models : move build_inp_out_ids outside loop (#17151)

* move build_inp_out_ids outside loop

* realign

.devops/s390x.Dockerfile

@@ -24,8 +24,9 @@ RUN --mount=type=cache,target=/root/.ccache \
         -DCMAKE_C_COMPILER_LAUNCHER=ccache \
         -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
         -DLLAMA_BUILD_TESTS=OFF \
-        -DGGML_BACKEND_DL=OFF \
         -DGGML_NATIVE=OFF \
+        -DGGML_BACKEND_DL=ON \
+        -DGGML_CPU_ALL_VARIANTS=ON \
         -DGGML_BLAS=ON \
         -DGGML_BLAS_VENDOR=OpenBLAS && \
     cmake --build build --config Release -j $(nproc) && \
@@ -103,6 +104,7 @@ FROM base AS light
 WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
 COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin
 
 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
@@ -116,6 +118,7 @@ ENV LLAMA_ARG_HOST=0.0.0.0
 WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
 COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
 
 EXPOSE 8080

.devops/vulkan.Dockerfile

@@ -1,4 +1,4 @@
-ARG UBUNTU_VERSION=24.04
+ARG UBUNTU_VERSION=25.10
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
@@ -7,32 +7,16 @@ FROM ubuntu:$UBUNTU_VERSION AS build
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget xz-utils
 
-# Install Vulkan SDK
-ARG VULKAN_VERSION=1.4.321.1
-RUN ARCH=$(uname -m) && \
-    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
-    mkdir -p /opt/vulkan && \
-    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
-    mv /tmp/${ARCH}/* /opt/vulkan/ && \
-    rm -rf /tmp/*
-
 # Install cURL and Vulkan SDK dependencies
 RUN apt install -y libcurl4-openssl-dev curl \
-    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
-
-# Set environment variables
-ENV VULKAN_SDK=/opt/vulkan
-ENV PATH=$VULKAN_SDK/bin:$PATH
-ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
-ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
-ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc
 
 # Build it
 WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \
@@ -50,7 +34,7 @@ RUN mkdir -p /app/full \
 FROM ubuntu:$UBUNTU_VERSION AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl libvulkan-dev \
+    && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.editorconfig

@@ -60,3 +60,11 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
+
+[benches/**]
+indent_style = unset
+indent_size = unset
+end_of_line = unset
+charset = unset
+trim_trailing_whitespace = unset
+insert_final_newline = unset

.github/labeler.yml

@@ -76,6 +76,10 @@ ggml:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/**
+model:
+    - changed-files:
+        - any-glob-to-any-file:
+            - src/models/**
 nix:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-linux-cross.yml

@@ -4,49 +4,49 @@ on:
   workflow_call:
 
 jobs:
-  ubuntu-24-riscv64-cpu-cross:
-    runs-on: ubuntu-24.04
+  # ubuntu-24-riscv64-cpu-cross:
+  #   runs-on: ubuntu-24.04
 
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo dpkg --add-architecture riscv64
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup Riscv
+  #       run: |
+  #         sudo dpkg --add-architecture riscv64
 
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
 
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 gcc-14-riscv64-linux-gnu \
+  #                 g++-14-riscv64-linux-gnu
 
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+  #                        -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+  #                        -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-          cmake --build build --config Release -j $(nproc)
+  #         cmake --build build --config Release -j $(nproc)
 
   # ubuntu-24-riscv64-vulkan-cross:
   #   runs-on: ubuntu-24.04

.github/workflows/build.yml

@@ -161,15 +161,16 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          DAWN_VERSION="v1.0.0"
+          DAWN_VERSION="v2.0.0"
           DAWN_OWNER="reeselevine"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-macos-latest-Release.tar.gz"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release.zip"
           echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          curl -L -o artifact.tar.gz \
+          curl -L -o artifact.zip \
             "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
           mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+          unzip artifact.zip
+          tar -xvf Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -387,6 +388,39 @@ jobs:
           cd build
           ctest -L main --verbose
 
+  ubuntu-24-cmake-vulkan-deb:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-24-cmake-vulkan-deb
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get install -y glslc libvulkan-dev libcurl4-openssl-dev
+
+      - name: Configure
+        id: cmake_configure
+        run: |
+          cmake -B build \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGGML_VULKAN=ON
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake --build build -j $(nproc)
+
   ubuntu-24-cmake-vulkan:
     runs-on: ubuntu-24.04
 
@@ -488,15 +522,16 @@ jobs:
         id: dawn-depends
         run: |
           sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v1.0.0"
+          DAWN_VERSION="v2.0.0"
           DAWN_OWNER="reeselevine"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-ubuntu-latest-Release.tar.gz"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release.zip"
           echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          curl -L -o artifact.tar.gz \
+          curl -L -o artifact.zip \
             "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
           mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+          unzip artifact.zip
+          tar -xvf Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -1272,6 +1307,81 @@ jobs:
           cd examples/llama.android
           ./gradlew build --no-daemon
 
+  android-ndk-build:
+    runs-on: ubuntu-latest
+
+    env:
+      OPENCL_VERSION: 2025.07.22
+
+    strategy:
+      matrix:
+        include:
+          - build: 'arm64-cpu'
+            defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_CURL=OFF -D GGML_OPENMP=OFF'
+          - build: 'arm64-snapdragon'
+            defines: '--preset arm64-android-snapdragon-release'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        run: |
+          mkdir opencl
+          curl -L -o opencl/clhpp.tar.gz      https://github.com/KhronosGroup/OpenCL-CLHPP/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          curl -L -o opencl/headers.tar.gz    https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          curl -L -o opencl/icd-loader.tar.gz https://github.com/KhronosGroup/OpenCL-ICD-Loader/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
+          tar -xaf opencl/headers.tar.gz    -C opencl
+          tar -xaf opencl/clhpp.tar.gz      -C opencl
+          tar -xaf opencl/icd-loader.tar.gz -C opencl
+          sudo cp -r opencl/OpenCL-Headers-${OPENCL_VERSION}/CL         ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+          sudo cp -r opencl/OpenCL-CLHPP-${OPENCL_VERSION}/include/CL/* ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/CL
+          cd opencl/OpenCL-ICD-Loader-${OPENCL_VERSION}
+          cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -DOPENCL_ICD_LOADER_HEADERS_DIR=${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=31 -DANDROID_STL=c++_shared
+          cmake --build build
+          sudo cp build/libOpenCL.so ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+          rm -rf opencl
+
+      - name: Install Hexagon SDK
+        id: install_hexsdk
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        env:
+          HEXSDK_VER: 6.4.0.2
+          HEXTLS_VER: 19.0.04
+        run: |
+          curl -L -o hex-sdk.tar.gz https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v$HEXSDK_VER/hexagon-sdk-v$HEXSDK_VER-amd64-lnx.tar.xz
+          mkdir hex-sdk
+          tar -xaf hex-sdk.tar.gz -C hex-sdk
+          ls -l hex-sdk
+          sudo mv hex-sdk /opt/hexagon
+          echo "HEXAGON_SDK_ROOT=/opt/hexagon/$HEXSDK_VER"                                     >> "$GITHUB_ENV"
+          echo "HEXAGON_TOOLS_ROOT=/opt/hexagon/$HEXSDK_VER/tools/HEXAGON_Tools/$HEXTLS_VER"   >> "$GITHUB_ENV"
+          echo "DEFAULT_HLOS_ARCH=64"                                                          >> "$GITHUB_ENV"
+          echo "DEFAULT_TOOLS_VARIANT=toolv19"                                                 >> "$GITHUB_ENV"
+          echo "DEFAULT_NO_QURT_INC=0"                                                         >> "$GITHUB_ENV"
+          echo "DEFAULT_DSP_ARCH=v73"                                                          >> "$GITHUB_ENV"
+
+      - name: Update CMake presets
+        id: update_presets
+        if: ${{ matrix.build == 'arm64-snapdragon' }}
+        run: |
+          cp docs/backend/hexagon/CMakeUserPresets.json .
+
+      - name: Build
+        id: ndk_build
+        run: |
+          cmake ${{ matrix.defines }} -B build
+          cmake --build build
+          cmake --install build --prefix pkg-adb/llama.cpp
+
+      - name: Test
+        id: cmake_test
+        run: |
+          echo "FIXME: test on devices"
+
   openEuler-latest-cmake-cann:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:

.github/workflows/docker.yml

@@ -40,7 +40,7 @@ jobs:
           # https://github.com/ggml-org/llama.cpp/issues/11888
           #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
+          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }

.github/workflows/release.yml

@@ -134,6 +134,8 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
+          - build: 's390x'
+            os: ubuntu-24.04-s390x
           # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
           # - build: 'arm64'
           #   os: ubuntu-22.04-arm

.github/workflows/update-ops-docs.yml

@@ -3,10 +3,12 @@ name: Update Operations Documentation
 on:
     push:
         paths:
+            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
     pull_request:
         paths:
+            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
 

CODEOWNERS

@@ -55,7 +55,7 @@
 /ggml/src/ggml-cuda/common.cuh          @slaren
 /ggml/src/ggml-cuda/fattn*              @JohannesGaessler
 /ggml/src/ggml-cuda/ggml-cuda.cu        @slaren
-/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler
+/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
 /ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
 /ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
 /ggml/src/ggml-cuda/mmvq.*              @JohannesGaessler
@@ -65,6 +65,7 @@
 /ggml/src/ggml-impl.h                   @ggerganov @slaren
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
+/ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov
@@ -88,6 +89,7 @@
 /src/llama-model-loader.*               @slaren
 /src/llama-model.*                      @CISC
 /src/llama-vocab.*                      @CISC
+/src/models/                            @CISC
 /tests/                                 @ggerganov
 /tests/test-backend-ops.cpp             @slaren
 /tests/test-thread-safety.cpp           @slaren

README.md

@@ -17,14 +17,13 @@ LLM inference in C/C++
 
 ## Hot topics
 
-- **[guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)**
-- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
+- **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
+- [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
+- [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
 - Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
-- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
 - Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
-- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
@@ -84,6 +83,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
+- [x] [Jamba](https://huggingface.co/ai21labs)
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
@@ -138,6 +138,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
+- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 
 #### Multimodal
 
@@ -187,6 +188,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
 - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 - Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
+- Go (no CGo needed): [hybridgroup/yzma](https://github.com/hybridgroup/yzma)
 
 </details>
 
@@ -278,6 +280,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
+| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
 
 ## Obtaining and quantizing models
 

benches/dgx-spark/aime25_openai__gpt-oss-120b-high_temp1.0_20251109_094547.json

@@ -0,0 +1,6 @@
+{
+  "chars": 2296.1916666666666,
+  "chars:std": 986.051306946325,
+  "score": 0.925,
+  "score:std": 0.26339134382131846
+}

benches/dgx-spark/dgx-spark.md

@@ -0,0 +1,264 @@
+## System info
+
+```bash
+uname --all
+Linux spark-17ed 6.11.0-1016-nvidia #16-Ubuntu SMP PREEMPT_DYNAMIC Sun Sep 21 16:52:46 UTC 2025 aarch64 aarch64 aarch64 GNU/Linux
+
+g++ --version
+g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
+
+nvidia-smi
+Sun Nov  2 10:43:25 2025
++-----------------------------------------------------------------------------------------+
+| NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
++-----------------------------------------+------------------------+----------------------+
+| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
+| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
+|                                         |                        |               MIG M. |
+|=========================================+========================+======================|
+|   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
+| N/A   35C    P8              4W /  N/A  | Not Supported          |      0%      Default |
+|                                         |                        |                  N/A |
++-----------------------------------------+------------------------+----------------------+
+```
+
+## ggml-org/gpt-oss-20b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.374 |  1369.01 |    0.383 |    83.64 |    0.757 |   719.01 |
+|   512 |     32 |    2 |   1088 |    0.274 |  3741.35 |    0.659 |    97.14 |    0.933 |  1166.66 |
+|   512 |     32 |    4 |   2176 |    0.526 |  3896.47 |    0.817 |   156.73 |    1.342 |  1621.08 |
+|   512 |     32 |    8 |   4352 |    1.044 |  3925.10 |    0.987 |   259.44 |    2.030 |  2143.56 |
+|   512 |     32 |   16 |   8704 |    2.076 |  3945.84 |    1.248 |   410.32 |    3.324 |  2618.60 |
+|   512 |     32 |   32 |  17408 |    4.170 |  3929.28 |    1.630 |   628.40 |    5.799 |  3001.76 |
+|  4096 |     32 |    1 |   4128 |    1.083 |  3782.66 |    0.394 |    81.21 |    1.477 |  2795.13 |
+|  4096 |     32 |    2 |   8256 |    2.166 |  3782.72 |    0.725 |    88.28 |    2.891 |  2856.14 |
+|  4096 |     32 |    4 |  16512 |    4.333 |  3780.88 |    0.896 |   142.82 |    5.230 |  3157.38 |
+|  4096 |     32 |    8 |  33024 |    8.618 |  3802.14 |    1.155 |   221.69 |    9.773 |  3379.08 |
+|  4096 |     32 |   16 |  66048 |   17.330 |  3781.73 |    1.598 |   320.34 |   18.928 |  3489.45 |
+|  4096 |     32 |   32 | 132096 |   34.671 |  3780.48 |    2.336 |   438.35 |   37.007 |  3569.51 |
+|  8192 |     32 |    1 |   8224 |    2.233 |  3668.56 |    0.438 |    72.98 |    2.671 |  3078.44 |
+|  8192 |     32 |    2 |  16448 |    4.425 |  3702.95 |    0.756 |    84.66 |    5.181 |  3174.95 |
+|  8192 |     32 |    4 |  32896 |    8.859 |  3698.64 |    0.967 |   132.38 |    9.826 |  3347.72 |
+|  8192 |     32 |    8 |  65792 |   17.714 |  3699.57 |    1.277 |   200.52 |   18.991 |  3464.35 |
+|  8192 |     32 |   16 | 131584 |   35.494 |  3692.84 |    1.841 |   278.12 |   37.335 |  3524.46 |
+|  8192 |     32 |   32 | 263168 |   70.949 |  3694.82 |    2.798 |   365.99 |   73.747 |  3568.53 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      3714.25 ± 20.36 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         86.58 ± 0.43 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |      3445.17 ± 17.85 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         81.72 ± 0.53 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      3218.78 ± 11.34 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.86 ± 0.64 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       2732.83 ± 7.17 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.57 ± 0.51 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      2119.75 ± 12.81 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         62.33 ± 0.24 |
+
+build: eeee367de (6989)
+
+## ggml-org/gpt-oss-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.571 |   897.18 |    0.543 |    58.96 |    1.113 |   488.60 |
+|   512 |     32 |    2 |   1088 |    0.593 |  1725.37 |    1.041 |    61.45 |    1.635 |   665.48 |
+|   512 |     32 |    4 |   2176 |    1.043 |  1963.15 |    1.334 |    95.95 |    2.377 |   915.36 |
+|   512 |     32 |    8 |   4352 |    2.099 |  1951.63 |    1.717 |   149.07 |    3.816 |  1140.45 |
+|   512 |     32 |   16 |   8704 |    4.207 |  1947.12 |    2.311 |   221.56 |    6.518 |  1335.35 |
+|   512 |     32 |   32 |  17408 |    8.422 |  1945.36 |    3.298 |   310.46 |   11.720 |  1485.27 |
+|  4096 |     32 |    1 |   4128 |    2.138 |  1915.88 |    0.571 |    56.09 |    2.708 |  1524.12 |
+|  4096 |     32 |    2 |   8256 |    4.266 |  1920.25 |    1.137 |    56.27 |    5.404 |  1527.90 |
+|  4096 |     32 |    4 |  16512 |    8.564 |  1913.02 |    1.471 |    86.99 |   10.036 |  1645.29 |
+|  4096 |     32 |    8 |  33024 |   17.092 |  1917.19 |    1.979 |   129.33 |   19.071 |  1731.63 |
+|  4096 |     32 |   16 |  66048 |   34.211 |  1915.65 |    2.850 |   179.66 |   37.061 |  1782.15 |
+|  4096 |     32 |   32 | 132096 |   68.394 |  1916.44 |    4.381 |   233.72 |   72.775 |  1815.13 |
+|  8192 |     32 |    1 |   8224 |    4.349 |  1883.45 |    0.620 |    51.65 |    4.969 |  1655.04 |
+|  8192 |     32 |    2 |  16448 |    8.674 |  1888.83 |    1.178 |    54.33 |    9.852 |  1669.48 |
+|  8192 |     32 |    4 |  32896 |   17.351 |  1888.55 |    1.580 |    81.01 |   18.931 |  1737.68 |
+|  8192 |     32 |    8 |  65792 |   34.743 |  1886.31 |    2.173 |   117.80 |   36.916 |  1782.20 |
+|  8192 |     32 |   16 | 131584 |   69.413 |  1888.29 |    3.297 |   155.28 |   72.710 |  1809.70 |
+|  8192 |     32 |   32 | 263168 |  138.903 |  1887.24 |    5.004 |   204.63 |  143.907 |  1828.73 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       1919.36 ± 5.01 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         60.40 ± 0.30 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       1825.30 ± 6.37 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         56.94 ± 0.29 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1739.19 ± 6.00 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         52.51 ± 0.42 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1536.75 ± 4.27 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         49.33 ± 0.27 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1255.85 ± 3.26 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         42.99 ± 0.18 |
+
+build: eeee367de (6989)
+
+## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.398 |  1285.90 |    0.530 |    60.41 |    0.928 |   586.27 |
+|   512 |     32 |    2 |   1088 |    0.386 |  2651.65 |    0.948 |    67.50 |    1.334 |   815.38 |
+|   512 |     32 |    4 |   2176 |    0.666 |  3076.37 |    1.209 |   105.87 |    1.875 |  1160.71 |
+|   512 |     32 |    8 |   4352 |    1.325 |  3091.39 |    1.610 |   158.98 |    2.935 |  1482.65 |
+|   512 |     32 |   16 |   8704 |    2.664 |  3075.58 |    2.150 |   238.19 |    4.813 |  1808.39 |
+|   512 |     32 |   32 |  17408 |    5.336 |  3070.31 |    2.904 |   352.59 |    8.240 |  2112.50 |
+|  4096 |     32 |    1 |   4128 |    1.444 |  2836.81 |    0.581 |    55.09 |    2.025 |  2038.81 |
+|  4096 |     32 |    2 |   8256 |    2.872 |  2852.14 |    1.084 |    59.06 |    3.956 |  2086.99 |
+|  4096 |     32 |    4 |  16512 |    5.744 |  2852.32 |    1.440 |    88.90 |    7.184 |  2298.47 |
+|  4096 |     32 |    8 |  33024 |   11.463 |  2858.68 |    2.068 |   123.78 |   13.531 |  2440.65 |
+|  4096 |     32 |   16 |  66048 |   22.915 |  2859.95 |    3.018 |   169.67 |   25.933 |  2546.90 |
+|  4096 |     32 |   32 | 132096 |   45.956 |  2852.10 |    4.609 |   222.18 |   50.565 |  2612.39 |
+|  8192 |     32 |    1 |   8224 |    3.063 |  2674.72 |    0.693 |    46.20 |    3.755 |  2189.92 |
+|  8192 |     32 |    2 |  16448 |    6.109 |  2681.87 |    1.214 |    52.71 |    7.323 |  2245.98 |
+|  8192 |     32 |    4 |  32896 |   12.197 |  2686.63 |    1.682 |    76.11 |   13.878 |  2370.30 |
+|  8192 |     32 |    8 |  65792 |   24.409 |  2684.94 |    2.556 |   100.17 |   26.965 |  2439.95 |
+|  8192 |     32 |   16 | 131584 |   48.753 |  2688.50 |    3.994 |   128.20 |   52.747 |  2494.64 |
+|  8192 |     32 |   32 | 263168 |   97.508 |  2688.42 |    6.528 |   156.86 |  104.037 |  2529.57 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2925.55 ± 4.25 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         62.80 ± 0.27 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2531.01 ± 6.79 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         55.86 ± 0.33 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       2244.39 ± 5.33 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         45.95 ± 0.33 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1783.17 ± 3.68 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.07 ± 0.10 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1241.90 ± 3.13 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         29.92 ± 0.06 |
+
+build: eeee367de (6989)
+
+## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.211 |  2421.57 |    1.055 |    30.33 |    1.266 |   429.57 |
+|   512 |     32 |    2 |   1088 |    0.419 |  2441.34 |    1.130 |    56.65 |    1.549 |   702.32 |
+|   512 |     32 |    4 |   2176 |    0.873 |  2345.54 |    1.174 |   108.99 |    2.048 |  1062.74 |
+|   512 |     32 |    8 |   4352 |    1.727 |  2371.85 |    1.254 |   204.22 |    2.980 |  1460.19 |
+|   512 |     32 |   16 |   8704 |    3.452 |  2373.22 |    1.492 |   343.16 |    4.944 |  1760.56 |
+|   512 |     32 |   32 |  17408 |    6.916 |  2368.93 |    1.675 |   611.51 |    8.591 |  2026.36 |
+|  4096 |     32 |    1 |   4128 |    1.799 |  2277.26 |    1.084 |    29.51 |    2.883 |  1431.91 |
+|  4096 |     32 |    2 |   8256 |    3.577 |  2290.01 |    1.196 |    53.50 |    4.774 |  1729.51 |
+|  4096 |     32 |    4 |  16512 |    7.172 |  2284.36 |    1.313 |    97.50 |    8.485 |  1946.00 |
+|  4096 |     32 |    8 |  33024 |   14.341 |  2284.96 |    1.520 |   168.46 |   15.860 |  2082.18 |
+|  4096 |     32 |   16 |  66048 |   28.675 |  2285.44 |    1.983 |   258.21 |   30.658 |  2154.33 |
+|  4096 |     32 |   32 | 132096 |   57.354 |  2285.32 |    2.640 |   387.87 |   59.994 |  2201.82 |
+|  8192 |     32 |    1 |   8224 |    3.701 |  2213.75 |    1.119 |    28.59 |    4.820 |  1706.34 |
+|  8192 |     32 |    2 |  16448 |    7.410 |  2211.19 |    1.272 |    50.31 |    8.682 |  1894.56 |
+|  8192 |     32 |    4 |  32896 |   14.802 |  2213.83 |    1.460 |    87.68 |   16.261 |  2022.96 |
+|  8192 |     32 |    8 |  65792 |   29.609 |  2213.35 |    1.781 |   143.74 |   31.390 |  2095.93 |
+|  8192 |     32 |   16 | 131584 |   59.229 |  2212.96 |    2.495 |   205.17 |   61.725 |  2131.79 |
+|  8192 |     32 |   32 | 263168 |  118.449 |  2213.15 |    3.714 |   275.75 |  122.162 |  2154.25 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2272.74 ± 4.68 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         30.66 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2107.80 ± 9.55 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         29.71 ± 0.05 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1937.80 ± 6.75 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         28.86 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1641.12 ± 1.78 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.24 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1296.02 ± 2.67 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.78 ± 0.03 |
+
+build: eeee367de (6989)
+
+## ggml-org/gemma-3-4b-it-qat-GGUF
+
+Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.094 |  5434.73 |    0.394 |    81.21 |    0.488 |  1114.15 |
+|   512 |     32 |    2 |   1088 |    0.168 |  6091.68 |    0.498 |   128.52 |    0.666 |  1633.41 |
+|   512 |     32 |    4 |   2176 |    0.341 |  6010.68 |    0.542 |   236.37 |    0.882 |  2466.43 |
+|   512 |     32 |    8 |   4352 |    0.665 |  6161.46 |    0.678 |   377.74 |    1.342 |  3241.72 |
+|   512 |     32 |   16 |   8704 |    1.323 |  6193.19 |    0.902 |   567.41 |    2.225 |  3911.74 |
+|   512 |     32 |   32 |  17408 |    2.642 |  6202.03 |    1.231 |   832.03 |    3.872 |  4495.36 |
+|  4096 |     32 |    1 |   4128 |    0.701 |  5840.49 |    0.439 |    72.95 |    1.140 |  3621.23 |
+|  4096 |     32 |    2 |   8256 |    1.387 |  5906.82 |    0.574 |   111.48 |    1.961 |  4210.12 |
+|  4096 |     32 |    4 |  16512 |    2.758 |  5940.33 |    0.651 |   196.58 |    3.409 |  4843.33 |
+|  4096 |     32 |    8 |  33024 |    5.491 |  5967.56 |    0.876 |   292.40 |    6.367 |  5187.12 |
+|  4096 |     32 |   16 |  66048 |   10.978 |  5969.58 |    1.275 |   401.69 |   12.253 |  5390.38 |
+|  4096 |     32 |   32 | 132096 |   21.944 |  5972.93 |    1.992 |   514.16 |   23.936 |  5518.73 |
+|  8192 |     32 |    1 |   8224 |    1.402 |  5841.91 |    0.452 |    70.73 |    1.855 |  4434.12 |
+|  8192 |     32 |    2 |  16448 |    2.793 |  5865.34 |    0.637 |   100.55 |    3.430 |  4795.51 |
+|  8192 |     32 |    4 |  32896 |    5.564 |  5889.64 |    0.770 |   166.26 |    6.334 |  5193.95 |
+|  8192 |     32 |    8 |  65792 |   11.114 |  5896.44 |    1.122 |   228.07 |   12.237 |  5376.51 |
+|  8192 |     32 |   16 | 131584 |   22.210 |  5901.38 |    1.789 |   286.15 |   24.000 |  5482.74 |
+|  8192 |     32 |   32 | 263168 |   44.382 |  5906.56 |    3.044 |   336.38 |   47.426 |  5549.02 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      5810.04 ± 21.71 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         84.54 ± 0.18 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       5288.04 ± 3.54 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         78.82 ± 1.37 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      4960.43 ± 16.64 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.13 ± 0.30 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |      4495.92 ± 31.11 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         72.37 ± 0.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      3746.90 ± 40.01 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         63.02 ± 0.20 |
+
+build: eeee367de (6989)
+

