(HIGH, "mmap : fd %u size %u pinned %u", fd, size, pinned);
-
+#if __HVX_ARCH__ > 73
             void *va = HAP_mmap2(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
+#else
+            if (size > HTP_MMAP_MAX_VMEM) { // HAP_mmap has a size limit of 2GB
+                FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) size);
+                abort(); // can't do much else at this point
+            }
+
+            void *va = HAP_mmap(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
+#endif
             if (va == (void*)-1) {
                 FARF(ERROR, "mmap failed : va %p fd %u size %u", va, fd, (uint32_t) size);
                 return AEE_EFAILED;
@@ -201,7 +214,11 @@ AEEResult htp_iface_munmap(remote_handle64 handle, int fd) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (fd < 0 || m->fd == fd) {
             FARF(HIGH, "unmmap : base %p fd %u size %u", (void*) m->base, m->fd, (uint32_t) m->size);
+#if __HVX_ARCH__ > 73
             HAP_munmap2((void *) m->base, m->size);
+#else
+            HAP_munmap((void *) m->base, m->size);
+#endif
             m->size   = 0;
             m->base   = NULL;
             m->fd     = -1;
@@ -324,6 +341,14 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
 
 #ifdef HTP_HAS_HMX
     ctx->hmx_enabled = use_hmx;
+    ctx->hmx_queue   = NULL;
+    if (use_hmx) {
+        ctx->hmx_queue = hmx_queue_create(16, ctx->vtcm_rctx);
+        if (!ctx->hmx_queue) {
+            FARF(ERROR, "hmx-queue-create failed");
+            ctx->hmx_enabled = false;
+        }
+    }
     FARF(HIGH, "HMX %s (use_hmx=%d)", ctx->hmx_enabled ? "enabled" : "disabled", use_hmx);
 #endif
 
@@ -389,7 +414,11 @@ AEEResult htp_iface_stop(remote_handle64 handle) {
     }
 
 #ifdef HTP_HAS_HMX
-    ctx->hmx_enabled = 0;
+    if (ctx->hmx_queue) {
+        hmx_queue_delete(ctx->hmx_queue);
+        ctx->hmx_queue = NULL;
+    }
+    ctx->hmx_enabled = false;
 #endif
 
     vtcm_free(ctx);
@@ -513,7 +542,11 @@ static inline bool reuse_buf(struct htp_context *ctx, uint32_t *m_reuse, struct
 static inline void drop_mmap(struct htp_context *ctx, struct htp_mmap *m) {
     if (m->size && !m->pinned) {
         FARF(HIGH, "unmap : fd %u base %p size %u pinned %u", m->fd, (void*) m->base, (uint32_t) m->size, m->pinned);
+#if __HVX_ARCH__ > 73
         HAP_munmap2((void *) m->base, m->size);
+#else
+        HAP_munmap((void *) m->base, m->size);
+#endif
         m->size = 0;
         m->base = 0;
         m->fd   = -1;
@@ -527,7 +560,16 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
     for (uint32_t i=0; i < HTP_MAX_MMAPS; i++) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
+#if __HVX_ARCH__ > 73
             void *va = HAP_mmap2(NULL, b->size, HAP_PROT_READ | HAP_PROT_WRITE, 0, b->fd, 0);
+#else
+            if (b->size > HTP_MMAP_MAX_VMEM) { // HAP_mmap has a size limit of 2GB
+                FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) b->size);
+                abort(); // can't do much else at this point
+            }
+
+            void *va = HAP_mmap(NULL, b->size, HAP_PROT_READ | HAP_PROT_WRITE, 0, b->fd, 0);
+#endif
             if (va == (void*)-1) {
                 FARF(ERROR, "mmap failed : va %p fd %u size %u", va, b->fd, (uint32_t) b->size);
                 abort(); // can't do much else at this point

ggml/src/ggml-impl.h

@@ -30,6 +30,8 @@ extern "C" {
 
 void ggml_print_backtrace(void);
 
+uint64_t ggml_graph_next_uid(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif
@@ -338,6 +340,10 @@ struct ggml_cgraph {
     struct ggml_hash_set visited_hash_set;
 
     enum ggml_cgraph_eval_order order;
+
+    // an optional identifier that can be utilized to recognize same graphs if two non-zero values match
+    // a value of 0 means it is not set and should be ignored
+    uint64_t uid;
 };
 
 // returns a slice of cgraph with nodes [i0, i1)

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

@@ -250,6 +250,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary(ggml_metal
                 case GGML_UNARY_OP_CEIL:        op_num = OP_UNARY_NUM_CEIL;        break;
                 case GGML_UNARY_OP_ROUND:       op_num = OP_UNARY_NUM_ROUND;       break;
                 case GGML_UNARY_OP_TRUNC:       op_num = OP_UNARY_NUM_TRUNC;       break;
+                case GGML_UNARY_OP_XIELU:       op_num = OP_UNARY_NUM_XIELU;       break;
                 default: GGML_ABORT("fatal error");
             } break;
         default: GGML_ABORT("fatal error");
@@ -1818,6 +1819,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale(ggml_met
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_roll(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_ROLL);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_roll_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_PAD);
 

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

@@ -152,6 +152,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_3d
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_roll              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_arange            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_timestep_embedding(ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_adamw    (ggml_metal_library_t lib, const struct ggml_tensor * op);

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

@@ -1043,6 +1043,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
                 case GGML_UNARY_OP_TRUNC:
+                case GGML_UNARY_OP_XIELU:
                     return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
                 default:
                     return false;
@@ -1137,6 +1138,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_ARGSORT:
         case GGML_OP_TOP_K:
         case GGML_OP_ARANGE:
+        case GGML_OP_ROLL:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
             // for new head sizes, add checks here
@@ -1159,6 +1161,23 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             if (op->src[1]->type != op->src[2]->type) {
                 return false;
             }
+            switch (op->src[1]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                    break;
+                case GGML_TYPE_BF16:
+                    if (!has_bfloat) {
+                        return false;
+                    }
+                    break;
+                default:
+                    return false;
+            }
             return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:

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

@@ -127,6 +127,7 @@
 #define OP_UNARY_NUM_CEIL        118
 #define OP_UNARY_NUM_ROUND       119
 #define OP_UNARY_NUM_TRUNC       120
+#define OP_UNARY_NUM_XIELU       121
 
 #define OP_SUM_ROWS_NUM_SUM_ROWS 10
 #define OP_SUM_ROWS_NUM_MEAN     11
@@ -1016,6 +1017,29 @@ typedef struct {
     int32_t  p1;
 } ggml_metal_kargs_pad_reflect_1d;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  ne0;
+    int64_t  ne1;
+    int64_t  ne2;
+    int64_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  s2;
+    int32_t  s3;
+} ggml_metal_kargs_roll;
+
 typedef struct {
     uint64_t nb1;
     int      dim;