ci/run.sh

@@ -75,7 +75,7 @@ if [ ! -z ${GG_BUILD_ROCM} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DAMDGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then

common/CMakeLists.txt

@@ -56,6 +56,8 @@ add_library(${TARGET} STATIC
     common.h
     console.cpp
     console.h
+    download.cpp
+    download.h
     http.h
     json-partial.cpp
     json-partial.h

common/arg.h

@@ -59,8 +59,8 @@ struct common_arg {
     common_arg & set_sparam();
     bool in_example(enum llama_example ex);
     bool is_exclude(enum llama_example ex);
-    bool get_value_from_env(std::string & output);
-    bool has_value_from_env();
+    bool get_value_from_env(std::string & output) const;
+    bool has_value_from_env() const;
     std::string to_string();
 };
 

common/chat-parser.cpp

@@ -432,7 +432,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         if (is_arguments_path({})) {
             // Entire JSON is the arguments and was parsed fully.
             return consume_json_result {
-                partial->json.dump(),
+                partial->json.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true),
                 /* .is_partial = */ false,
             };
         }
@@ -444,7 +444,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
     std::vector<std::string> path;
     std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
         if (is_arguments_path(path)) {
-            auto arguments = j.dump();
+            auto arguments = j.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true);
             if (is_partial() && !partial->healing_marker.marker.empty()) {
                 auto idx = arguments.find(partial->healing_marker.json_dump_marker);
                 if (idx != std::string::npos) {

common/chat.cpp

@@ -9,8 +9,11 @@
 #include <minja/chat-template.hpp>
 #include <minja/minja.hpp>
 
+#include <algorithm>
 #include <cstdio>
+#include <cctype>
 #include <exception>
+#include <functional>
 #include <iostream>
 #include <optional>
 #include <stdexcept>
@@ -310,7 +313,6 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
-            jmsg["thinking"] = msg.reasoning_content; // gpt-oss
         }
         if (!msg.tool_name.empty()) {
             jmsg["name"] = msg.tool_name;
@@ -640,6 +642,7 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_SEED_OSS: return "Seed-OSS";
         case COMMON_CHAT_FORMAT_NEMOTRON_V2: return "Nemotron V2";
         case COMMON_CHAT_FORMAT_APERTUS: return "Apertus";
+        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS: return "LFM2 with JSON tools";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -986,6 +989,126 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     return data;
 }
 
+
+// Case-insensitive find
+static size_t ifind_string(const std::string & haystack, const std::string & needle, size_t pos = 0) {
+    auto it = std::search(
+        haystack.begin() + pos, haystack.end(),
+        needle.begin(), needle.end(),
+        [](char a, char b) { return std::tolower(a) == std::tolower(b); }
+    );
+    return (it == haystack.end()) ? std::string::npos : std::distance(haystack.begin(), it);
+}
+
+static common_chat_params common_chat_params_init_lfm2(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+    const auto is_json_schema_provided = !inputs.json_schema.is_null();
+    const auto is_grammar_provided = !inputs.grammar.empty();
+    const auto are_tools_provided = inputs.tools.is_array() && !inputs.tools.empty();
+
+    // the logic requires potentially modifying the messages
+    auto tweaked_messages = inputs.messages;
+
+    auto replace_json_schema_marker = [](json & messages) -> bool {
+        static std::string marker1 = "force json schema.\n";
+        static std::string marker2 = "force json schema.";
+
+        if (messages.empty() || messages.at(0).at("role") != "system") {
+            return false;
+        }
+
+        std::string content = messages.at(0).at("content");
+
+        for (const auto & marker : {marker1, marker2}) {
+            const auto pos = ifind_string(content, marker);
+            if (pos != std::string::npos) {
+                content.replace(pos, marker.length(), "");
+                // inject modified content back into the messages
+                messages.at(0).at("content") = content;
+                return true;
+            }
+        }
+
+        return false;
+    };
+
+    // Lfm2 model does not natively work with json, but can generally understand the tools structure
+    //
+    // Example of the pytorch dialog structure:
+    //     <|startoftext|><|im_start|>system
+    //     List of tools: <|tool_list_start|>[{"name": "get_candidate_status", "description": "Retrieves the current status of a candidate in the recruitment process", "parameters": {"type": "object", "properties": {"candidate_id": {"type": "string", "description": "Unique identifier for the candidate"}}, "required": ["candidate_id"]}}]<|tool_list_end|><|im_end|>
+    //     <|im_start|>user
+    //     What is the current status of candidate ID 12345?<|im_end|>
+    //     <|im_start|>assistant
+    //     <|tool_call_start|>[get_candidate_status(candidate_id="12345")]<|tool_call_end|>Checking the current status of candidate ID 12345.<|im_end|>
+    //     <|im_start|>tool
+    //     <|tool_response_start|>{"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"}<|tool_response_end|><|im_end|>
+    //     <|im_start|>assistant
+    //     The candidate with ID 12345 is currently in the "Interview Scheduled" stage for the position of Clinical Research Associate, with an interview date set for 2023-11-20.<|im_end|>
+    //
+    // For the llama server compatibility with json tools semantic,
+    // the client can add "Follow json schema." line into the system message prompt to force the json output.
+    //
+    if (are_tools_provided && (is_json_schema_provided || is_grammar_provided)) {
+        // server/utils.hpp prohibits that branch for the custom grammar anyways
+        throw std::runtime_error("Tools call must not use \"json_schema\" or \"grammar\", use non-tool invocation if you want to use custom grammar");
+    } else if (are_tools_provided && replace_json_schema_marker(tweaked_messages)) {
+        LOG_INF("%s: Using tools to build a grammar\n", __func__);
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            auto schemas = json::array();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                schemas.push_back({
+                    {"type", "object"},
+                    {"properties", {
+                        {"name", {
+                            {"type", "string"},
+                            {"const", function.at("name")},
+                        }},
+                        {"arguments", function.at("parameters")},
+                    }},
+                    {"required", json::array({"name", "arguments", "id"})},
+                });
+            });
+            auto schema = json {
+                {"type", "array"},
+                {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
+                {"minItems", 1},
+            };
+            if (!inputs.parallel_tool_calls) {
+                schema["maxItems"] = 1;
+            }
+
+            builder.add_rule("root", "\"<|tool_call_start|>\"" + builder.add_schema("tool_calls", schema) + "\"<|tool_call_end|>\"");
+        });
+        // model has no concept of tool selection mode choice,
+        // if the system prompt rendered correctly it will produce a tool call
+        // the grammar goes inside the tool call body
+        data.grammar_lazy = true;
+        data.grammar_triggers = {{COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL, "\\s*<\\|tool_call_start\\|>\\s*\\["}};
+        data.preserved_tokens = {"<|tool_call_start|>", "<|tool_call_end|>"};
+        data.format = COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS;
+    } else if (are_tools_provided && (!is_json_schema_provided && !is_grammar_provided)) {
+        LOG_INF("%s: Using tools without json schema or grammar\n", __func__);
+        // output those tokens
+        data.preserved_tokens = {"<|tool_call_start|>", "<|tool_call_end|>"};
+    } else if (is_json_schema_provided) {
+        LOG_INF("%s: Using provided json schema to build a grammar\n", __func__);
+        data.grammar = json_schema_to_grammar(inputs.json_schema);
+    } else if (is_grammar_provided) {
+        LOG_INF("%s: Using provided grammar\n", __func__);
+        data.grammar = inputs.grammar;
+    } else {
+        LOG_INF("%s: Using content relying on the template\n", __func__);
+    }
+
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
+    LOG_DBG("%s: Prompt: %s\n", __func__, data.prompt.c_str());
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_magistral(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -1686,7 +1809,23 @@ static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
 
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs);
+
+    // Copy reasoning to the "thinking" field as expected by the gpt-oss template
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string();
+        auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array();
+
+        if (has_reasoning_content && has_tool_calls) {
+            auto adjusted_message = msg;
+            adjusted_message["thinking"] = msg.at("reasoning_content");
+            adjusted_messages.push_back(adjusted_message);
+        } else {
+            adjusted_messages.push_back(msg);
+        }
+    }
+
+    auto prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
 
     // Check if we need to replace the return token with end token during
     // inference and without generation prompt. For more details see:
@@ -2499,6 +2638,71 @@ static void common_chat_parse_apertus(common_chat_msg_parser & builder) {
     builder.add_content(builder.consume_rest());
 }
 
+
+static void common_chat_parse_lfm2(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // LFM2 format: <|tool_call_start|>[{"name": "get_current_time", "arguments": {"location": "Paris"}}]<|tool_call_end|>
+    static const common_regex tool_call_start_regex(regex_escape("<|tool_call_start|>"));
+    static const common_regex tool_call_end_regex(regex_escape("<|tool_call_end|>"));
+
+    // Loop through all tool calls
+    while (auto res = builder.try_find_regex(tool_call_start_regex, std::string::npos, /* add_prelude_to_content= */ true)) {
+        builder.move_to(res->groups[0].end);
+
+        // Parse JSON array format: [{"name": "...", "arguments": {...}}]
+        auto tool_calls_data = builder.consume_json();
+
+        // Consume end marker
+        builder.consume_spaces();
+        if (!builder.try_consume_regex(tool_call_end_regex)) {
+            throw common_chat_msg_partial_exception("Expected <|tool_call_end|>");
+        }
+
+        // Process each tool call in the array
+        if (tool_calls_data.json.is_array()) {
+            for (const auto & tool_call : tool_calls_data.json) {
+                if (!tool_call.is_object()) {
+                    throw common_chat_msg_partial_exception("Tool call must be an object");
+                }
+
+                if (!tool_call.contains("name")) {
+                    throw common_chat_msg_partial_exception("Tool call missing 'name' field");
+                }
+
+                std::string function_name = tool_call.at("name");
+                std::string arguments = "{}";
+
+                if (tool_call.contains("arguments")) {
+                    if (tool_call.at("arguments").is_object()) {
+                        arguments = tool_call.at("arguments").dump();
+                    } else if (tool_call.at("arguments").is_string()) {
+                        arguments = tool_call.at("arguments");
+                    }
+                }
+
+                if (!builder.add_tool_call(function_name, "", arguments)) {
+                    throw common_chat_msg_partial_exception("Incomplete tool call");
+                }
+            }
+        } else {
+            throw common_chat_msg_partial_exception("Expected JSON array for tool calls");
+        }
+
+        // Consume any trailing whitespace after this tool call
+        builder.consume_spaces();
+    }
+
+    // Consume any remaining content after all tool calls
+    auto remaining = builder.consume_rest();
+    if (!string_strip(remaining).empty()) {
+        builder.add_content(remaining);
+    }
+}
+
 static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) {
     // Parse thinking tags first - this handles the main reasoning content
     builder.try_parse_reasoning("<seed:think>", "</seed:think>");
@@ -2748,6 +2952,12 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_apertus(tmpl, params);
     }
 
+    // LFM2 (w/ tools)
+    if (src.find("List of tools: <|tool_list_start|>[") != std::string::npos &&
+        src.find("]<|tool_list_end|>") != std::string::npos) {
+        return common_chat_params_init_lfm2(tmpl, params);
+    }
+
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
     if ((params.tools.is_array() && params.json_schema.is_object())) {
@@ -2926,6 +3136,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_APERTUS:
             common_chat_parse_apertus(builder);
             break;
+        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS:
+            common_chat_parse_lfm2(builder);
+            break;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }

common/chat.h

@@ -116,6 +116,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_SEED_OSS,
     COMMON_CHAT_FORMAT_NEMOTRON_V2,
     COMMON_CHAT_FORMAT_APERTUS,
+    COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };

common/common.cpp

@@ -908,6 +908,39 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
+std::vector<common_file_info> fs_list_files(const std::string & path) {
+    std::vector<common_file_info> files;
+    if (path.empty()) return files;
+
+    std::filesystem::path dir(path);
+    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
+        return files;
+    }
+
+    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
+        try {
+            // Only include regular files (skip directories)
+            const auto & p = entry.path();
+            if (std::filesystem::is_regular_file(p)) {
+                common_file_info info;
+                info.path = p.string();
+                info.name = p.filename().string();
+                try {
+                    info.size = static_cast<size_t>(std::filesystem::file_size(p));
+                } catch (const std::filesystem::filesystem_error &) {
+                    info.size = 0;
+                }
+                files.push_back(std::move(info));
+            }
+        } catch (const std::filesystem::filesystem_error &) {
+            // skip entries we cannot inspect
+            continue;
+        }
+    }
+
+    return files;
+}
+
 
 //
 // Model utils

common/common.h

@@ -406,6 +406,8 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
+    int image_min_tokens = -1;
+    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -426,7 +428,7 @@ 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
     int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
-    int32_t cache_ram_mib     = 8192;         // 0 = no limit, 1 = 1 MiB, etc.
+    int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -458,7 +460,8 @@ struct common_params {
     float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
-    bool is_pp_shared = false;
+    bool is_pp_shared   = false;
+    bool is_tg_separate = false;
 
     std::vector<int32_t> n_pp;
     std::vector<int32_t> n_tg;
@@ -505,6 +508,10 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
+
+    bool has_speculative() const {
+        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
+    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -605,6 +612,13 @@ bool fs_create_directory_with_parents(const std::string & path);
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
 
+struct common_file_info {
+    std::string path;
+    std::string name;
+    size_t      size = 0; // in bytes
+};
+std::vector<common_file_info> fs_list_files(const std::string & path);
+
 //
 // Model utils
 //

common/download.h

@@ -0,0 +1,55 @@
+#pragma once
+
+#include <string>
+
+struct common_params_model;
+
+//
+// download functionalities
+//
+
+struct common_cached_model_info {
+    std::string manifest_path;
+    std::string user;
+    std::string model;
+    std::string tag;
+    size_t      size = 0; // GGUF size in bytes
+    std::string to_string() const {
+        return user + "/" + model + ":" + tag;
+    }
+};
+
+struct common_hf_file_res {
+    std::string repo; // repo name with ":tag" removed
+    std::string ggufFile;
+    std::string mmprojFile;
+};
+
+/**
+ * Allow getting the HF file from the HF repo with tag (like ollama), for example:
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
+ * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
+ *
+ * Return pair of <repo, file> (with "repo" already having tag removed)
+ *
+ * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
+ */
+common_hf_file_res common_get_hf_file(
+    const std::string & hf_repo_with_tag,
+    const std::string & bearer_token,
+    bool offline);
+
+// returns true if download succeeded
+bool common_download_model(
+    const common_params_model & model,
+    const std::string & bearer_token,
+    bool offline);
+
+// returns list of cached models
+std::vector<common_cached_model_info> common_list_cached_models();
+
+// resolve and download model from Docker registry
+// return local path to downloaded model file
+std::string common_docker_resolve_model(const std::string & docker);

common/json-partial.cpp

@@ -5,6 +5,7 @@
 #include <nlohmann/json.hpp>
 
 #include <string>
+#include <regex>
 
 using json = nlohmann::ordered_json;
 
@@ -168,6 +169,47 @@ bool common_json_parse(
                 }
             }
 
+            // Matches a potentially partial unicode escape sequence, e.g. \u, \uX, \uXX, \uXXX, \uXXXX
+            static const std::regex partial_unicode_regex(R"(\\u(?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F])?)?)?)?$)");
+
+            auto is_high_surrogate = [&](const std::string & s) {
+                // Check if a partial of a high surrogate (U+D800-U+DBFF)
+                return s.length() >= 4 &&
+                    s[0] == '\\' && s[1] == 'u' &&
+                    std::tolower(s[2]) == 'd' &&
+                    (s[3] == '8' || s[3] == '9' || std::tolower(s[3]) == 'a' || std::tolower(s[3]) == 'b');
+            };
+
+            // Initialize the unicode marker to a low surrogate to handle the edge case
+            // where a high surrogate (U+D800-U+DBFF) is immediately followed by a
+            // backslash (\)
+            std::string unicode_marker_padding = "udc00";
+            std::smatch last_unicode_seq;
+
+            if (std::regex_search(str, last_unicode_seq, partial_unicode_regex)) {
+                std::smatch second_last_seq;
+                std::string prelude = str.substr(0, last_unicode_seq.position());
+
+                // Pad the escape sequence with 0s until it forms a complete sequence of 6 characters
+                unicode_marker_padding = std::string(6 - last_unicode_seq.length(), '0');
+
+                if (is_high_surrogate(last_unicode_seq.str())) {
+                    // If the sequence is a partial match for a high surrogate, add a low surrogate (U+DC00-U+UDFF)
+                    unicode_marker_padding += "\\udc00";
+                } else if (std::regex_search(prelude, second_last_seq, partial_unicode_regex)) {
+                    if (is_high_surrogate(second_last_seq.str())) {
+                        // If this follows a high surrogate, pad it to be a low surrogate
+                        if (last_unicode_seq.length() == 2) {
+                            unicode_marker_padding = "dc00";
+                        } else if (last_unicode_seq.length() == 3) {
+                            unicode_marker_padding = "c00";
+                        } else {
+                            // The original unicode_marker_padding is already padded with 0s
+                        }
+                    }
+                }
+            }
+
             const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
 
             if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
@@ -186,6 +228,9 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
                     // Was inside an object value string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
+                    // Was inside an object value string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
                 } else {
                     // find last :
                     auto last_pos = str.find_last_of(':');
@@ -205,6 +250,9 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
                     // Was inside an array value string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
+                    // Was inside an array value string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
                 } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
                     // Had just finished a value
                     str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
@@ -230,6 +278,9 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
                     // Was inside an object key string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\": 1" + closing)) {
+                    // Was inside an object key string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\": 1" + closing;
                 } else {
                     auto last_pos = str.find_last_of(':');
                     if (last_pos == std::string::npos) {

common/json-schema-to-grammar.cpp

@@ -41,9 +41,9 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
-    auto has_min = min_value != std::numeric_limits<int>::min();
-    auto has_max = max_value != std::numeric_limits<int>::max();
+static void _build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
+    auto has_min = min_value != std::numeric_limits<int64_t>::min();
+    auto has_max = max_value != std::numeric_limits<int64_t>::max();
 
     auto digit_range = [&](char from, char to) {
         out << "[";
@@ -159,7 +159,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
     if (has_min) {
         if (min_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
             out << ") | [0] | [1-9] ";
             more_digits(0, decimals_left - 1);
         } else if (min_value == 0) {
@@ -194,7 +194,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
             }
             digit_range(c, c);
             out << " (";
-            _build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
+            _build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
             out << ")";
             if (c < '9') {
                 out << " | ";
@@ -216,7 +216,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
             _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
         } else {
             out << "\"-\" (";
-            _build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
             out << ")";
         }
         return;
@@ -601,7 +601,10 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        auto it = ref.find('#');
+        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
+        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
+        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -774,11 +777,24 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_null() || !target.contains(sel)) {
+                            if (target.is_object() && target.contains(sel)) {
+                                target = target[sel];
+                            } else if (target.is_array()) {
+                                size_t sel_index;
+                                try {
+                                    sel_index = std::stoul(sel);
+                                } catch (const std::invalid_argument & e) {
+                                    sel_index = target.size();
+                                }
+                                if (sel_index >= target.size()) {
+                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                    return;
+                                }
+                                target = target[sel_index];
+                            } else {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
-                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }
@@ -925,17 +941,17 @@ public:
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
         } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
-            int min_value = std::numeric_limits<int>::min();
-            int max_value = std::numeric_limits<int>::max();
+            int64_t min_value = std::numeric_limits<int64_t>::min();
+            int64_t max_value = std::numeric_limits<int64_t>::max();
             if (schema.contains("minimum")) {
-                min_value = schema["minimum"].get<int>();
+                min_value = schema["minimum"].get<int64_t>();
             } else if (schema.contains("exclusiveMinimum")) {
-                min_value = schema["exclusiveMinimum"].get<int>() + 1;
+                min_value = schema["exclusiveMinimum"].get<int64_t>() + 1;
             }
             if (schema.contains("maximum")) {
-                max_value = schema["maximum"].get<int>();
+                max_value = schema["maximum"].get<int64_t>();
             } else if (schema.contains("exclusiveMaximum")) {
-                max_value = schema["exclusiveMaximum"].get<int>() - 1;
+                max_value = schema["exclusiveMaximum"].get<int64_t>() - 1;
             }
             std::stringstream out;
             out << "(";

convert_hf_to_gguf_update.py

@@ -139,8 +139,9 @@ models = [
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
-    {"name": "llada-moe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
+    {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
+    {"name": "minimax-m2",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/MiniMaxAI/MiniMax-M2", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -435,7 +436,7 @@ for model in models:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
         else:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
+    except (OSError, TypeError) as e:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
 

docs/backend/OPENCL.md

@@ -39,18 +39,23 @@ The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adren
 | Adreno 830 (Snapdragon 8 Elite)      | Support |
 | Adreno X85 (Snapdragon X Elite)      | Support |
 
+> A6x GPUs with a recent driver and compiler are supported; they are usually found in IoT platforms.
+However, A6x GPUs in phones are likely not supported due to the outdated driver and compiler.
+
 ## DataType Supports
 
 | DataType               | Status                     |
 |:----------------------:|:--------------------------:|
 | Q4_0                   | Support                    |
 | Q6_K                   | Support, but not optimized |
+| Q8_0                   | Support                    |
+| MXFP4                  | Support                    |
 
 ## Model Preparation
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration.
+You can refer to the general [llama-quantize tool](/tools/quantize/README.md) for steps to convert a model in Hugging Face safetensor format to GGUF with quantization.
 
-Currently we support `Q4_0` quantization and have optimize for it. To achieve best performance on Adreno GPU, add `--pure` to `llama-quantize`. For example,
+Currently we support `Q4_0` quantization and have optimized for it. To achieve best performance on Adreno GPU, add `--pure` to `llama-quantize` (i.e., make all weights in `Q4_0`). For example,
 
 ```sh
 ./llama-quantize --pure ggml-model-qwen2.5-3b-f16.gguf ggml-model-qwen-3b-Q4_0.gguf Q4_0
@@ -58,6 +63,17 @@ Currently we support `Q4_0` quantization and have optimize for it. To achieve be
 
 Since `Q6_K` is also supported, `Q4_0` quantization without `--pure` will also work. However, the performance will be worse compared to pure `Q4_0` quantization.
 
+### `MXFP4` MoE Models
+
+OpenAI gpt-oss models are MoE models in `MXFP4`. The quantized model will be in `MXFP4_MOE`, a mixture of `MXFP4` and `Q8_0`.
+For this quantization, there is no need to specify `--pure`.
+For gpt-oss-20b model, you can directly [download](https://huggingface.co/ggml-org/gpt-oss-20b-GGUF) the quantized GGUF file in `MXFP4_MOE` from Hugging Face.
+
+Although it is possible to quantize gpt-oss-20b model in pure `Q4_0` (all weights in `Q4_0`), it is not recommended since `MXFP4` has been optimized for MoE while `Q4_0` is not. In addition, accuracy should degrade with such pure `Q4_0` quantization.
+Hence, using the default `MXFP4_MOE` quantization (see the link above) is recommended for this model.
+
+> Note that the `Q4_0` model found [here](https://huggingface.co/unsloth/gpt-oss-20b-GGUF/blob/main/gpt-oss-20b-Q4_0.gguf) is a mixture of `Q4_0`, `Q8_0` and `MXFP4` and gives better performance than `MXFP4_MOE` quantization.
+
 ## CMake Options
 
 The OpenCL backend has the following CMake options that control the behavior of the backend.
@@ -146,10 +162,13 @@ A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the followi
 * Ninja
 * Visual Studio 2022
 * Powershell 7
+* Python
 
 Visual Studio provides necessary headers and libraries although it is not directly used for building.
 Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
 
+> Note that building using Visual Studio's cl compiler is not supported. Clang must be used. Clang depends on libraries provided by Visual Studio to work. Therefore, Visual Studio must be installed. Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
+
 Powershell 7 is used for the following commands.
 If an older version of Powershell is used, these commands may not work as they are.
 
@@ -201,9 +220,12 @@ ninja
 
 ## Known Issues
 
-- Currently OpenCL backend does not work on Adreno 6xx GPUs.
+- Flash attention does not always improve performance.
+- Currently OpenCL backend works on A6xx GPUs with recent drivers and compilers (usually found in IoT platforms).
+  However, it does not work on A6xx GPUs found in phones with old drivers and compilers.
 
 ## TODO
 
 - Optimization for Q6_K
 - Support and optimization for Q4_K
+- Improve flash attention

docs/backend/hexagon/CMakeUserPresets.json

@@ -0,0 +1,49 @@
+{
+  "version": 4,
+  "configurePresets": [
+    {
+        "name": "arm64-android-snapdragon",
+        "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "ANDROID_ABI":      "arm64-v8a",
+            "ANDROID_PLATFORM": "android-31",
+            "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
+            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
+            "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
+            "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
+            "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
+            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "PREBUILT_LIB_DIR": "android_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "ON",
+            "GGML_HEXAGON":     "ON",
+            "LLAMA_CURL":       "OFF"
+        }
+    },
+
+    {
+        "name": "arm64-windows-snapdragon",
+        "inherits": [ "base", "arm64-windows-llvm" ],
+        "cacheVariables": {
+            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "PREBUILT_LIB_DIR": "windows_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "ON",
+            "GGML_HEXAGON":     "ON",
+            "LLAMA_CURL":       "OFF"
+        }
+    },
+
+    { "name": "arm64-android-snapdragon-debug"  , "inherits": [ "base", "arm64-android-snapdragon", "debug" ] },
+    { "name": "arm64-android-snapdragon-release", "inherits": [ "base", "arm64-android-snapdragon", "release" ] },
+
+    { "name": "arm64-windows-snapdragon-debug"  , "inherits": [ "base", "arm64-windows-snapdragon", "debug" ] },
+    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] }
+  ]
+}

docs/backend/hexagon/README.md

@@ -0,0 +1,239 @@
+# Snapdragon-based Android devices
+
+## How to Build
+
+The easiest way to build llama.cpp for a Snapdragon-based Android device is using the toolchain Docker image (see github.com/snapdragon-toolchain).
+This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
+
+This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
+
+```
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.3
+[d]/> cd /workspace
+```
+
+The rest of the Android build process assumes that you're running inside the toolchain container.
+Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
+
+```
+[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
+
+[d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
+Preset CMake variables:
+  ANDROID_ABI="arm64-v8a"
+  ...
+  CMAKE_TOOLCHAIN_FILE="/opt/android-ndk-r28b/build/cmake/android.toolchain.cmake"
+  GGML_HEXAGON="ON"
+  GGML_OPENCL="ON"
+  GGML_OPENMP="OFF"
+  HEXAGON_SDK_ROOT="/opt/hexagon/6.4.0.2"
+...
+-- Including OpenCL backend
+-- Including Hexagon backend
+...
+-- Build files have been written to: /workspace/build-snapdragon
+
+[d]/workspace> cmake --build build-snapdragon
+...
+[144/356] Performing build step for 'htp-v73'
+[1/16] Generating htp_iface_skel.c, htp_iface_stub.c, htp_iface.h
+[2/16] Building C object CMakeFiles/ggml-htp-v73.dir/hvx-sigmoid.c.obj
+[3/16] Building C object CMakeFiles/ggml-htp-v73.dir/htp-dma.c.obj
+[4/16] Building C object CMakeFiles/ggml-htp-v73.dir/worker-pool.c.obj
+...
+-- Installing: /workspace/build-snapdragon/ggml/src/ggml-hexagon/libggml-htp-v73.so
+-- Installing: /workspace/build-snapdragon/ggml/src/ggml-hexagon/libggml-htp-v75.so
+...
+```
+
+To generate an installable "package" simply use cmake --install:
+
+```
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-adb/llama.cpp
+-- Install configuration: "Release"
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-cpu.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-opencl.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-hexagon.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v73.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v75.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v79.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v81.so
+-- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml.so
+...
+-- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-bench
+-- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-cli
+...
+```
+
+## How to Install
+
+For this step, your device needs to be configured for on-device development.
+Please see https://developer.android.com/studio/debug/dev-options for details.
+
+Once ADB is enabled, use `adb push` to install `pkg-snapdragon` on the device.
+**Note that the toolchain Docker image doesn't have ADB and doesn't set up the ADB bridge. Please use native ADB on the host.**
+
+```
+~/src/llama.cpp$ adb push pkg-adb/llama.cpp /data/local/tmp/
+pkg-adb/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
+pkg-adb/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
+pkg-adb/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
+102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
+```
+
+At this point, you should also install some models:
+
+```
+~/src/llama.cpp$ wget https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf
+...
+2025-10-11 12:04:52 (10.7 MB/s) - ‘Llama-3.2-1B-Instruct-Q4_0.gguf’ saved [773025920/773025920]
+
+~/src/llama.cpp$ adb push Llama-3.2-1B-Instruct-Q4_0.gguf /data/local/tmp/gguf
+Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920 bytes in 19.250s)
+```
+
+## How to Run
+
+The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
+
+llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU (HTP0-4).
+You can select which backend to run the model on using the `D=` variable, which maps to the `--device` option.
+
+Hexagon NPU behaves as a "GPU" device when it comes to `-ngl` and other offload-related options.
+
+Here are some examples of running various llama.cpp tools via ADB.
+
+Simple question for Llama-3.2-1B
+
+```
+~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-cli.sh -no-cnv -p "what is the most popular cookie in the world?"
+...
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v79
+ggml-hex: allocating new session: HTP0
+ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v79.so?htp_iface_skel_handle_invoke&_modver=1.0&_dom=cdsp&_session=0 handle 0xb4000072c7955e50
+...
+load_tensors: offloading output layer to GPU
+load_tensors: offloaded 17/17 layers to GPU
+load_tensors:          CPU model buffer size =   225.49 MiB
+load_tensors:         HTP0 model buffer size =     0.26 MiB
+load_tensors:  HTP0-REPACK model buffer size =   504.00 MiB
+...
+I hope this helps you understand the world's most popular cookies! [end of text]
+...
+llama_perf_sampler_print:    sampling time =      30.08 ms /   487 runs   (    0.06 ms per token, 16191.77 tokens per second)
+llama_perf_context_print:        load time =     617.94 ms
+llama_perf_context_print: prompt eval time =      80.76 ms /    11 tokens (    7.34 ms per token,   136.21 tokens per second)
+llama_perf_context_print:        eval time =    9210.59 ms /   475 runs   (   19.39 ms per token,    51.57 tokens per second)
+llama_perf_context_print:       total time =    9454.92 ms /   486 tokens
+llama_perf_context_print:    graphs reused =        473
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                  439 =   225 +     136 +      77                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                  504 =   504 +       0 +       0                |
+```
+
+Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
+
+```
+~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-cli.sh -f surfing.txt -no-cnv
+...
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v81
+ggml-hex: allocating new session: HTP0
+ggml-hex: allocating new session: HTP1
+...
+load_tensors: offloading output layer to GPU
+load_tensors: offloaded 17/17 layers to GPU
+load_tensors:          CPU model buffer size =   143.86 MiB
+load_tensors:         HTP1 model buffer size =     0.23 MiB
+load_tensors:  HTP1-REPACK model buffer size =  1575.00 MiB
+load_tensors:         HTP0 model buffer size =     0.28 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2025.00 MiB
+...
+llama_context:        CPU  output buffer size =     0.19 MiB
+llama_kv_cache:       HTP1 KV buffer size =   238.00 MiB
+llama_kv_cache:       HTP0 KV buffer size =   306.00 MiB
+llama_kv_cache: size =  544.00 MiB (  8192 cells,  16 layers,  1/1 seqs), K (q8_0):  272.00 MiB, V (q8_0):  272.00 MiB
+llama_context:       HTP0 compute buffer size =    15.00 MiB
+llama_context:       HTP1 compute buffer size =    15.00 MiB
+llama_context:        CPU compute buffer size =    24.56 MiB
+...
+llama_perf_context_print: prompt eval time =    1730.57 ms /   212 tokens (    8.16 ms per token,   122.50 tokens per second)
+llama_perf_context_print:        eval time =    5624.75 ms /   257 runs   (   21.89 ms per token,    45.69 tokens per second)
+llama_perf_context_print:       total time =    7377.33 ms /   469 tokens
+llama_perf_context_print:    graphs reused =        255
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                  742 =   144 +     544 +      54                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 1575 =  1575 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2025 =  2025 +       0 +       0                |
+```
+
+Op test for MUL_MAT
+
+```
+~/src/llama.cpp$ HB=0 ./scripts/snapdragon/adb/run-tool.sh test-backend-ops -b HTP0 -o MUL_MAT
+...
+Backend 2/3: HTP0
+Device description: Hexagon
+Device memory: 2048 MB (2048 MB free)
+MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
+MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
+MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
+
+~/src/llama.cpp-hexagon$ M=Llama-3.2-1B-Instruct-Q4_0.gguf ./scripts/snapdragon/adb/run-bench.sh -p 128 -n 64
+...
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v79
+ggml-hex: allocating new session: HTP0
+ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v79.so?htp_iface_skel_handle_invoke&_modver=1.0&_dom=cdsp&_session=0 handle 0xb400007d4b231090
+| model          |       size | params | backend    | ngl | threads | n_batch | mmap |  test |           t/s |
+| ---------------| ---------: | -----: | ---------- | --: | ------: | ------: | ---: | ----: | ------------: |
+| llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 | pp128 | 169.42 ± 1.75 |
+| llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 |  tg64 |  51.54 ± 1.13 |
+
+build: 6a8cf8914 (6733)
+```
+
+## Environment variables
+
+- `GGML_HEXAGON_NDEV=1`
+  Controls the number of devices/sessions to allocate. The default is 1.
+  Most quantized models under 4B fit into a single session; an 8B model needs two, and a 20B model needs four.
+
+- `GGML_HEXAGON_NHVX=0`
+  Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
+
+- `GGML_HEXAGON_HOSTBUF=1`
+  Controls whether the Hexagon backend allocates host buffers. By default, all buffers except for REPACK are host buffers.
+  This option is required for testing Ops that require REPACK buffers (MUL_MAT and MUL_MAT_ID).
+
+- `GGML_HEXAGON_VERBOSE=1`
+  Enables verbose logging of Ops from the backend. Example output:
+
+  ```
+  ggml-hex: HTP0 graph-compute n_nodes 2
+  ggml-hex: HTP0 matmul : blk.27.ffn_up.weight x ffn_norm-27 -> ffn_up-27 : 3072:8192 x 3072:1 -> 8192:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x1
+  ggml-hex: HTP0 matmul : blk.27.ffn_gate.weight x ffn_norm-27 -> ffn_gate-27 : 3072:8192 x 3072:1 -> 8192:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x3
+  ggml-hex: HTP0 graph-compute n_nodes 1
+  ggml-hex: HTP0 matmul : blk.27.ffn_down.weight x ffn_gate_par-27 -> ffn_out-27 : 8192:3072 x 8192:1 -> 3072:1 : q4_0 x f32 -> f32 : HTP0 x HTP0 -> HTP0 : flags 0x0
+  ggml-hex: HTP0 get-tensor result_output : data 0x7592487000 offset 0 size 513024
+  ```
+
+- `GGML_HEXAGON_PROFILE=1`
+  Generates a host-side profile for the ggml-hexagon Ops.
+
+- `GGML_HEXAGON_OPMASK=0x0`
+  Allows enabling specific stages of the processing pipeline:
+
+  - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
+  - `0x2` Enable Dynamic Quantizer (if needed for the Op)
+  - `0x4` Enable Op Compute (MUL_MAT, etc.)
+
+  Examples:
+
+      `GGML_HEXAGON_OPMASK=0x1 llama-cli ...` - Ops are enqueued but NPU-side processing is stubbed out
+      `GGML_HEXAGON_OPMASK=0x3 llama-cli ...` - NPU performs dynamic quantization and skips the rest
+      `GGML_HEXAGON_OPMASK=0x7 llama-cli ...` - Full queuing and processing of Ops (default)

docs/backend/hexagon/developer.md

@@ -0,0 +1,109 @@
+# Hexagon backend developer details
+
+## Backend libraries
+
+The Hexagon backend consist of two parts:
+
+  - `libggml-hexagon`
+    This is the regular CPU-side GGML backend library, either shared or statically linked
+
+  - `libggml-htp-vNN`
+    This is the NPU-side (HTP stands for Hexagon Tensor Processor) shared library that contains the Op dispatcher and kernels.
+    The correct library is selected automatically at runtime based on the HW version.
+
+Here is an example of the build artifacts
+
+```
+~/src/llama.cpp$ ls -l pkg-adb/llama.cpp/lib/libggml*
+pkg-adb/llama.cpp/lib/libggml-base.so
+pkg-adb/llama.cpp/lib/libggml-cpu.so
+pkg-adb/llama.cpp/lib/libggml-hexagon.so      <<< CPU library
+pkg-adb/llama.cpp/lib/libggml-htp-v73.so      <<< HTP op/kernels for Hexagon v73
+pkg-adb/llama.cpp/lib/libggml-htp-v75.so
+pkg-adb/llama.cpp/lib/libggml-htp-v79.so
+pkg-adb/llama.cpp/lib/libggml-htp-v81.so
+```
+
+## Memory buffers
+
+Hexagon NPU backend takes advantage of the Snapdragon's unified memory model where all buffers are fully accessible by the CPU and GPU.
+The NPU does have a dedicated tightly-coupled memory called VTCM but that memory is used only for intermediate data (e.g. dynamically
+quantized tensors) or temporary data (chunks of the weight tensors fetched via DMA).
+
+Please note that currently the Hexagon backend does not implement SET/GET_ROWS Ops because there is no advantage in offloading those
+to the NPU at this point.
+
+The backend does allocates non-host buffers for the tensors with datatypes that require repacking: Q4_0, Q8_0, MXFP4.
+From the MMU perspective these buffers are still regular buffers (normal access by the CPU) they are marked as non-host simply to force
+the repacking.
+
+## Large model handling
+
+Hexagon NPU session (aka Process Domain (PD) in the Hexagon docs) is limited to a memory mapping of around 3.5GB.
+In llama.cpp/GGML the Hexagon session is mapped to a single GGML backend device (HTP0, HTP1, etc).
+
+In order to map models larger than 3.5GB we need to allocate multiple devices and split the model.
+For this we're taking advantage of the llama.cpp/GGML multi-GPU layer-splitting support.
+Each Hexagon device behaves like a GPU from the offload and model splitting perspective.
+
+Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
+
+```
+M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-cli.sh -no-cnv -f surfing.txt -n 32
+...
+LD_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+ADSP_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+GGML_HEXAGON_NDEV=4 ./bin/llama-cli --no-mmap -m /data/local/tmp/llama.cpp/../gguf/gpt-oss-20b-Q4_0.gguf
+      -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 --device HTP0,HTP1,HTP2,HTP3 -no-cnv -f surfing.txt
+...
+llama_model_loader: - type  f32:  289 tensors
+llama_model_loader: - type q4_0:   96 tensors
+llama_model_loader: - type q8_0:    2 tensors
+llama_model_loader: - type mxfp4:  72 tensors
+...
+load_tensors: offloaded 25/25 layers to GPU
+load_tensors:          CPU model buffer size =  1182.09 MiB
+load_tensors:         HTP1 model buffer size =     6.64 MiB
+load_tensors:  HTP1-REPACK model buffer size =  2505.94 MiB
+load_tensors:         HTP3 model buffer size =     5.55 MiB
+load_tensors:  HTP3-REPACK model buffer size =  2088.28 MiB
+load_tensors:         HTP0 model buffer size =     7.75 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2923.59 MiB
+load_tensors:         HTP2 model buffer size =     6.64 MiB
+load_tensors:  HTP2-REPACK model buffer size =  2505.94 MiB
+...
+llama_context: n_ctx_per_seq (8192) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
+llama_context:        CPU  output buffer size =     0.77 MiB
+llama_kv_cache_iswa: creating non-SWA KV cache, size = 8192 cells
+llama_kv_cache:       HTP1 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP3 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP0 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP2 KV buffer size =    25.50 MiB
+llama_kv_cache: size =  102.00 MiB (  8192 cells,  12 layers,  1/1 seqs), K (q8_0):   51.00 MiB, V (q8_0):   51.00 MiB
+llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
+llama_kv_cache:       HTP1 KV buffer size =     0.80 MiB
+llama_kv_cache:       HTP3 KV buffer size =     0.53 MiB
+llama_kv_cache:       HTP0 KV buffer size =     1.06 MiB
+llama_kv_cache:       HTP2 KV buffer size =     0.80 MiB
+llama_kv_cache: size =    3.19 MiB (   256 cells,  12 layers,  1/1 seqs), K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
+llama_context:       HTP0 compute buffer size =    16.06 MiB
+llama_context:       HTP1 compute buffer size =    16.06 MiB
+llama_context:       HTP2 compute buffer size =    16.06 MiB
+llama_context:       HTP3 compute buffer size =    16.06 MiB
+llama_context:        CPU compute buffer size =    98.19 MiB
+...
+llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
+llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.39 ms per token, 18.39 tokens per second)
+llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
+llama_perf_context_print:    graphs reused =         30
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP2 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP3 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 2505 =  2505 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP3-REPACK        |                 2088 =  2088 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2923 =  2923 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP2-REPACK        |                 2505 =  2505 +       0 +       0                |
+```

docs/build.md

@@ -178,6 +178,48 @@ GeForce RTX 3070      8.6
 cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="86;89"
 ```
 
+### Overriding the CUDA Version
+
+If you have multiple CUDA installations on your system and want to compile llama.cpp for a specific one, e.g. for CUDA 11.7 installed under `/opt/cuda-11.7`:
+
+```bash
+cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_COMPILER=/opt/cuda-11.7/bin/nvcc -DCMAKE_INSTALL_RPATH="/opt/cuda-11.7/lib64;\$ORIGIN" -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
+```
+
+#### Fixing Compatibility Issues with Old CUDA and New glibc
+
+If you try to use an old CUDA version (e.g. v11.7) with a new glibc version you can get errors like this:
+
+```
+/usr/include/bits/mathcalls.h(83): error: exception specification is
+  incompatible with that of previous function "cospi"
+
+
+  /opt/cuda-11.7/bin/../targets/x86_64-linux/include/crt/math_functions.h(5545):
+  here
+```
+
+It seems the least bad solution is to patch the CUDA installation to declare the correct signatures.
+Replace the following lines in `/path/to/your/cuda/installation/targets/x86_64-linux/include/crt/math_functions.h`:
+
+```C++
+// original lines
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 cospi(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  cospif(float x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 sinpi(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  sinpif(float x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x);
+
+// edited lines
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 cospi(double x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  cospif(float x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 sinpi(double x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  sinpif(float x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x) noexcept (true);
+```
+
 ### Runtime CUDA environmental variables
 
 You may set the [cuda environmental variables](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) at runtime.
@@ -261,10 +303,12 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
 
+  Note: `GPU_TARGETS` is optional, omitting it will build the code for all GPUs in the current system.
+
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
   The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager.
@@ -282,17 +326,17 @@ You can download it from your Linux distro's package manager or from here: [ROCm
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
   HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build -- -j 16
   ```
 
 - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
   ```bash
   set PATH=%HIP_PATH%\bin;%PATH%
-  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+  cmake -S . -B build -G Ninja -DGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
   cmake --build build
   ```
-  Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+  If necessary, adapt `GPU_TARGETS` to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
   Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
 
 

docs/docker.md

@@ -7,9 +7,9 @@
 ## Images
 We have three Docker images available for this project:
 
-1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
 
 Additionally, there the following images, similar to the above:
 

docs/ops.md

@@ -22,16 +22,17 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -41,6 +42,7 @@ Legend:
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -51,7 +53,7 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
@@ -65,23 +67,24 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -92,19 +95,22 @@ Legend:
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
+|                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

docs/ops/CPU.csv

@@ -59,6 +59,14 @@
 "CPU","EXP","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
+"CPU","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","ABS","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
 "CPU","ABS","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","SGN","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
@@ -119,6 +127,14 @@
 "CPU","EXP","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
 "CPU","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
+"CPU","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
+"CPU","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","1","yes","CPU"
 "CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","1","yes","CPU"

docs/ops/Vulkan.csv

@@ -3263,27 +3263,27 @@
 "Vulkan0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=0","support","1","yes","Vulkan"
 "Vulkan0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=1","support","1","yes","Vulkan"
 "Vulkan0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","0","no","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
-"Vulkan0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
+"Vulkan0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=1,n_seqs=1","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=1","support","1","yes","Vulkan"
 "Vulkan0","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=4","support","1","yes","Vulkan"

examples/embedding/README.md

@@ -38,6 +38,7 @@ The above command will output space-separated float values.
 |            | multiple embeddings          | $[[x_1,...,x_n],[x_1,...,x_n],...,[x_1,...,x_n]]$
 | 'json'     | openai style                 |
 | 'json+'    | add cosine similarity matrix |
+| 'raw'      | plain text output            |
 
 ### --embd-separator $"string"$
 | $"string"$   | |

examples/embedding/embedding.cpp

@@ -70,6 +70,29 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     }
 }
 
+// plain, pipe-friendly output: one embedding per line
+static void print_raw_embeddings(const float * emb,
+                                 int n_embd_count,
+                                 int n_embd,
+                                 const llama_model * model,
+                                 enum llama_pooling_type pooling_type,
+                                 int embd_normalize) {
+    const uint32_t n_cls_out = llama_model_n_cls_out(model);
+    const bool is_rank = (pooling_type == LLAMA_POOLING_TYPE_RANK);
+    const int cols = is_rank ? std::min<int>(n_embd, (int) n_cls_out) : n_embd;
+
+    for (int j = 0; j < n_embd_count; ++j) {
+        for (int i = 0; i < cols; ++i) {
+            if (embd_normalize == 0) {
+                LOG("%1.0f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
+            } else {
+                LOG("%1.7f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
+            }
+        }
+        LOG("\n");
+    }
+}
+
 int main(int argc, char ** argv) {
     common_params params;
 
@@ -372,6 +395,8 @@ int main(int argc, char ** argv) {
         }
 
         if (notArray) LOG("\n}\n");
+    } else if (params.embd_out == "raw") {
+        print_raw_embeddings(emb, n_embd_count, n_embd, model, pooling_type, params.embd_normalize);
     }
 
     LOG("\n");

examples/gguf/gguf.cpp

@@ -184,8 +184,13 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
             const size_t size   = gguf_get_tensor_size  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
+            const auto   type   = gguf_get_tensor_type  (ctx, i);
 
-            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
+            const char * type_name  = ggml_type_name(type);
+            const size_t type_size  = ggml_type_size(type);
+            const size_t n_elements = size / type_size;
+
+            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu, type = %s, n_elts = %zu\n", __func__, i, name, size, offset, type_name, n_elements);
         }
     }
 

examples/json_schema_to_grammar.py

@@ -371,8 +371,17 @@ class SchemaConverter:
                         raise ValueError(f'Unsupported ref {ref}')
 
                     for sel in ref.split('#')[-1].split('/')[1:]:
-                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
-                        target = target[sel]
+                        assert target is not None, f'Error resolving ref {ref}: {sel} not in {target}'
+                        if isinstance(target, list):
+                            try:
+                                sel_index = int(sel)
+                            except ValueError:
+                                raise ValueError(f'Error resolving ref {ref}: {sel} not in {target}')
+                            assert 0 <= sel_index < len(target), f'Error resolving ref {ref}: {sel} not in {target}'
+                            target = target[sel_index]
+                        else:
+                            assert sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                            target = target[sel]
 
                     self._refs[ref] = target
                 else:
@@ -547,7 +556,8 @@ class SchemaConverter:
 
 
     def _resolve_ref(self, ref):
-        ref_name = ref.split('/')[-1]
+        ref_fragment = ref.split('#')[-1]
+        ref_name = 'ref' + re.sub(r'[^a-zA-Z0-9-]+', '-', ref_fragment)
         if ref_name not in self._rules and ref not in self._refs_being_resolved:
             self._refs_being_resolved.add(ref)
             resolved = self._refs[ref]

examples/model-conversion/scripts/causal/run-org-model.py

@@ -138,7 +138,10 @@ if model_path is None:
         "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
     )
 
-config = AutoConfig.from_pretrained(model_path)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
 
 print("Model type:       ", config.model_type)
 print("Vocab size:       ", config.vocab_size)
@@ -147,10 +150,6 @@ print("Number of layers: ", config.num_hidden_layers)
 print("BOS token id:     ", config.bos_token_id)
 print("EOS token id:     ", config.eos_token_id)
 
-print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path)
-config = AutoConfig.from_pretrained(model_path)
-
 if unreleased_model_name:
     model_name_lower = unreleased_model_name.lower()
     unreleased_module_path = (
@@ -171,7 +170,7 @@ if unreleased_model_name:
         exit(1)
 else:
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", offload_folder="offload"
+        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=config
     )
 
 for name, module in model.named_modules():

ggml/CMakeLists.txt

@@ -168,7 +168,7 @@ option(GGML_RV_ZFH           "ggml: enable riscv zfh"        ON)
 option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
 option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
-option(GGML_VXE              "ggml: enable vxe"              ON)
+option(GGML_VXE              "ggml: enable vxe"              ${GGML_NATIVE})
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
@@ -251,6 +251,8 @@ option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adr
 set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
                                             "gmml: OpenCL API version to target")
 
+option(GGML_HEXAGON                         "ggml: enable Hexagon backend"                    OFF)
+
 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
 

ggml/include/ggml-hexagon.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_BACKEND_API ggml_backend_t ggml_backend_hexagon_init(void);
+
+GGML_BACKEND_API bool ggml_backend_is_hexagon(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_hexagon_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-rpc.h

@@ -21,8 +21,7 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const c
 GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total);
 
 GGML_BACKEND_API void ggml_backend_rpc_start_server(const char * endpoint, const char * cache_dir,
-                                                    size_t n_threads, size_t n_devices,
-                                                    ggml_backend_dev_t * devices, size_t * free_mem, size_t * total_mem);
+                                                    size_t n_threads, size_t n_devices, ggml_backend_dev_t * devices);
 
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_add_server(const char * endpoint);

ggml/include/ggml.h

@@ -242,6 +242,7 @@
 #define GGML_ROPE_TYPE_NEOX   2
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
+#define GGML_ROPE_TYPE_IMROPE 40 // binary: 101000
 
 #define GGML_MROPE_SECTIONS   4
 
@@ -577,6 +578,10 @@ extern "C" {
         GGML_UNARY_OP_EXP,
         GGML_UNARY_OP_GELU_ERF,
         GGML_UNARY_OP_XIELU,
+        GGML_UNARY_OP_FLOOR,
+        GGML_UNARY_OP_CEIL,
+        GGML_UNARY_OP_ROUND,
+        GGML_UNARY_OP_TRUNC,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -1151,6 +1156,46 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_floor(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_floor_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_ceil(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_ceil_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_round(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_round_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+     /**
+     * Truncates the fractional part of each element in the tensor (towards zero).
+     * For example: trunc(3.7) = 3.0, trunc(-2.9) = -2.0
+     * Similar to std::trunc in C/C++.
+     */
+
+    GGML_API struct ggml_tensor * ggml_trunc(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_trunc_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+
+
     // xIELU activation function
     // x = x * (c_a(alpha_n) + c_b(alpha_p, beta) * sigmoid(beta * x)) + eps * (x > 0)
     // where c_a = softplus and c_b(a, b) = softplus(a) + b are constraining functions
@@ -2063,6 +2108,7 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+        GGML_SCALE_MODE_BICUBIC  = 2,
 
         GGML_SCALE_MODE_COUNT
     };

ggml/src/CMakeLists.txt

@@ -304,6 +304,14 @@ function(ggml_add_cpu_backend_variant tag_name)
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        foreach (feat VXE2 NNPA)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
@@ -371,6 +379,13 @@ if (GGML_CPU_ALL_VARIANTS)
         else()
             message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(z15    Z15 VXE2)
+            ggml_add_cpu_backend_variant(z16    Z16 VXE2 NNPA)
+        else()
+            message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()
@@ -390,6 +405,7 @@ ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
 ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
+ggml_add_backend(Hexagon)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-alloc.c

@@ -226,16 +226,23 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
     }
 
     if (best_fit_block == -1) {
-        // no suitable block found, try the last block (this will grow a chunks size)
+        // no suitable block found, try the last block (this may grow a chunks size)
+        int64_t best_reuse = INT64_MIN;
         for (int c = 0; c < alloc->n_chunks; ++c) {
             struct tallocr_chunk * chunk = alloc->chunks[c];
             if (chunk->n_free_blocks > 0) {
                 struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1];
                 max_avail = MAX(max_avail, block->size);
-                if (block->size >= size) {
+                int64_t reuse_factor = chunk->max_size - block->offset - size;
+                // reuse_factor < 0 : amount of extra memory that needs to be allocated
+                // reuse_factor = 0 : allocated free space exactly matches tensor size
+                // reuse_factor > 0 : superfluous memory that will remain unused
+                bool better_reuse = best_reuse < 0 && reuse_factor > best_reuse;
+                bool better_fit = reuse_factor >= 0 && reuse_factor < best_reuse;
+                if (block->size >= size && (better_reuse || better_fit)) {
                     best_fit_chunk = c;
                     best_fit_block = chunk->n_free_blocks - 1;
-                    break;
+                    best_reuse = reuse_factor;
                 }
             }
         }
@@ -268,7 +275,7 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, addr, tensor);
     size_t cur_max = addr.offset + size;
-    if (cur_max > alloc->max_size[addr.chunk]) {
+    if (cur_max > chunk->max_size) {
         // sort allocated_tensors by chunk/offset
         for (int i = 0; i < 1024; i++) {
             for (int j = i + 1; j < 1024; j++) {
@@ -598,6 +605,26 @@ static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
+// free the extra space at the end if the new tensor is smaller
+static void ggml_gallocr_free_extra_space(ggml_gallocr_t galloc, struct ggml_tensor * node, struct ggml_tensor * parent) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+
+    size_t parent_size = ggml_backend_buft_get_alloc_size(galloc->bufts[p_hn->buffer_id], parent);
+    size_t node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+
+    GGML_ASSERT(parent_size >= node_size);
+
+    if (parent_size > node_size) {
+        struct ggml_dyn_tallocr * p_alloc = galloc->buf_tallocs[p_hn->buffer_id];
+        struct buffer_address p_addr = p_hn->addr;
+        p_addr.offset += node_size;
+        size_t extra_size = parent_size - node_size;
+        AT_PRINTF("freeing extra %zu bytes from parent %s for %s\n", extra_size, parent->name, node->name);
+        ggml_dyn_tallocr_free_tensor(p_alloc, p_addr, extra_size, parent);
+    }
+}
+
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
     GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
@@ -643,6 +670,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                             hn->addr = p_hn->addr;
                             p_hn->allocated = false; // avoid freeing the parent
                             view_src_hn->allocated = false;
+                            ggml_gallocr_free_extra_space(galloc, node, view_src);
                             return;
                         }
                     } else {
@@ -650,6 +678,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                         hn->buffer_id = p_hn->buffer_id;
                         hn->addr = p_hn->addr;
                         p_hn->allocated = false; // avoid freeing the parent
+                        ggml_gallocr_free_extra_space(galloc, node, parent);
                         return;
                     }
                 }

ggml/src/ggml-backend-reg.cpp

@@ -57,6 +57,10 @@
 #include "ggml-opencl.h"
 #endif
 
+#ifdef GGML_USE_HEXAGON
+#include "ggml-hexagon.h"
+#endif
+
 #ifdef GGML_USE_BLAS
 #include "ggml-blas.h"
 #endif
@@ -199,6 +203,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
+#ifdef GGML_USE_HEXAGON
+        register_backend(ggml_backend_hexagon_reg());
+#endif
 #ifdef GGML_USE_CANN
         register_backend(ggml_backend_cann_reg());
 #endif
@@ -598,6 +605,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("sycl", silent, dir_path);
     ggml_backend_load_best("vulkan", silent, dir_path);
     ggml_backend_load_best("opencl", silent, dir_path);
+    ggml_backend_load_best("hexagon", silent, dir_path);
     ggml_backend_load_best("musa", silent, dir_path);
     ggml_backend_load_best("cpu", silent, dir_path);
     // check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend

ggml/src/ggml-cann/acl_tensor.cpp

@@ -51,28 +51,31 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
     return ACL_DT_UNDEFINED;
 }
 
-aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
-                                   size_t* nb, int64_t dims, aclFormat format,
-                                   size_t offset) {
+aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
+                                    int64_t *           ne,
+                                    size_t *            nb,
+                                    int64_t             dims,
+                                    aclFormat           format,
+                                    size_t              offset) {
     // If tensor is bcasted, Up to GGML_MAX_DIMS additional dimensions will be
     // added.
     int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
 
     if (ne == nullptr) {
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
-            acl_ne[i] = tensor->ne[i];
+            acl_ne[i]     = tensor->ne[i];
             // The step size of acl is in elements.
             acl_stride[i] = tensor->nb[i] / ggml_element_size(tensor);
         }
     } else {
         // With bcast
         for (int i = 0; i < dims; i++) {
-            acl_ne[i] = ne[i];
+            acl_ne[i]     = ne[i];
             acl_stride[i] = nb[i] / ggml_element_size(tensor);
         }
     }
 
-    int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
+    int64_t final_dims      = (dims == 0 ? GGML_MAX_DIMS : dims);
     int64_t acl_storage_len = 1;
     for (int i = 0; i < final_dims; i++) {
         acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
@@ -84,15 +87,13 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     std::reverse(acl_ne, acl_ne + final_dims);
     std::reverse(acl_stride, acl_stride + final_dims);
 
-    aclTensor* acl_tensor = aclCreateTensor(
-        acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
-        elem_offset, format, &acl_storage_len, 1,
-        tensor->data);
+    aclTensor * acl_tensor = aclCreateTensor(acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
+                                             elem_offset, format, &acl_storage_len, 1, tensor->data);
 
     return acl_tensor;
 }
 
-bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
+bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1) {
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (t1->ne[i] != t0->ne[i] && t1->ne[i] != 1) {
             return true;
@@ -101,15 +102,16 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     return false;
 }
 
-int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
-                                  const ggml_tensor* src1,
-                                  int64_t* bcast_src0_ne,
-                                  int64_t* bcast_src1_ne, size_t* bcast_src0_nb,
-                                  size_t* bcast_src1_nb) {
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
+                                  const ggml_tensor * src1,
+                                  int64_t *           bcast_src0_ne,
+                                  int64_t *           bcast_src1_ne,
+                                  size_t *            bcast_src0_nb,
+                                  size_t *            bcast_src1_nb) {
     GGML_ASSERT(ggml_can_repeat(src1, src0));
     int bcast_dim_cnt = 0;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        int64_t nr = src0->ne[i] / src1->ne[i];
+        int64_t nr                   = src0->ne[i] / src1->ne[i];
         bcast_src0_ne[bcast_dim_cnt] = src0->ne[i] / nr;
         bcast_src1_ne[bcast_dim_cnt] = src1->ne[i];
         bcast_src0_nb[bcast_dim_cnt] = src0->nb[i];
@@ -119,21 +121,26 @@ int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
             // Need to add an extra dim.
             bcast_src0_ne[bcast_dim_cnt] = nr;
             bcast_src1_ne[bcast_dim_cnt] = 1;
-            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] *
-                                           bcast_src0_ne[bcast_dim_cnt - 1];
-            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] *
-                                           bcast_src1_ne[bcast_dim_cnt - 1];
+            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] * bcast_src0_ne[bcast_dim_cnt - 1];
+            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] * bcast_src1_ne[bcast_dim_cnt - 1];
             bcast_dim_cnt++;
         }
     }
     return bcast_dim_cnt;
 }
 
-int64_t ggml_cann_get_mulmat_bcast_shape(
-    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
-    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
-    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
-    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb) {
+int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
+                                         const int64_t * weight_ne,
+                                         const int64_t * dst_ne,
+                                         const size_t *  input_nb,
+                                         const size_t *  weight_nb,
+                                         const size_t *  dst_nb,
+                                         int64_t *       bcast_input_ne,
+                                         int64_t *       bcast_weight_ne,
+                                         int64_t *       bcast_dst_ne,
+                                         size_t *        bcast_input_nb,
+                                         size_t *        bcast_weight_nb,
+                                         size_t *        bcast_dst_nb) {
     // input and dst shoule in same shape, except first two dims.
     GGML_ASSERT(input_ne[2] == dst_ne[2]);
     GGML_ASSERT(input_ne[3] == dst_ne[3]);
@@ -148,34 +155,30 @@ int64_t ggml_cann_get_mulmat_bcast_shape(
         // Do not use bcast in the first two dimensions because we only support
         // the bcast batch dimension. Just copy them.
         if (i < 2 || nr == 1) {
-            bcast_input_ne[bcast_dim_cnt] = input_ne[i];
+            bcast_input_ne[bcast_dim_cnt]  = input_ne[i];
             bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
-            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i];
+            bcast_dst_ne[bcast_dim_cnt]    = dst_ne[i];
 
-            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
+            bcast_input_nb[bcast_dim_cnt]  = input_nb[i];
             bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
-            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
+            bcast_dst_nb[bcast_dim_cnt]    = dst_nb[i];
             bcast_dim_cnt++;
         } else {
             // Need to add an extra dim.
-            bcast_input_ne[bcast_dim_cnt] = nr;
-            bcast_dst_ne[bcast_dim_cnt] = nr;
+            bcast_input_ne[bcast_dim_cnt]  = nr;
+            bcast_dst_ne[bcast_dim_cnt]    = nr;
             bcast_weight_ne[bcast_dim_cnt] = 1;
-            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
-            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
+            bcast_input_nb[bcast_dim_cnt]  = input_nb[i];
+            bcast_dst_nb[bcast_dim_cnt]    = dst_nb[i];
             bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
             bcast_dim_cnt++;
 
-            bcast_input_ne[bcast_dim_cnt] = input_ne[i] / nr;
-            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i] / nr;
+            bcast_input_ne[bcast_dim_cnt]  = input_ne[i] / nr;
+            bcast_dst_ne[bcast_dim_cnt]    = dst_ne[i] / nr;
             bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
-            bcast_input_nb[bcast_dim_cnt] = bcast_input_nb[bcast_dim_cnt - 1] *
-                                            bcast_input_ne[bcast_dim_cnt - 1];
-            bcast_dst_nb[bcast_dim_cnt] = bcast_dst_nb[bcast_dim_cnt - 1] *
-                                          bcast_dst_ne[bcast_dim_cnt - 1];
-            bcast_weight_nb[bcast_dim_cnt] =
-                bcast_weight_nb[bcast_dim_cnt - 1] *
-                bcast_weight_ne[bcast_dim_cnt - 1];
+            bcast_input_nb[bcast_dim_cnt]  = bcast_input_nb[bcast_dim_cnt - 1] * bcast_input_ne[bcast_dim_cnt - 1];
+            bcast_dst_nb[bcast_dim_cnt]    = bcast_dst_nb[bcast_dim_cnt - 1] * bcast_dst_ne[bcast_dim_cnt - 1];
+            bcast_weight_nb[bcast_dim_cnt] = bcast_weight_nb[bcast_dim_cnt - 1] * bcast_weight_ne[bcast_dim_cnt - 1];
             bcast_dim_cnt++;
         }
     }

ggml/src/ggml-cann/acl_tensor.h

@@ -62,10 +62,12 @@ aclDataType ggml_cann_type_mapping(ggml_type type);
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
-aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
-                             size_t* nb = nullptr, int64_t dims = 0,
-                             aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
+                                    int64_t *           ne     = nullptr,
+                                    size_t *            nb     = nullptr,
+                                    int64_t             dims   = 0,
+                                    aclFormat           format = ACL_FORMAT_ND,
+                                    size_t              offset = 0);
 
 /**
  * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
@@ -87,12 +89,15 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
-template<typename TYPE>
-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   TYPE type_size, int64_t* ne, TYPE* nb,
-                                   int64_t dims,
-                                   aclFormat format = ACL_FORMAT_ND,
-                                   size_t offset = 0) {
+template <typename TYPE>
+aclTensor * ggml_cann_create_tensor(void *      data_ptr,
+                                    aclDataType dtype,
+                                    TYPE        type_size,
+                                    int64_t *   ne,
+                                    TYPE *      nb,
+                                    int64_t     dims,
+                                    aclFormat   format = ACL_FORMAT_ND,
+                                    size_t      offset = 0) {
     int64_t tmp_ne[GGML_MAX_DIMS * 2];
     int64_t tmp_stride[GGML_MAX_DIMS * 2];
 
@@ -109,9 +114,8 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
     std::reverse(tmp_ne, tmp_ne + dims);
     std::reverse(tmp_stride, tmp_stride + dims);
 
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
+    aclTensor * acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size, format, &acl_storage_len, 1, data_ptr);
 
     return acl_tensor;
 }
@@ -132,7 +136,7 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
  *          to 1. If such a dimension is found, broadcasting is required to align t1
  *          with t0 for element-wise operations.
  */
-bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
+bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1);
 
 /**
  * @brief   Computes broadcast shapes and strides for two ggml_tensors.
@@ -187,19 +191,21 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
  *  dim1 in a inserted dim, should add nb for dim1,
  *  and all other nb moves to next in order.
  */
-int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
-                        int64_t* bcast_ne_src0, int64_t* bcast_ne_src1,
-                        size_t* bcast_nb_src0, size_t* bcast_nb_src1);
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
+                                  const ggml_tensor * src1,
+                                  int64_t *           bcast_ne_src0,
+                                  int64_t *           bcast_ne_src1,
+                                  size_t *            bcast_nb_src0,
+                                  size_t *            bcast_nb_src1);
 
 // Bcast macro to avoid duplicate code.
-#define BCAST_SHAPE(src0, src1)                                              \
-    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                            \
-    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                            \
-    size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2];                             \
-    size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2];                             \
-    int64_t bcast_dims = ggml_cann_get_bcast_shape(                          \
-        src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, bcast_##src0##_nb, \
-        bcast_##src1##_nb);
+#define BCAST_SHAPE(src0, src1)                                                                      \
+    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                                                    \
+    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                                                    \
+    size_t  bcast_##src0##_nb[GGML_MAX_DIMS * 2];                                                    \
+    size_t  bcast_##src1##_nb[GGML_MAX_DIMS * 2];                                                    \
+    int64_t bcast_dims = ggml_cann_get_bcast_shape(src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, \
+                                                   bcast_##src0##_nb, bcast_##src1##_nb);
 
 #define BCAST_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
 
@@ -233,26 +239,31 @@ int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* sr
  *       before cast dim.
  * @sa ggml_cann_get_bcast_shape
  */
-int64_t ggml_cann_get_mulmat_bcast_shape(
-    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
-    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
-    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
-    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb);
+int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
+                                         const int64_t * weight_ne,
+                                         const int64_t * dst_ne,
+                                         const size_t *  input_nb,
+                                         const size_t *  weight_nb,
+                                         const size_t *  dst_nb,
+                                         int64_t *       bcast_input_ne,
+                                         int64_t *       bcast_weight_ne,
+                                         int64_t *       bcast_dst_ne,
+                                         size_t *        bcast_input_nb,
+                                         size_t *        bcast_weight_nb,
+                                         size_t *        bcast_dst_nb);
 
 // Bcast macro to avoid duplicate code.
-#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                         \
-    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                      \
-    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                     \
-    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                        \
-    size_t bcast_##input##_nb[GGML_MAX_DIMS * 2];                       \
-    size_t bcast_##weight##_nb[GGML_MAX_DIMS * 2];                      \
-    size_t bcast_##dst##_nb[GGML_MAX_DIMS * 2];                         \
-    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(              \
-        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, \
-        bcast_##input##_ne, bcast_##weight##_ne, bcast_##dst##_ne,      \
-        bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
+#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                                                                  \
+    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                                                               \
+    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                                                              \
+    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                                                                 \
+    size_t  bcast_##input##_nb[GGML_MAX_DIMS * 2];                                                               \
+    size_t  bcast_##weight##_nb[GGML_MAX_DIMS * 2];                                                              \
+    size_t  bcast_##dst##_nb[GGML_MAX_DIMS * 2];                                                                 \
+    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(                                                       \
+        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, bcast_##input##_ne, bcast_##weight##_ne, \
+        bcast_##dst##_ne, bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
 
-#define BCAST_MUL_MAT_PARAM(tensor) \
-    bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
+#define BCAST_MUL_MAT_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
 
 #endif  // CANN_ACL_TENSOR_H

ggml/src/ggml-cann/aclnn_ops.h

@@ -62,7 +62,7 @@
  * @param   dst The ggml tensor representing the destination, which op is
  *              GGML_OP_REPEAT and specifies the desired dimensions.
  */
-void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
@@ -82,7 +82,7 @@ void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result of the Leaky ReLU
  *            activation is stored, which op is `GGML_OP_LEAKY_RELU`
  */
-void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_leaky_relu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief    Concatenates multiple tensors along a specified dimension using the
@@ -97,7 +97,7 @@ void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @attention tensorList length should be 2 and the dimension using for concat
  *            default to 1.
  */
-void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_concat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Generates a sequence of evenly spaced values within a specified
@@ -113,7 +113,7 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *            `start`, 'stop' and 'step' are in dst->op_params and dst->op is
  *            `GGML_OP_ARANGE`.
  */
-void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_arange(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies a clamp operation to the elements of a ggml tensor using the
@@ -131,7 +131,7 @@ void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the clamped values will be stored.
  *            dst->op is `GGML_OP_CLAMP`, `min` and `max` value is in dst->params.
  */
-void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_clamp(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Scales the elements of a ggml tensor by a constant factor using the
@@ -148,7 +148,7 @@ void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the scaled values will be stored.
  *            dst->op is `GGML_OP_SCALE` and `scale` value is in dst->params.
  */
-void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_scale(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Sorts the elements of a ggml tensor and returns the indices that
@@ -163,7 +163,7 @@ void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the sorted indices will be stored.
  *            dst->op is `GGML_OP_ARGSORT`.
  */
-void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_argsort(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the Layer Normalization for a ggml tensor using the CANN
@@ -185,7 +185,7 @@ void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the normalized values will be stored.
  * @attention `Var` defaults to dst->ne[0].
  */
-void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief  Computes the Group Normalization for a ggml tensor using the CANN
@@ -209,7 +209,7 @@ void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @attention eps defaults to 1e-6f.
  */
-void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the accumulation of tensors using the CANN backend.
@@ -228,7 +228,7 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the accumulated values will be stored.
  *            `inplace` is in dst->params, and dst->op is `GGML_OP_ACC`.
  */
-void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_acc(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the sum of elements along the last dimension of a ggml tensor
@@ -244,7 +244,7 @@ void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @attention `reduce_dims` defaults to 3, which means the last dimension.
  */
-void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the sum of elements in a ggml tensor.
@@ -258,7 +258,7 @@ void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  */
 
-void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
@@ -274,8 +274,7 @@ void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the upsampled values will be stored.
  *            dst->op is `GGML_OP_UPSCALE`.
  */
-void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
-                                  ggml_tensor* dst);
+void ggml_cann_upsample_nearest2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Pads a ggml tensor to match the dimensions of the destination tensor
@@ -290,7 +289,7 @@ void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
  * @param dst The destination tensor, which specifies the target dimensions for
  *            padding. dst->op is `GGML_OP_PAD`.
  */
-void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_pad(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Executes a 2D pooling operation on a ggml tensor using the CANN
@@ -307,7 +306,7 @@ void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor on which the pooling operation is to be
  *            performed. dst->op is `GGML_OP_POOL_2D`.
  */
-void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_pool2d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Duplicates a ggml tensor using the CANN backend.
@@ -326,7 +325,7 @@ void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *            different shape and dst is no-contiguous.
  * @note:     This func need to simplify.
  */
-void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_dup(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the Root Mean Square (RMS) normalization of a ggml tensor
@@ -348,7 +347,7 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the normalized values will be stored.
  *            dst->op is `GGML_OP_RMS_NORM`.
  */
-void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_rms_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies a diagonal mask to the tensor with a specified value.
@@ -363,7 +362,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *            `GGML_OP_DIAG_MASK`
  * @param value The value to use for masking.
  */
-void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, float value);
+void ggml_cann_diag_mask(ggml_backend_cann_context & ctx, ggml_tensor * dst, float value);
 
 /**
  * @brief   Performs an image-to-column transformation on the input tensor.
@@ -378,7 +377,7 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, float
  * @param dst The destination tensor that stores the result of the operation.
  *            dst->op is `GGML_OP_IM2COL`.
  */
-void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_im2col(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes time step embeddings using sine and cosine functions.
@@ -392,10 +391,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result of the embedding operation
  *            will be stored. dst->op is `GGML_OP_TIMESTEP_EMBEDDING`.
  */
-void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_timestep_embedding(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 // @see ggml_cann_dup.
-void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the softmax activation with optional masking.
@@ -417,7 +416,7 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored. dst->op is
  *            `GGML_OP_SOFTMAX`.
  */
-void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Extracts specific rows from a tensor based on indices.
@@ -429,7 +428,7 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the extracted rows will be stored.
  */
-void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Writes specific rows into a tensor at positions specified by indices.
@@ -441,7 +440,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the specified rows will be updated.
  */
-void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_set_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Executes matrix multiplication for the given tensor.
@@ -454,7 +453,7 @@ void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor for storing the result of the matrix
  *            multiplication. dst->op is `GGML_OP_MUL_MAT`.
  */
-void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_mul_mat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief Applies Rotary Positional Embedding (RoPE) to the input tensor.
@@ -477,7 +476,7 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @note The function currently does not support cases where the freq_scale is
  *       not equal 1.
  */
-void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the index of the maximum value along the specified dimension
@@ -492,7 +491,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the indices of the maximum values will
  *            be stored. dst->op is `GGML_OP_ARGMAX`.
  */
-void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief Adds two tensors element-wise and stores the result in a destination
@@ -509,8 +508,10 @@ void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param acl_src1 The second source tensor.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
-    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
+void aclnn_add(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src0,
+               aclTensor *                 acl_src1,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Sub two tensors element-wise and stores the result in a destination
@@ -527,8 +528,10 @@ void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
  * @param acl_src1 The second source tensor.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
-    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
+void aclnn_sub(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src0,
+               aclTensor *                 acl_src1,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Performs element-wise multiplication of two tensors and stores the
@@ -546,8 +549,10 @@ void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
  * @param acl_other The second tensor for element-wise multiplication.
  * @param acl_dst The destination tensor where the result will be stored.
  */
-void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
+void aclnn_mul(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src,
+               aclTensor *                 acl_other,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Matrix division, optionally in-place.
@@ -567,8 +572,10 @@ void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param inplace Flag indicating whether to perform the operation in-place on
  * `acl_src`.
  */
-void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
+void aclnn_div(ggml_backend_cann_context & ctx,
+               aclTensor *                 acl_src,
+               aclTensor *                 acl_other,
+               aclTensor *                 acl_dst = nullptr);
 
 /**
  * @brief Applies element-wise cosine function to the elements of a tensor.
@@ -584,8 +591,7 @@ void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param acl_dst The destination tensor where the cosine results will be
  * stored.
  */
-void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_dst);
+void aclnn_cos(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst);
 
 /**
  * @brief Applies element-wise sine function to the elements of a tensor.
@@ -602,8 +608,7 @@ void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param acl_src The source tensor on which the sine function will be applied.
  * @param acl_dst The destination tensor where the sine results will be stored.
  */
-void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-    aclTensor* acl_dst);
+void aclnn_sin(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst);
 
 /**
  * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one
@@ -621,8 +626,12 @@ void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
  * @param acl_dst  Output pointer to the created ACL tensor corresponding to dst.
  */
-void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
-    aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst);
+void bcast_shape(ggml_tensor * src0,
+                 ggml_tensor * src1,
+                 ggml_tensor * dst,
+                 aclTensor **  acl_src0,
+                 aclTensor **  acl_src1,
+                 aclTensor **  acl_dst);
 
 /**
  * @brief   Computes the 1D transposed convolution (deconvolution) of a ggml
@@ -637,7 +646,7 @@ void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
  * @param dst The destination tensor where the transposed convolution result
  * will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`.
  */
-void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_conv_transpose_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies the ELU (Exponential Linear Unit) activation to a ggml tensor
@@ -662,7 +671,7 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
  * @param dst The destination tensor where the ELU-activated result will be stored.
  *            dst->op is expected to be `GGML_OP_ELU`.
  */
-void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_elu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Computes the mean of a ggml tensor element-wise using the CANN backend.
@@ -677,7 +686,7 @@ void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the mean result will be stored.
  *            dst->op is expected to be `GGML_OP_MEAN`.
  */
-void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_mean(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies 1D reflect padding to a ggml tensor using the CANN backend.
@@ -692,7 +701,7 @@ void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the padded result will be stored.
  *            dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
  */
-void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_pad_reflect_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Counts the number of equal elements in two ggml tensors using the CANN backend.
@@ -708,7 +717,7 @@ void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
  */
-void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Applies the Step activation function to a ggml tensor using the CANN backend.
@@ -723,7 +732,7 @@ void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_STEP`.
  */
-void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Performs the Flash Attention extended operator using the CANN backend.
@@ -738,59 +747,46 @@ void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  * @param dst The destination tensor where the result will be stored.
  *            dst->op is expected to be `GGML_OP_FLASH_ATTN_EXT`.
  */
-void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /*
  * @brief A generic wrapper for ACL resources with custom deleter support.
  */
-using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+using any_acl_resource = std::unique_ptr<void, std::function<void(void *)>>;
 
 /**
  * @brief Trait structure used to define how to destroy a given ACL resource type.
  *
  * @tparam T ACL resource type.
  */
-template<typename T>
-struct acl_resource_traits;
+template <typename T> struct acl_resource_traits;
 
 /**
  * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
  */
-template<>
-struct acl_resource_traits<aclTensor> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
-    }
+template <> struct acl_resource_traits<aclTensor> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyTensor(static_cast<aclTensor *>(p))); }
 };
 
 /**
  * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
  */
-template<>
-struct acl_resource_traits<aclIntArray> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
-    }
+template <> struct acl_resource_traits<aclIntArray> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray *>(p))); }
 };
 
 /**
  * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
  */
-template<>
-struct acl_resource_traits<aclScalar> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
-    }
+template <> struct acl_resource_traits<aclScalar> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyScalar(static_cast<aclScalar *>(p))); }
 };
 
 /**
  * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
  */
-template<>
-struct acl_resource_traits<aclTensorList> {
-    static void destroy(void* p) {
-        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
-    }
+template <> struct acl_resource_traits<aclTensorList> {
+    static void destroy(void * p) { ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList *>(p))); }
 };
 
 /**
@@ -800,14 +796,8 @@ struct acl_resource_traits<aclTensorList> {
  * @param ptr Raw pointer to ACL resource.
  * @return any_acl_resource Smart pointer that handles destruction.
  */
-template<typename T>
-any_acl_resource make_acl_resource(T* ptr) {
-    return any_acl_resource(
-        static_cast<void*>(ptr),
-        [](void* p) {
-            acl_resource_traits<T>::destroy(p);
-        }
-    );
+template <typename T> any_acl_resource make_acl_resource(T * ptr) {
+    return any_acl_resource(static_cast<void *>(ptr), [](void * p) { acl_resource_traits<T>::destroy(p); });
 }
 
 /**
@@ -817,8 +807,7 @@ any_acl_resource make_acl_resource(T* ptr) {
  * @param vec Target vector to hold ACL resources.
  * @param args Raw pointers to ACL resources.
  */
-template<typename... Args>
-void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+template <typename... Args> void register_acl_resources(std::vector<any_acl_resource> & vec, Args *... args) {
     (vec.emplace_back(make_acl_resource(args)), ...);
 }
 
@@ -826,39 +815,36 @@ void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
  * @brief Task class that wraps the execution of an aclnn function call.
  */
 class aclnn_task : public cann_task {
-    public:
-        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
-                   uint64_t workspace_size, aclOpExecutor * executor,
-                   aclrtStream stream) :
-            aclnn_func_(aclnn_func),
-            workspace_addr_(workspace_addr),
-            workspace_size_(workspace_size),
-            executor_(executor),
-            stream_(stream) {}
-        virtual void run_task() override {
-            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
-        }
-    private:
-        aclnn_func_t aclnn_func_;
-        void *          workspace_addr_;
-        uint64_t        workspace_size_;
-        aclOpExecutor * executor_;
-        aclrtStream     stream_;
+  public:
+    aclnn_task(aclnn_func_t    aclnn_func,
+               void *          workspace_addr,
+               uint64_t        workspace_size,
+               aclOpExecutor * executor,
+               aclrtStream     stream) :
+        aclnn_func_(aclnn_func),
+        workspace_addr_(workspace_addr),
+        workspace_size_(workspace_size),
+        executor_(executor),
+        stream_(stream) {}
+
+    virtual void run_task() override { ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_)); }
+  private:
+    aclnn_func_t    aclnn_func_;
+    void *          workspace_addr_;
+    uint64_t        workspace_size_;
+    aclOpExecutor * executor_;
+    aclrtStream     stream_;
 };
 
 /**
  * @brief Task class that releases ACL resources after usage.
  */
 class release_resource_task : public cann_task {
-public:
-    release_resource_task(std::vector<any_acl_resource>&& resources){
-        resource_ = std::move(resources);
-    }
+  public:
+    release_resource_task(std::vector<any_acl_resource> && resources) { resource_ = std::move(resources); }
 
-    virtual void run_task() override {
-        resource_.clear();
-    }
-private:
+    virtual void run_task() override { resource_.clear(); }
+  private:
     std::vector<any_acl_resource> resource_;
 };
 
@@ -866,38 +852,40 @@ private:
  * @brief Task class for performing asynchronous memory copy operations.
  */
 class async_memcpy_task : public cann_task {
-public:
-    async_memcpy_task(void* dst, const void* src, size_t size,
-                      aclrtMemcpyKind kind, aclrtStream stream)
-        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+  public:
+    async_memcpy_task(void * dst, const void * src, size_t size, aclrtMemcpyKind kind, aclrtStream stream) :
+        dst_(dst),
+        src_(src),
+        size_(size),
+        kind_(kind),
+        stream_(stream) {}
 
-    virtual void run_task() override {
-        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
-    }
-private:
-    void* dst_;
-    const void* src_;
-    size_t size_;
+    virtual void run_task() override { ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_)); }
+  private:
+    void *          dst_;
+    const void *    src_;
+    size_t          size_;
     aclrtMemcpyKind kind_;
-    aclrtStream stream_;
+    aclrtStream     stream_;
 };
 
 /**
  * @brief Task class for performing asynchronous memory set operations.
  */
 class async_memset_task : public cann_task {
-    public:
-    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
-            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+  public:
+    async_memset_task(void * buffer, size_t size, int32_t value, aclrtStream stream) :
+        buffer_(buffer),
+        size_(size),
+        value_(value),
+        stream_(stream) {}
 
-        virtual void run_task() override {
-            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
-        }
-    private:
-        void* buffer_;
-        size_t size_;
-        int32_t value_;
-        aclrtStream stream_;
+    virtual void run_task() override { ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_)); }
+  private:
+    void *      buffer_;
+    size_t      size_;
+    int32_t     value_;
+    aclrtStream stream_;
 };
 
 /**
@@ -918,25 +906,24 @@ class async_memset_task : public cann_task {
  * same stream are executed in queue order.
  */
 
-#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
-    do {                                                                                    \
-        uint64_t        workspaceSize = 0;                                                  \
-        aclOpExecutor * executor;                                                           \
-        void *          workspaceAddr = nullptr;                                            \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
-        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
-        if (workspaceSize > 0) {                                                            \
-            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
-            workspaceAddr = workspace_allocator.get();                                      \
-        }                                                                                   \
-        if (CTX.async_mode) {                                                               \
-            auto task =                                                                     \
-                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
-                    executor, CTX.stream()); \
-            CTX.task_queue.submit_task(std::move(task));                                    \
-        } else {                                                                            \
-            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
-        }                                                                                   \
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                                                  \
+    do {                                                                                                            \
+        uint64_t        workspaceSize = 0;                                                                          \
+        aclOpExecutor * executor;                                                                                   \
+        void *          workspaceAddr = nullptr;                                                                    \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));                        \
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */                          \
+        if (workspaceSize > 0) {                                                                                    \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);                                    \
+            workspaceAddr = workspace_allocator.get();                                                              \
+        }                                                                                                           \
+        if (CTX.async_mode) {                                                                                       \
+            auto task =                                                                                             \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize, executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                                            \
+        } else {                                                                                                    \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));                        \
+        }                                                                                                           \
     } while (0)
 
 /**
@@ -947,11 +934,10 @@ class async_memset_task : public cann_task {
  * @param ctx Backend context which manages task submission and async mode.
  * @param args Pointers to ACL resources to be released.
  */
-template <typename... Args>
-void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+template <typename... Args> void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
     std::vector<any_acl_resource> resources;
     register_acl_resources(resources, std::forward<Args>(args)...);
-    if(ctx.async_mode) {
+    if (ctx.async_mode) {
         auto task = std::make_unique<release_resource_task>(std::move(resources));
         ctx.task_queue.submit_task(std::move(task));
     }
@@ -966,8 +952,11 @@ void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... arg
  * @param len Size of memory to copy (in bytes).
  * @param kind Type of memory copy (host-to-device, device-to-host, etc).
  */
-inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
-                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx,
+                                   void *                      dst,
+                                   const void *                src,
+                                   size_t                      len,
+                                   aclrtMemcpyKind             kind) {
     if (ctx.async_mode) {
         auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
         ctx.task_queue.submit_task(std::move(task));
@@ -976,8 +965,11 @@ inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
     }
 }
 
-inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
-                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx,
+                                   void *                      dst,
+                                   const void *                src,
+                                   size_t                      len,
+                                   aclrtMemcpyKind             kind) {
     if (ctx->async_mode) {
         auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
         ctx->task_queue.submit_task(std::move(task));
@@ -994,8 +986,7 @@ inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
  * @param size Size of the memory buffer (in bytes).
  * @param value Value to set in the buffer.
  */
-inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
-                                   size_t size, int value) {
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer, size_t size, int value) {
     if (ctx.async_mode) {
         auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
         ctx.task_queue.submit_task(std::move(task));
@@ -1029,7 +1020,7 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  * @param dst The destination tensor where the expert-weighted token outputs are stored.
  *            Expected to be of shape [M, K, N, 1].
  */
-void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
@@ -1041,20 +1032,14 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @param tensor Pointer to the target ggml_tensor object (const-qualified).
  */
-static bool is_matmul_weight(const ggml_tensor* tensor) {
-    std::string name = ggml_get_name(tensor);
-    static const std::unordered_set<std::string> weight_suffixes{
-        "output.weight",
-        "attn_q.weight",
-        "attn_k.weight",
-        "attn_v.weight",
-        "attn_output.weight",
-        "ffn_gate.weight",
-        "ffn_up.weight",
-        "ffn_down.weight"
-    };
+static bool is_matmul_weight(const ggml_tensor * tensor) {
+    std::string                                  name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{ "output.weight",      "attn_q.weight",
+                                                                  "attn_k.weight",      "attn_v.weight",
+                                                                  "attn_output.weight", "ffn_gate.weight",
+                                                                  "ffn_up.weight",      "ffn_down.weight" };
 
-    for (const auto& suffix : weight_suffixes) {
+    for (const auto & suffix : weight_suffixes) {
         if (name.find(suffix) != std::string::npos) {
             return true;
         }
@@ -1078,14 +1063,13 @@ static bool is_matmul_weight(const ggml_tensor* tensor) {
  * @param ctx The CANN backend context used to manage execution and resources.
  * @param dst The destination tensor.
  */
-template <auto binary_op>
-void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src0 = dst->src[0];
-    ggml_tensor* src1 = dst->src[1];
+template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
 
-    aclTensor* acl_src0;
-    aclTensor* acl_src1;
-    aclTensor* acl_dst;
+    aclTensor * acl_src0;
+    aclTensor * acl_src1;
+    aclTensor * acl_dst;
 
     // Need bcast
     bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
@@ -1094,7 +1078,6 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
-
 /**
  * @brief Applies a unary operation to an input tensor using the CANN backend.
  *
@@ -1107,12 +1090,12 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @param ctx The CANN backend context for managing resources and execution.
  * @param dst The destination tensor. Its src[0] is treated as the input tensor.
  */
-template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
-    void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src = dst->src[0];
+template <void unary_op(ggml_backend_cann_context &, aclTensor *, aclTensor *)>
+void ggml_cann_op_unary(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
 
-    aclTensor* acl_src = ggml_cann_create_tensor(src);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor * acl_src = ggml_cann_create_tensor(src);
+    aclTensor * acl_dst = ggml_cann_create_tensor(dst);
 
     unary_op(ctx, acl_src, acl_dst);
     ggml_cann_release_resources(ctx, acl_src, acl_dst);
@@ -1138,9 +1121,9 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
  *
  * @see GGML_CANN_CALL_OP_UNARY
  */
-void ggml_cann_op_unary(
-    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
-    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_op_unary(std::function<void(ggml_backend_cann_context &, aclTensor *, aclTensor *)> unary_op,
+                        ggml_backend_cann_context &                                                ctx,
+                        ggml_tensor *                                                              dst);
 
 /**
  * @brief Applies a gated (GLU-style) unary operation using the CANN backend.
@@ -1172,9 +1155,9 @@ void ggml_cann_op_unary(
  *
  * @see GGML_CANN_CALL_OP_UNARY_GATED
  */
-void ggml_cann_op_unary_gated(
-    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
-    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, aclTensor *, aclTensor *)> unary_op,
+                              ggml_backend_cann_context &                                                ctx,
+                              ggml_tensor *                                                              dst);
 
 /**
  * @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
@@ -1197,16 +1180,13 @@ void ggml_cann_op_unary_gated(
  * @see ggml_cann_op_unary
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                              \
-    do {                                                              \
-        auto lambda = [](ggml_backend_cann_context& ctx,              \
-            aclTensor* acl_src,                                       \
-            aclTensor* acl_dst) {                                     \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
-        };                                                            \
-        ggml_cann_op_unary(lambda, ctx, dst);                         \
-    }                                                                 \
-    while (0)
+#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                                                              \
+    do {                                                                                              \
+        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
+        };                                                                                            \
+        ggml_cann_op_unary(lambda, ctx, dst);                                                         \
+    } while (0)
 
 /**
  * @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
@@ -1229,15 +1209,12 @@ void ggml_cann_op_unary_gated(
  * @see ggml_cann_op_unary_gated
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                        \
-    do {                                                              \
-        auto lambda = [](ggml_backend_cann_context& ctx,              \
-            aclTensor* acl_src,                                       \
-            aclTensor* acl_dst) {                                     \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
-        };                                                            \
-        ggml_cann_op_unary_gated(lambda, ctx, dst);                   \
-    }                                                                 \
-    while (0)
+#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                                                        \
+    do {                                                                                              \
+        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
+        };                                                                                            \
+        ggml_cann_op_unary_gated(lambda, ctx, dst);                                                   \
+    } while (0)
 
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -44,7 +44,7 @@
 #include "../include/ggml.h"
 #include "../ggml-impl.h"
 
-#define MATRIX_ROW_PADDING 512
+#define MATRIX_ROW_PADDING    512
 #define GGML_CANN_MAX_STREAMS 8
 
 /**
@@ -56,8 +56,7 @@
  * @param line The line number at which the error occurred.
  * @param msg The error message.
  */
-[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
-                                  const char* file, int line, const char* msg);
+[[noreturn]] void ggml_cann_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
 
 /**
  * @brief Checks the result of a CANN function call and invokes the error
@@ -89,25 +88,24 @@ struct ggml_cann_device_info {
      * @brief Information about a single CANN device.
      */
     struct cann_device_info {
-        int cc;                 /**< Compute capability.                   */
+        int    cc;              /**< Compute capability.                   */
         size_t smpb;            /**< Maximum shared memory per block.      */
-        bool vmm;               /**< Virtual memory support.               */
+        bool   vmm;             /**< Virtual memory support.               */
         size_t vmm_granularity; /**< Granularity of virtual memory.        */
         size_t total_vram;      /**< Total video RAM available on the device. */
     };
 
-    cann_device_info devices[GGML_CANN_MAX_DEVICES] =
-        {}; /**< Array of CANN device information. */
+    cann_device_info devices[GGML_CANN_MAX_DEVICES] = {}; /**< Array of CANN device information. */
 };
 
-const ggml_cann_device_info& ggml_cann_info();
+const ggml_cann_device_info & ggml_cann_info();
 
-void ggml_cann_set_device(int32_t device);
+void    ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
-std::optional<std::string> get_env(const std::string& name);
-bool parse_bool(const std::string& value);
-int parse_integer(const std::string& value);
+std::optional<std::string> get_env(const std::string & name);
+bool                       parse_bool(const std::string & value);
+int                        parse_integer(const std::string & value);
 
 /**
  * @brief Abstract base class for memory pools used by CANN.
@@ -126,7 +124,7 @@ struct ggml_cann_pool {
      *                     will be stored.
      * @return             Pointer to the allocated memory block.
      */
-    virtual void* alloc(size_t size, size_t* actual_size) = 0;
+    virtual void * alloc(size_t size, size_t * actual_size) = 0;
 
     /**
      * @brief Frees a previously allocated memory block.
@@ -136,16 +134,16 @@ struct ggml_cann_pool {
      * @note Note that all CANN opertors are running async. Make sure memory is
      *       still avaiable before this operator finished.
      */
-    virtual void free(void* ptr, size_t size) = 0;
+    virtual void free(void * ptr, size_t size) = 0;
 };
 
 /**
  * @brief RAII wrapper for managing memory allocations from a CANN memory pool.
  */
 struct ggml_cann_pool_alloc {
-    ggml_cann_pool* pool = nullptr; /**< Pointer to the memory pool. */
-    void* ptr = nullptr;    /**< Pointer to the allocated memory block. */
-    size_t actual_size = 0; /**< Actual size of the allocated memory block. */
+    ggml_cann_pool * pool        = nullptr; /**< Pointer to the memory pool. */
+    void *           ptr         = nullptr; /**< Pointer to the allocated memory block. */
+    size_t           actual_size = 0;       /**< Actual size of the allocated memory block. */
 
     /**
      * @brief Default constructor.
@@ -156,16 +154,14 @@ struct ggml_cann_pool_alloc {
      * @brief Constructor that initializes the memory pool.
      * @param pool Reference to the memory pool.
      */
-    explicit ggml_cann_pool_alloc(ggml_cann_pool& pool) : pool(&pool) {}
+    explicit ggml_cann_pool_alloc(ggml_cann_pool & pool) : pool(&pool) {}
 
     /**
      * @brief Constructor that initializes the memory pool and allocates memory.
      * @param pool Reference to the memory pool.
      * @param size Size of the memory block to allocate.
      */
-    ggml_cann_pool_alloc(ggml_cann_pool& pool, size_t size) : pool(&pool) {
-        alloc(size);
-    }
+    ggml_cann_pool_alloc(ggml_cann_pool & pool, size_t size) : pool(&pool) { alloc(size); }
 
     /**
      * @brief Destructor that frees the allocated memory block.
@@ -181,7 +177,7 @@ struct ggml_cann_pool_alloc {
      * @param size Size of the memory block to allocate.
      * @return Pointer to the allocated memory block.
      */
-    void* alloc(size_t size) {
+    void * alloc(size_t size) {
         GGML_ASSERT(pool != nullptr);
         GGML_ASSERT(ptr == nullptr);
         ptr = pool->alloc(size, &this->actual_size);
@@ -194,7 +190,7 @@ struct ggml_cann_pool_alloc {
      * @param size Size of the memory block to allocate.
      * @return Pointer to the allocated memory block.
      */
-    void* alloc(ggml_cann_pool& pool, size_t size) {
+    void * alloc(ggml_cann_pool & pool, size_t size) {
         this->pool = &pool;
         return alloc(size);
     }
@@ -203,25 +199,25 @@ struct ggml_cann_pool_alloc {
      * @brief Gets the pointer to the allocated memory block.
      * @return Pointer to the allocated memory block.
      */
-    void* get() { return ptr; }
+    void * get() { return ptr; }
 
     // Deleted copy constructor
-    ggml_cann_pool_alloc(const ggml_cann_pool_alloc&) = delete;
+    ggml_cann_pool_alloc(const ggml_cann_pool_alloc &) = delete;
 
     // Deleted move constructor
-    ggml_cann_pool_alloc(ggml_cann_pool_alloc&&) = delete;
+    ggml_cann_pool_alloc(ggml_cann_pool_alloc &&) = delete;
 
     // Deleted copy assignment operator
-    ggml_cann_pool_alloc& operator=(const ggml_cann_pool_alloc&) = delete;
+    ggml_cann_pool_alloc & operator=(const ggml_cann_pool_alloc &) = delete;
 
     // Deleted move assignment operator
-    ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
+    ggml_cann_pool_alloc & operator=(ggml_cann_pool_alloc &&) = delete;
 };
 
 /**
  * @brief Function pointer type for ACLNN operator calls.
  */
-using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+using aclnn_func_t = aclnnStatus (*)(void *, uint64_t, aclOpExecutor *, aclrtStream);
 
 /**
  * @brief Base class for all CANN tasks to be submitted to the task queue.
@@ -229,7 +225,7 @@ using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStrea
  * Users should override the run_task() method with actual task logic.
  */
 class cann_task {
-public:
+  public:
     virtual void run_task() {}
 };
 
@@ -237,16 +233,20 @@ public:
  * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
  */
 class cann_task_queue {
-public:
+  public:
     /**
      * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
      *
      * @param capacity Queue capacity. Must be a power of 2.
      * @param device Target device ID (used for context setting).
      */
-    explicit cann_task_queue(size_t capacity, int32_t device)
-        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
-          running_(false), device_(device) {
+    explicit cann_task_queue(size_t capacity, int32_t device) :
+        buffer_(capacity),
+        capacity_(capacity),
+        head_(0),
+        tail_(0),
+        running_(false),
+        device_(device) {
         GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
         mask_ = capacity_ - 1;
     }
@@ -257,7 +257,7 @@ public:
      * @param item Unique pointer to the task.
      * @return true if the task was successfully enqueued, false if the queue was full.
      */
-    bool enqueue(std::unique_ptr<cann_task>&& item) {
+    bool enqueue(std::unique_ptr<cann_task> && item) {
         size_t next_tail = (tail_ + 1) & mask_;
 
         if (next_tail == head_) {
@@ -276,17 +276,16 @@ public:
      *
      * @param task Task to be submitted.
      */
-    void submit_task(std::unique_ptr<cann_task>&& task) {
-        while(!enqueue(std::move(task))) {
+    void submit_task(std::unique_ptr<cann_task> && task) {
+        while (!enqueue(std::move(task))) {
             std::this_thread::yield();
             continue;
         }
 
         if (!running_) {
             running_ = true;
-            thread_ = std::thread(&cann_task_queue::execute, this);
+            thread_  = std::thread(&cann_task_queue::execute, this);
         }
-
     }
 
     /**
@@ -309,7 +308,7 @@ public:
         }
     }
 
-private:
+  private:
     /**
      * @brief Worker thread function that continuously dequeues and executes tasks.
      */
@@ -317,7 +316,7 @@ private:
         ggml_cann_set_device(device_);
 
         while (running_) {
-            if(head_ == tail_) {
+            if (head_ == tail_) {
                 std::this_thread::yield();
                 continue;
             }
@@ -330,24 +329,24 @@ private:
     }
 
     std::vector<std::unique_ptr<cann_task>> buffer_;
-    const size_t capacity_;
-    size_t mask_;
-    size_t head_;
-    size_t tail_;
-    bool running_;
-    std::thread thread_;
-    int32_t device_;
+    const size_t                            capacity_;
+    size_t                                  mask_;
+    size_t                                  head_;
+    size_t                                  tail_;
+    bool                                    running_;
+    std::thread                             thread_;
+    int32_t                                 device_;
 };
 
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
     // dst tensor
-    void * node_address;
+    void *  node_address;
     int64_t ne[GGML_MAX_DIMS];
-    size_t nb[GGML_MAX_DIMS];
+    size_t  nb[GGML_MAX_DIMS];
 
     // src tensor
-    void * src_address[GGML_MAX_SRC];
+    void *  src_address[GGML_MAX_SRC];
     int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
     size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
 
@@ -376,13 +375,11 @@ struct ggml_cann_graph {
  * move existing graphs to the front (most recently used), and clear the cache.
  */
 struct ggml_cann_graph_lru_cache {
-    size_t capacity;  /**< Maximum number of graphs in the cache. */
+    size_t capacity;                         /**< Maximum number of graphs in the cache. */
 
-    std::list<ggml_cann_graph*> cache_list; /**< List storing cached graphs as raw pointers. */
+    std::list<ggml_cann_graph *> cache_list; /**< List storing cached graphs as raw pointers. */
 
-    ggml_cann_graph_lru_cache() {
-        capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12"));
-    }
+    ggml_cann_graph_lru_cache() { capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12")); }
 
     /**
      * @brief Push a new graph to the front of the cache.
@@ -390,11 +387,11 @@ struct ggml_cann_graph_lru_cache {
      * @param new_node Pointer to the new ggml_cann_graph to cache.
      *        Ownership is transferred to the cache (cache will delete it).
      */
-    void push(ggml_cann_graph* new_node) {
+    void push(ggml_cann_graph * new_node) {
         if (cache_list.size() >= capacity) {
-            ggml_cann_graph* old = cache_list.back();
+            ggml_cann_graph * old = cache_list.back();
             cache_list.pop_back();
-            delete old; // free the old graph
+            delete old;  // free the old graph
         }
         cache_list.push_front(new_node);
     }
@@ -403,7 +400,7 @@ struct ggml_cann_graph_lru_cache {
      * @brief Move an existing graph to the front of the cache.
      * @param node Pointer to the ggml_cann_graph to move.
      */
-    void move_to_front(ggml_cann_graph* node) {
+    void move_to_front(ggml_cann_graph * node) {
         cache_list.remove(node);
         cache_list.push_front(node);
     }
@@ -421,92 +418,89 @@ struct ggml_cann_graph_lru_cache {
     /**
      * @brief Destructor that clears the cache and frees all cached graphs.
      */
-    ~ggml_cann_graph_lru_cache() {
-        clear();
-    }
+    ~ggml_cann_graph_lru_cache() { clear(); }
 };
 #endif  // USE_ACL_GRAPH
 
 struct ggml_cann_rope_cache {
     ~ggml_cann_rope_cache() {
-        if(theta_scale_cache != nullptr) {
+        if (theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
-        if(sin_cache != nullptr) {
+        if (sin_cache != nullptr) {
             ACL_CHECK(aclrtFree(sin_cache));
         }
-        if(cos_cache != nullptr) {
+        if (cos_cache != nullptr) {
             ACL_CHECK(aclrtFree(cos_cache));
         }
     }
 
-    void* theta_scale_cache = nullptr;
+    void *  theta_scale_cache  = nullptr;
     int64_t theta_scale_length = 0;
     // sin/cos cache, used only to accelerate first layer on each device
-    void* sin_cache = nullptr;
-    void* cos_cache = nullptr;
-    int64_t position_length = 0;
+    void *  sin_cache          = nullptr;
+    void *  cos_cache          = nullptr;
+    int64_t position_length    = 0;
     // Properties to check before reusing the sincos cache
-    bool cached = false;
-    float ext_factor = 0.0f;
-    float theta_scale = 0.0f;
-    float freq_scale = 0.0f;
-    float attn_factor = 0.0f;
-    bool is_neox = false;
+    bool    cached             = false;
+    float   ext_factor         = 0.0f;
+    float   theta_scale        = 0.0f;
+    float   freq_scale         = 0.0f;
+    float   attn_factor        = 0.0f;
+    bool    is_neox            = false;
 };
 
 struct ggml_cann_tensor_cache {
     ~ggml_cann_tensor_cache() {
-        if(cache != nullptr) {
+        if (cache != nullptr) {
             ACL_CHECK(aclrtFree(cache));
         }
     }
 
-    void* cache = nullptr;
-    int64_t size = 0;
+    void *  cache = nullptr;
+    int64_t size  = 0;
 };
 
 /**
  * @brief Context for managing CANN backend operations.
  */
 struct ggml_backend_cann_context {
-    int32_t device;                  /**< Device ID. */
-    std::string name;                /**< Name of the device. */
-    std::string description;         /**< Description of the device. */
-    aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    int32_t     device;               /**< Device ID. */
+    std::string name;                 /**< Name of the device. */
+    std::string description;          /**< Description of the device. */
+    aclrtEvent  copy_event = nullptr; /**< Event for managing copy operations. */
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
     ggml_cann_graph_lru_cache graph_lru_cache;
-    bool acl_graph_mode = true;
+    bool                      acl_graph_mode = true;
 #endif
-    cann_task_queue task_queue;
-    bool async_mode;
+    cann_task_queue        task_queue;
+    bool                   async_mode;
     // Rope Cache
-    ggml_cann_rope_cache rope_cache;
+    ggml_cann_rope_cache   rope_cache;
     // Constant Pool
     ggml_cann_tensor_cache rms_norm_one_tensor_cache;
     ggml_cann_tensor_cache rms_norm_zero_tensor_cache;
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = { nullptr }; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
-    explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
+    explicit ggml_backend_cann_context(int device) :
+        device(device),
+        name("CANN" + std::to_string(device)),
+        task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
 
         async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
-        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
-            device, async_mode ? "ON" : "OFF");
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__, device, async_mode ? "ON" : "OFF");
 #ifdef USE_ACL_GRAPH
         acl_graph_mode = parse_bool(get_env("GGML_CANN_ACL_GRAPH").value_or("on"));
-        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n",
-              __func__, device,
-              acl_graph_mode ? "GRAPH" : "EAGER",
-              acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
+        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n", __func__, device, acl_graph_mode ? "GRAPH" : "EAGER",
+                      acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
 #endif
     }
 
@@ -549,8 +543,7 @@ struct ggml_backend_cann_context {
     aclrtStream stream() { return stream(0); }
 
     // TODO: each stream should have a memory pool.
-    std::unique_ptr<ggml_cann_pool>
-        mem_pool; /**< Memory pool for the device. */
+    std::unique_ptr<ggml_cann_pool> mem_pool; /**< Memory pool for the device. */
 
     /**
      * @brief Create a new memory pool for a given device.
@@ -563,7 +556,7 @@ struct ggml_backend_cann_context {
      * @brief Get or create the memory pool for the context.
      * @return Reference to the memory pool.
      */
-    ggml_cann_pool& pool() {
+    ggml_cann_pool & pool() {
         if (mem_pool == nullptr) {
             mem_pool = new_pool_for_device(device);
         }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -126,25 +126,36 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 )
                 if (NOT ARM_MCPU_RESULT)
                     string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
+                    string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}")
+
+                    # on some old GCC we need to read -march=
+                    if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native")
+                        set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}")
+                    elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native")
+                        set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}")
+                    endif()
                 endif()
-                if ("${ARM_MCPU_FLAG}" STREQUAL "")
-                    set(ARM_MCPU_FLAG -mcpu=native)
-                    message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
+
+                if ("${ARM_NATIVE_FLAG}" STREQUAL "")
+                    set(ARM_NATIVE_FLAG -mcpu=native)
+                    message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used")
+                else()
+                    message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}")
                 endif()
 
                 include(CheckCXXSourceRuns)
 
                 function(check_arm_feature tag code)
                     set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-                    set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
+                    set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
                     check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
                     if (GGML_MACHINE_SUPPORTS_${tag})
-                        set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
+                        set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}" PARENT_SCOPE)
                     else()
-                        set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
+                        set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}")
                         check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
                         if (GGML_MACHINE_SUPPORTS_no${tag})
-                            set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
+                            set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}" PARENT_SCOPE)
                         endif()
                     endif()
                     set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
@@ -155,7 +166,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 check_arm_feature(sve     "#include <arm_sve.h>\nint main()  { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
                 check_arm_feature(sme     "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
 
-                list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
+                list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
@@ -466,33 +477,56 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
-        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
-        file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
-        string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/s390/quants.c)
 
-        # TODO: Separation to determine activation of VX/VXE/VXE2
-        if (${S390X_M} MATCHES "8561|8562")
-            message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15)
-        elseif (${S390X_M} MATCHES "3931")
-            message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16)
-        elseif (${S390X_M} MATCHES "9175|9176")
-            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
-            #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
-            message(STATUS "z17 target")
-            list(APPEND ARCH_FLAGS -march=arch15)
-        else()
-            message(STATUS "Unknown target")
-            message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
-            list(APPEND ARCH_FLAGS -march=native -mtune=native)
+        # for native compilation
+        if (GGML_NATIVE)
+            # check machine level to determine target
+            file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
+            string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
+
+            # TODO: Separation to determine activation of VX/VXE/VXE2
+            if (${S390X_M} MATCHES "8561|8562")
+                message(STATUS "z15 target")
+                list(APPEND ARCH_FLAGS -march=z15)
+            elseif (${S390X_M} MATCHES "3931")
+                message(STATUS "z16 target")
+                list(APPEND ARCH_FLAGS -march=z16)
+            elseif (${S390X_M} MATCHES "9175|9176")
+                # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+                #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
+                message(STATUS "z17 target")
+                list(APPEND ARCH_FLAGS -march=arch15)
+            else()
+                message(STATUS "Unknown target")
+                message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
+                list(APPEND ARCH_FLAGS -march=native -mtune=native)
+            endif()
+        # for cross-compilation
+        elseif(GGML_CPU_ALL_VARIANTS)
+            # range through IBM z15 to z17
+            # NOTE: update when a new hardware level is released
+            foreach (ZHW RANGE 15 17)
+                if(DEFINED GGML_INTERNAL_Z${ZHW})
+                    message(STATUS "z${ZHW} cross-compile target")
+                    list(APPEND ARCH_FLAGS -march=z${ZHW})
+                endif()
+            endforeach()
         endif()
 
-        if (GGML_VXE)
-            message(STATUS "VX/VXE/VXE2 enabled")
+        if (GGML_VXE OR GGML_INTERNAL_VXE2)
+            message(STATUS "VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
-            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+            list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
         endif()
+
+        if (GGML_INTERNAL_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
+        endif()
+
+        ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
         list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)

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

@@ -2044,6 +2044,26 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 }
 
+#ifdef __ARM_FEATURE_SVE
+static inline svuint32_t ggml_decode_q4scales_and_mins_for_mmla(const uint32_t * vx_scales) {
+    const svbool_t pg_all   = svptrue_pat_b32(SV_VL4);
+    const svbool_t pg_false = svpfalse_b();            // 0x0000
+    const svbool_t pg_lo_8  = svwhilelt_b8_s32(0,  8); // 0x00ff
+    const svbool_t pg_odd   = svzip1_b32(pg_false, pg_lo_8);
+
+    svuint32_t vutmp_hi, vutmp_lo;
+    svuint32_t vx01 = svld1_u32(pg_lo_8, vx_scales);
+    vutmp_hi = svzip1_u32(vx01, vx01);
+    vutmp_hi = svlsr_n_u32_m(pg_odd, vutmp_hi, 2);
+    vutmp_hi = svreinterpret_u32_u64(svand_n_u64_x(pg_all, svreinterpret_u64_u32(vutmp_hi), UINT64_C(0x303030303f3f3f3f)));
+    const svuint32_t vx2 = svdup_u32(vx_scales[2]);
+    vutmp_lo = svlsr_u32_x(pg_all, vx2, svreinterpret_u32_s32(svindex_s32(-2, 2)));
+    vutmp_lo = svand_n_u32_z(pg_odd, vutmp_lo, UINT32_C(0x0f0f0f0f));
+    svuint32_t vutmp = svorr_u32_z(pg_all, vutmp_hi, vutmp_lo);
+    return vutmp;
+}
+#endif
+
 void ggml_vec_dot_q4_K_q8_K(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) {
     assert(n % QK_K == 0);
 #ifdef __ARM_FEATURE_MATMUL_INT8
@@ -2066,8 +2086,220 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     static const uint32_t kmask3 = 0x03030303;
 
     uint32_t utmp[4];
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+#endif
 
-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        svbool_t pg32_2 = svptrue_pat_b32(SV_VL2);
+
+        const block_q4_K * GGML_RESTRICT vx0 = vx;
+        const block_q8_K * GGML_RESTRICT vy0 = vy;
+        const block_q4_K * GGML_RESTRICT vx1 = (const block_q4_K *) ((const uint8_t*)vx + bx);
+        const block_q8_K * GGML_RESTRICT vy1 = (const block_q8_K *) ((const uint8_t*)vy + by);
+
+        union {
+            uint32_t u32[8];
+            uint64_t u64[4];
+        } new_utmp;
+
+        svfloat32_t sumf1 = svdup_n_f32(0);
+
+        switch (vector_length) {
+            case 128:
+                {
+                    svbool_t pg_false = svpfalse_b();
+                    svbool_t pg_lo_8  = svwhilelt_b8_s32(0,  8);
+                    svbool_t vmins_mask1= svzip1_b32(pg_lo_8, pg_false);
+                    svbool_t vmins_mask2 = svzip1_b32(pg_false, pg_lo_8);
+                    svbool_t pg128_all  = svptrue_pat_b8(SV_VL16);
+                    for (int i = 0; i < nb; ++i) {
+                        svfloat32_t vy_d = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg128_all, vy_d, vx_d);
+                        svfloat32_t vx_dmins = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].dmin)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].dmin)));
+                        svfloat32_t vy_dmins = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t svdmins = svmul_n_f32_x(pg128_all, svmul_f32_x(pg128_all, vy_dmins, vx_dmins), -1);
+                        const uint8_t * GGML_RESTRICT q4_0 = vx0[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_0 = vy0[i].qs;
+                        const uint8_t * GGML_RESTRICT q4_1 = vx1[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_1 = vy1[i].qs;
+                        svint16_t lo = svld1_s16(pg128_all, vy0[i].bsums + 0);
+                        svint16_t hi = svld1_s16(pg128_all, vy0[i].bsums + 8);
+                        svint16_t sum_tmp1 = svuzp1_s16(lo, hi);
+                        svint16_t sum_tmp2 = svuzp2_s16(lo, hi);
+                        svint16_t svq8sums_0 = svadd_s16_x(pg128_all, sum_tmp1, sum_tmp2);
+                        lo = svld1_s16(pg128_all, vy1[i].bsums + 0);
+                        hi = svld1_s16(pg128_all, vy1[i].bsums + 8);
+                        sum_tmp1 = svuzp1(lo, hi);
+                        sum_tmp2 = svuzp2(lo, hi);
+                        svint16_t svq8sums_1 = svadd_s16_x(pg128_all, sum_tmp1, sum_tmp2);
+                        svuint32_t decoded_scales0 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx0[i].scales);
+                        svuint32_t decoded_scales1 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx1[i].scales);
+                        svuint32x2_t decoded_scales = svcreate2_u32(decoded_scales0, decoded_scales1);
+                        svst2_u32(pg128_all, new_utmp.u32, decoded_scales);
+                        svint16_t svmins8_0 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u32(svuzp1_u32(svld1_u32(vmins_mask1, new_utmp.u32+4), svdup_n_u32(0)))));
+                        svint16_t svmins8_1 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u32(svuzp2_u32(svld1_u32(vmins_mask2, new_utmp.u32+4), svdup_n_u32(0)))));
+                        svint32_t svsumfs_tmp1 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_0, svmins8_0));
+                        svint32_t svsumfs_tmp2 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_0, svmins8_1));
+                        svint32_t svsumfs_tmp3 = svtrn1_s32(svsumfs_tmp1, svsumfs_tmp2);
+                        svint32_t svsumfs_tmp4 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_1, svmins8_0));
+                        svint32_t svsumfs_tmp5 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_1, svmins8_1));
+                        svint32_t svsumfs_tmp6 = svtrn1_s32(svsumfs_tmp4, svsumfs_tmp5);
+                        svint32_t svsumfs_tmp7 = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(svsumfs_tmp3), svreinterpret_s64_s32(svsumfs_tmp6)));
+                        svint32_t svsumfs_tmp8 = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(svsumfs_tmp3), svreinterpret_s64_s32(svsumfs_tmp6)));
+                        svint32_t svsumfs_tmp = svadd_s32_x(pg128_all, svsumfs_tmp7, svsumfs_tmp8);
+                        svint32_t svscales, sumi1, sumi2;
+                        svint32_t acc_sumif1 = svdup_n_s32(0);
+                        svint32_t acc_sumif2 = svdup_n_s32(0);
+                        svint8_t q4bytes_0_l, q4bytes_0_h, q4bytes_1_l, q4bytes_1_h, l0, l1, l2, l3,
+                                 q8bytes_0_h, q8bytes_0_l, q8bytes_1_h, q8bytes_1_l, r0, r1, r2, r3;
+#pragma GCC unroll 1
+                        for (int j = 0; j < QK_K/64; ++j) {
+                            q4bytes_0_l = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0), 0xf));
+                            q4bytes_1_l = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1), 0xf));
+                            q4bytes_0_h = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0+16), 0xf));
+                            q4bytes_1_h = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1+16), 0xf));
+                            l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            l3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            q8bytes_0_h = svld1_s8(pg128_all, q8_0);
+                            q8bytes_1_h = svld1_s8(pg128_all, q8_1);
+                            q8bytes_0_l = svld1_s8(pg128_all, q8_0+16);
+                            q8bytes_1_l = svld1_s8(pg128_all, q8_1+16);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            sumi1 = svmmla_s32(svmmla_s32(svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg128_all, svlsl_n_u32_x(pg128_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-1)), 24));
+                            acc_sumif1 = svmla_s32_x(pg128_all, acc_sumif1, svscales, sumi1);
+
+                            q4bytes_0_l = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0), 4));
+                            q4bytes_1_l = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1), 4));
+                            q4bytes_0_h = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0+16), 4));
+                            q4bytes_1_h = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1+16), 4));
+                            l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            l3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            q8bytes_0_h = svld1_s8(pg128_all, q8_0+32);
+                            q8bytes_1_h = svld1_s8(pg128_all, q8_1+32);
+                            q8bytes_0_l = svld1_s8(pg128_all, q8_0+48);
+                            q8bytes_1_l = svld1_s8(pg128_all, q8_1+48);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            sumi2 = svmmla_s32(svmmla_s32(svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg128_all, svlsl_n_u32_x(pg128_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-2)), 24));
+                            acc_sumif2 = svmla_s32_x(pg128_all, acc_sumif2, svscales, sumi2);
+                            q4_0 += 32; q4_1 += 32; q8_0 += 64; q8_1 += 64;
+                        }
+                        sumf1 = svmla_f32_x(pg128_all,
+                                svmla_f32_x(pg128_all,
+                                    sumf1,
+                                    svcvt_f32_x(pg128_all,
+                                        svadd_s32_x(pg128_all, acc_sumif1, acc_sumif2)),
+                                    svsuper_block_scales),
+                                svdmins,
+                                svcvt_f32_s32_x(pg128_all, svsumfs_tmp));
+                    }  //end of for nb
+                } // end of case 128
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    const svbool_t pg8_16 = svptrue_pat_b8(SV_VL16);
+                    const svbool_t pg256_all = svptrue_pat_b8(SV_ALL);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT q4_0 = vx0[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_0 = vy0[i].qs;
+                        const uint8_t * GGML_RESTRICT q4_1 = vx1[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_1 = vy1[i].qs;
+                        svint32_t svscales, sumi1, sumi2;
+                        svint32_t acc_sumif1 = svdup_n_s32(0);
+                        svint32_t acc_sumif2 = svdup_n_s32(0);
+                        svint8_t l0, l1, l2, l3, r0, r1, r2, r3;
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat64_t vy_d_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_d = svreinterpret_f32_f64(svuzp1_f64(vy_d_tmp, vy_d_tmp));
+                        svfloat32_t svsuper_block_scales = svmul_f32_z(pg32_4, vy_d, vx_d);
+                        svfloat32_t vx_dmins = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].dmin)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].dmin)));
+                        svfloat64_t vy_dmins_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_dmins = svreinterpret_f32_f64(svuzp1_f64(vy_dmins_tmp, vy_dmins_tmp));
+                        svfloat32_t svdmins = svmul_n_f32_x(pg32_4, svmul_f32_x(pg32_4, vx_dmins, vy_dmins), -1);
+                        svint16_t rc1 = svuzp1_s16(svld1_s16(pg256_all, vy0[i].bsums), svld1_s16(pg256_all, vy1[i].bsums));
+                        svint16_t rc2 = svuzp2_s16(svld1_s16(pg256_all, vy0[i].bsums), svld1_s16(pg256_all, vy1[i].bsums));
+                        svint16_t svq8sums = svadd_s16_x(pg256_all, rc1, rc2);
+                        svuint32_t decoded_scales0 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx0[i].scales);
+                        svuint32_t decoded_scales1 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx1[i].scales);
+                        svuint32x2_t decoded_scales = svcreate2_u32(decoded_scales0, decoded_scales1);
+                        svst2_u32(pg8_16, new_utmp.u32, decoded_scales);
+                        svint16_t new_svq8sums_0 = svreinterpret_s16_u64(svtrn1_u64(svreinterpret_u64_s16(svq8sums), svreinterpret_u64_s16(svq8sums)));
+                        svint16_t new_svq8sums_1 = svreinterpret_s16_u64(svtrn2_u64(svreinterpret_u64_s16(svq8sums), svreinterpret_u64_s16(svq8sums)));
+                        svuint64_t new_mins_0 = svdup_u64(new_utmp.u64[2]);
+                        svuint64_t new_mins_1 = svdup_u64(new_utmp.u64[3]);
+                        svint16_t new_svmins8_0 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u64(new_mins_0)));
+                        svint16_t new_svmins8_1 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u64(new_mins_1)));
+                        svint64_t dot_prod_0 = svdot_s64(svdup_s64(0), new_svmins8_0, new_svq8sums_0);
+                        svint64_t dot_prod_1 = svdot_s64(dot_prod_0, new_svmins8_1, new_svq8sums_1);
+                        svfloat32_t converted_dot_prod_1 = svcvt_f32_s64_x(pg256_all, dot_prod_1);
+                        svfloat32_t svsumfs_tmp = svuzp1_f32(converted_dot_prod_1, converted_dot_prod_1);
+
+#pragma GCC unroll 1
+                        for (int j = 0; j < QK_K/64; ++j) {
+                            svuint8_t q4bytes_0 = svand_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_0), 0xf);
+                            svuint8_t q4bytes_1 = svand_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_1), 0xf);
+                            svuint8_t q4bytes_2 = svlsr_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_0), 4);
+                            svuint8_t q4bytes_3 = svlsr_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_1), 4);
+                            l0 = svreinterpret_s8_u64(svzip1_u64(svreinterpret_u64_u8(q4bytes_0), svreinterpret_u64_u8(q4bytes_1)));
+                            l1 = svreinterpret_s8_u64(svzip2_u64(svreinterpret_u64_u8(q4bytes_0), svreinterpret_u64_u8(q4bytes_1)));
+                            l2 = svreinterpret_s8_u64(svzip1_u64(svreinterpret_u64_u8(q4bytes_2), svreinterpret_u64_u8(q4bytes_3)));
+                            l3 = svreinterpret_s8_u64(svzip2_u64(svreinterpret_u64_u8(q4bytes_2), svreinterpret_u64_u8(q4bytes_3)));
+                            svint8_t q8bytes_0 = svld1_s8(pg256_all, q8_0);
+                            svint8_t q8bytes_1 = svld1_s8(pg256_all, q8_1);
+                            svint8_t q8bytes_2 = svld1_s8(pg256_all, q8_0+32);
+                            svint8_t q8bytes_3 = svld1_s8(pg256_all, q8_1+32);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_2), svreinterpret_s64_s8(q8bytes_3)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_2), svreinterpret_s64_s8(q8bytes_3)));
+                            sumi1 = svmmla(svmmla(svdup_n_s32(0), r0, l0), r1, l1);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg256_all, svlsl_n_u32_x(pg256_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-1)), 24));
+                            acc_sumif1 = svmla_s32_x(pg256_all, acc_sumif1, svscales, sumi1);
+                            sumi2 = svmmla(svmmla(svdup_n_s32(0), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg256_all, svlsl_n_u32_x(pg256_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-2)), 24));
+                            acc_sumif2 = svmla_s32_x(pg256_all, acc_sumif2, svscales, sumi2);
+                            q4_0 += 32; q4_1 += 32; q8_0 += 64; q8_1 += 64;
+                        }
+                        svint32_t acc_sumif = svadd_s32_x(pg256_all, acc_sumif1, acc_sumif2);
+                        svint32_t swap_acc_sumif = svext_s32(acc_sumif, acc_sumif, 4);
+                        acc_sumif = svadd_s32_x(pg32_4, acc_sumif, swap_acc_sumif);
+                        sumf1 = svmla_f32_x(pg32_4,
+                                svmla_f32_x(pg32_4,
+                                    sumf1,
+                                    svcvt_f32_x(pg32_4, acc_sumif),
+                                    svsuper_block_scales),
+                                svdmins,
+                                svsumfs_tmp);
+                    } // end of for nb
+                } // end of case 256-512
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+
+        svst1_f32(pg32_2, s, sumf1);
+        svst1_f32(pg32_2, s + bs, svreinterpret_f32_u8(svext_u8(svreinterpret_u8_f32(sumf1), svdup_n_u8(0), 8)));
+
+        return;
+    }
+#elif defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q4_K * GGML_RESTRICT x0 = x;
         const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
@@ -2235,7 +2467,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT q4 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
-        const int vector_length = ggml_cpu_get_sve_cnt()*8;
         const svuint8_t m4b = svdup_n_u8(0xf);
         const svint32_t mzero = svdup_n_s32(0);
         svint32_t sumi1 = svdup_n_s32(0);
@@ -2480,7 +2711,201 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     const int nb = n / QK_K;
 
-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+#endif
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const svbool_t pg32_2 = svptrue_pat_b32(SV_VL2);
+
+        svfloat32_t sum = svdup_n_f32(0);
+
+        const block_q6_K * GGML_RESTRICT vx0 = vx;
+        const block_q8_K * GGML_RESTRICT vy0 = vy;
+        const block_q6_K * GGML_RESTRICT vx1 = (const block_q6_K *) ((const uint8_t*)vx + bx);
+        const block_q8_K * GGML_RESTRICT vy1 = (const block_q8_K *) ((const uint8_t*)vy + by);
+
+        switch (vector_length) {
+            case 128:
+                {
+                    const svbool_t pg128_all = svptrue_pat_b8(SV_ALL);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT ql0 = vx0[i].ql;
+                        const uint8_t * GGML_RESTRICT qh0 = vx0[i].qh;
+                        const uint8_t * GGML_RESTRICT ql1 = vx1[i].ql;
+                        const uint8_t * GGML_RESTRICT qh1 = vx1[i].qh;
+                        const int8_t  * GGML_RESTRICT q80 = vy0[i].qs;
+                        const int8_t  * GGML_RESTRICT q81 = vy1[i].qs;
+
+                        const int8_t * GGML_RESTRICT scale0 = vx0[i].scales;
+                        const int8_t * GGML_RESTRICT scale1 = vx1[i].scales;
+
+                        svfloat32_t vy_d = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg128_all, vy_d, vx_d);
+                        // process q8sum summation 128 bit route
+                        const svint16_t q8sums_01 = svld1_s16(pg128_all, vy0[i].bsums);
+                        const svint16_t q8sums_02 = svld1_s16(pg128_all, vy0[i].bsums + 8);
+                        const svint16_t q8sums_11 = svld1_s16(pg128_all, vy1[i].bsums);
+                        const svint16_t q8sums_12 = svld1_s16(pg128_all, vy1[i].bsums + 8);
+                        const svint64x2_t q6scales_0_tmp = svld2_s64(pg128_all, (const int64_t *)scale0);
+                        const svint16_t q6scales_01 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_0_tmp, 0)));
+                        const svint16_t q6scales_02 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_0_tmp, 1)));
+                        const svint64x2_t q6scales_1_tmp = svld2_s64(pg128_all, (const int64_t *)scale1);
+                        const svint16_t q6scales_11 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_1_tmp, 0)));
+                        const svint16_t q6scales_12 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_1_tmp, 1)));
+                        const svint64_t prod = svdup_n_s64(0);
+
+                        svint32_t isum_tmp1 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_01, q6scales_01), q8sums_02, q6scales_02));
+                        svint32_t isum_tmp2 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_01, q6scales_11), q8sums_02, q6scales_12));
+                        svint32_t isum_tmp3 = svtrn1_s32(isum_tmp1, isum_tmp2);
+                        svint32_t isum_tmp4 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_11, q6scales_01), q8sums_12, q6scales_02));
+                        svint32_t isum_tmp5 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_11, q6scales_11), q8sums_12, q6scales_12));
+                        svint32_t isum_tmp6 = svtrn1_s32(isum_tmp4, isum_tmp5);
+                        svint32_t isum_tmp7 = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(isum_tmp3), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp8 = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(isum_tmp3), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t svisum_mins = svadd_s32_x(pg128_all, isum_tmp7, isum_tmp8);
+
+                        // process mmla
+                        svint8_t  l0, l1, r0, r1;
+                        svint32_t isum_tmp = svdup_n_s32(0);
+                        for (int j = 0; j < QK_K/128; ++j) {
+                            for (int k = 0; k < 8; ++k) {
+                                svuint8_t qhbits_0 = svld1_u8(pg128_all, qh0+16*(k%2));
+                                svuint8_t qhbits_1 = svld1_u8(pg128_all, qh1+16*(k%2));
+                                svuint8_t q6bits_0 = svld1_u8(pg128_all, ql0+16*(k%4));
+                                svuint8_t q6bits_1 = svld1_u8(pg128_all, ql1+16*(k%4));
+                                const int ql_pos = (k/4)*4;
+                                svuint8_t q6bytes_0_lo = (ql_pos < 4) ? svand_n_u8_x(pg128_all, q6bits_0, 0xf) : svlsr_n_u8_x(pg128_all, q6bits_0, 4);
+                                svuint8_t q6bytes_1_lo = (ql_pos < 4) ? svand_n_u8_x(pg128_all, q6bits_1, 0xf) : svlsr_n_u8_x(pg128_all, q6bits_1, 4);
+                                const int qh_pos = (k/2)*2;
+                                svuint8_t q6bytes_0_hi = svand_n_u8_x(pg128_all, qhbits_0, 0x3 << qh_pos);
+                                svuint8_t q6bytes_1_hi = svand_n_u8_x(pg128_all, qhbits_1, 0x3 << qh_pos);
+                                svint8_t  q6bytes_0, q6bytes_1;
+                                if (qh_pos <= 4) {
+                                    q6bytes_0 = svreinterpret_s8_u8(svmla_n_u8_x(pg128_all, q6bytes_0_lo, q6bytes_0_hi, 1 << (4 - qh_pos)));
+                                    q6bytes_1 = svreinterpret_s8_u8(svmla_n_u8_x(pg128_all, q6bytes_1_lo, q6bytes_1_hi, 1 << (4 - qh_pos)));
+                                } else {
+                                    q6bytes_0 = svreinterpret_s8_u8(svorr_u8_x(pg128_all, q6bytes_0_lo, svlsr_n_u8_x(pg128_all, q6bytes_0_hi, (qh_pos - 4))));
+                                    q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg128_all, q6bytes_1_lo, svlsr_n_u8_x(pg128_all, q6bytes_1_hi, (qh_pos - 4))));
+                                }
+                                svint8_t  q8bytes_0 = svld1_s8(pg128_all, q80+16*(k%8));
+                                svint8_t  q8bytes_1 = svld1_s8(pg128_all, q81+16*(k%8));
+                                l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                svint32_t svscale = svzip1_s32(svdup_n_s32(scale0[k]), svdup_n_s32(scale1[k]));
+                                isum_tmp = svmla_s32_x(pg128_all, isum_tmp, svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), svscale);
+                            }
+                            qh0 += 32;  qh1 += 32;
+                            ql0 += 64;  ql1 += 64;
+                            q80 += 128; q81 += 128;
+                            scale0 += 8; scale1 += 8;
+                        }
+                        sum = svmla_f32_x(pg128_all, sum,
+                                svcvt_f32_x(pg128_all, svmla_s32_x(pg128_all, isum_tmp,
+                                        svisum_mins, svdup_n_s32(-32))),
+                                svsuper_block_scales);
+                    }
+                } // end of case 128
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg256_all = svptrue_pat_b8(SV_ALL);
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT ql0 = vx0[i].ql;
+                        const uint8_t * GGML_RESTRICT qh0 = vx0[i].qh;
+                        const uint8_t * GGML_RESTRICT ql1 = vx1[i].ql;
+                        const uint8_t * GGML_RESTRICT qh1 = vx1[i].qh;
+                        const int8_t  * GGML_RESTRICT q80 = vy0[i].qs;
+                        const int8_t  * GGML_RESTRICT q81 = vy1[i].qs;
+
+                        const int8_t * GGML_RESTRICT scale0 = vx0[i].scales;
+                        const int8_t * GGML_RESTRICT scale1 = vx1[i].scales;
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat64_t vy_d_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_d = svreinterpret_f32_f64(svuzp1_f64(vy_d_tmp, vy_d_tmp));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg32_4, vy_d, vx_d);
+                        // process q8sum summation 256 bit route
+                        const svint16_t q8sums_0 = svld1_s16(pg256_all, vy0[i].bsums);
+                        const svint16_t q8sums_1 = svld1_s16(pg256_all, vy1[i].bsums);
+                        const svint16_t q6scales_0 = svunpklo_s16(svld1_s8(pg256_all, scale0));
+                        const svint16_t q6scales_1 = svunpklo_s16(svld1_s8(pg256_all, scale1));
+                        const svint64_t prod = svdup_n_s64(0);
+                        svint32_t isum_tmp1  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_0, q6scales_0));
+                        svint32_t isum_tmp2  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_0, q6scales_1));
+                        svint32_t isum_tmp3  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_1, q6scales_0));
+                        svint32_t isum_tmp4  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_1, q6scales_1));
+                        svint32_t isum_tmp5  = svtrn1_s32(isum_tmp1, isum_tmp2);
+                        svint32_t isum_tmp6  = svtrn1_s32(isum_tmp3, isum_tmp4);
+                        svint32_t isum_tmp7  = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(isum_tmp5), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp8  = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(isum_tmp5), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp9  = svadd_s32_x(pg256_all, isum_tmp7, isum_tmp8);
+                        svint32_t isum_tmp10 = svreinterpret_s32_u8(svext_u8(svreinterpret_u8_s32(isum_tmp9), svreinterpret_u8_s32(isum_tmp9), 16));
+                        svint32_t svisum_mins = svadd_s32_z(pg32_4, isum_tmp9, isum_tmp10);
+
+                        // process mmla
+                        svint8_t l0, l1, r0, r1;
+                        svint32_t isum_tmp = svdup_n_s32(0);
+                        for (int j = 0; j < QK_K/128; ++j) {
+                            for (int k = 0; k < 8; k+=2) { // process 2 block
+                                svuint8_t qhbits_0  = svld1_u8(pg256_all, qh0);
+                                svuint8_t qhbits_1  = svld1_u8(pg256_all, qh1);
+                                svuint8_t q6bits_0  = svld1_u8(pg256_all, ql0+32*((k%4)/2));
+                                svuint8_t q6bits_1  = svld1_u8(pg256_all, ql1+32*((k%4)/2));
+                                const int ql_pos = (k/4)*4;
+                                svuint8_t q6bytes_0_lo = (ql_pos < 4) ? svand_n_u8_x(pg256_all, q6bits_0, 0xf) : svlsr_n_u8_x(pg256_all, q6bits_0, 4);
+                                svuint8_t q6bytes_1_lo = (ql_pos < 4) ? svand_n_u8_x(pg256_all, q6bits_1, 0xf) : svlsr_n_u8_x(pg256_all, q6bits_1, 4);
+                                const int qh_pos = (k/2)*2;
+                                svuint8_t q6bytes_0_hi = svand_n_u8_x(pg256_all, qhbits_0, 0x3 << qh_pos);
+                                svuint8_t q6bytes_1_hi = svand_n_u8_x(pg256_all, qhbits_1, 0x3 << qh_pos);
+                                svint8_t  q6bytes_0, q6bytes_1;
+                                if (qh_pos <= 4) {
+                                    q6bytes_0 = svreinterpret_s8_u8(svmla_n_u8_x(pg256_all, q6bytes_0_lo, q6bytes_0_hi, 1 << (4 - qh_pos)));
+                                    q6bytes_1 = svreinterpret_s8_u8(svmla_n_u8_x(pg256_all, q6bytes_1_lo, q6bytes_1_hi, 1 << (4 - qh_pos)));
+                                } else {
+                                    q6bytes_0 = svreinterpret_s8_u8(svorr_u8_x(pg256_all, q6bytes_0_lo, svlsr_n_u8_x(pg256_all, q6bytes_0_hi, (qh_pos - 4))));
+                                    q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg256_all, q6bytes_1_lo, svlsr_n_u8_x(pg256_all, q6bytes_1_hi, (qh_pos - 4))));
+                                }
+                                svint8_t  q8bytes_0 = svld1_s8(pg256_all, q80+32*(k/2));
+                                svint8_t  q8bytes_1 = svld1_s8(pg256_all, q81+32*(k/2));
+                                l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                svint32_t svscale0 = svzip1_s32(svdup_n_s32(scale0[k]), svdup_n_s32(scale1[k]));
+                                svint32_t svscale1 = svzip1_s32(svdup_n_s32(scale0[k+1]), svdup_n_s32(scale1[k+1]));
+                                isum_tmp = svmla_s32_x(pg256_all, isum_tmp, svmmla_s32(svdup_n_s32(0), r0, l0), svscale0);
+                                isum_tmp = svmla_s32_x(pg256_all, isum_tmp, svmmla_s32(svdup_n_s32(0), r1, l1), svscale1);
+                            }
+                            qh0 += 32;  qh1 += 32;
+                            ql0 += 64;  ql1 += 64;
+                            q80 += 128; q81 += 128;
+                            scale0 += 8; scale1 += 8;
+                        } // end of for
+                        svint32_t swap_isum_tmp = svext_s32(isum_tmp, isum_tmp, 4);
+                        isum_tmp = svadd_s32_x(pg32_4, isum_tmp, swap_isum_tmp);
+                        sum = svmla_f32_x(pg32_4, sum,
+                                svcvt_f32_x(pg32_4, svmla_s32_x(pg32_4, isum_tmp,
+                                        svisum_mins, svdup_n_s32(-32))),
+                                svsuper_block_scales);
+                    }
+                } // end of case 256
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        } // end of switch
+
+        svst1_f32(pg32_2, s, sum);
+        svst1_f32(pg32_2, s + bs, svreinterpret_f32_u8(svext_u8(svreinterpret_u8_f32(sum), svdup_n_u8(0), 8)));
+
+        return;
+    }
+#elif defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q6_K * GGML_RESTRICT x0 = x;
         const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
@@ -2594,27 +3019,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             // adjust bias, apply superblock scale
             {
                 int32_t bias[4];
-#ifdef __ARM_FEATURE_SVE
-                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
-                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
-                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
-                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
-                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
-                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
-                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
-                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
-                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
-                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
-                const svint64_t zero = svdup_n_s64(0);
-                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
-                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
-                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
-                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
-#else
                 // NEON doesn't support int16 dot product, fallback to separated mul and add
                 const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
                 const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
@@ -2646,7 +3050,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                            vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
                 bias[3] = vaddvq_s32(prod);
 
-#endif
                 const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
 
                 const float32x4_t superblock_scale = {
@@ -2672,7 +3075,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 
 #ifdef __ARM_FEATURE_SVE
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
     float sum = 0;
     svuint8_t m4b = svdup_n_u8(0xf);
     svint32_t vzero = svdup_n_s32(0);

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

@@ -700,7 +700,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+        const float ft0 = GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d);
+        const __m128 d_0_1 = (__m128)(v4f32){ft0, ft0, ft0, ft0};
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -714,11 +715,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         bx_1 = __lsx_vsub_b(bx_1, off);
         const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
 
-        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
+        const float ft1 = GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d);
+        const __m128 d_2_3 = (__m128)(v4f32){ft1, ft1, ft1, ft1};
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 

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

@@ -580,16 +580,19 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
             uint8_t *patmp = atmp;
             int vsums;
-            int tmp;
+            int tmp, t1, t2, t3, t4, t5, t6, t7;
             __asm__ __volatile__(
                 "vsetivli zero, 16, e8, m1\n\t"
                 "vmv.v.x v8, zero\n\t"
+                "lb zero, 15(%[sc])\n\t"
                 "vle8.v v1, (%[sc])\n\t"
+                "vle8.v v2, (%[bsums])\n\t"
+                "addi %[tmp], %[bsums], 16\n\t"
                 "vand.vi v0, v1, 0xF\n\t"
                 "vsrl.vi v1, v1, 4\n\t"
+                "vle8.v v3, (%[tmp])\n\t"
                 "vse8.v v0, (%[scale])\n\t"
                 "vsetivli zero, 16, e16, m2\n\t"
-                "vle16.v v2, (%[bsums])\n\t"
                 "vzext.vf2 v0, v1\n\t"
                 "vwmul.vv v4, v0, v2\n\t"
                 "vsetivli zero, 16, e32, m4\n\t"
@@ -608,46 +611,89 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             for (int j = 0; j < QK_K/128; ++j) {
                 __asm__ __volatile__(
-                    "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "lb zero, 31(%[q2])\n\t"
+                    "addi %[tmp], %[q2], 16\n\t"
+                    "addi %[t1], %[q8], 16\n\t"
+                    "vsetivli zero, 16, e8, m1\n\t"
                     "vle8.v v0, (%[q2])\n\t"
+                    "vle8.v v1, (%[tmp])\n\t"
                     "vsrl.vi v2, v0, 2\n\t"
+                    "vsrl.vi v3, v1, 2\n\t"
                     "vsrl.vi v4, v0, 4\n\t"
-                    "vsrl.vi v6, v0, 6\n\t"
-                    "vand.vi v0, v0, 0x3\n\t"
-                    "vand.vi v2, v2, 0x3\n\t"
-                    "vand.vi v4, v4, 0x3\n\t"
-                    "vsetvli zero, %[vl128], e8, m8\n\t"
+                    "addi %[tmp], %[q8], 32\n\t"
                     "vle8.v v8, (%[q8])\n\t"
-                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vle8.v v9, (%[t1])\n\t"
+                    "addi %[t1], %[t1], 32\n\t"
+                    "vsrl.vi v5, v1, 4\n\t"
+                    "vsrl.vi v6, v0, 6\n\t"
+                    "vsrl.vi v7, v1, 6\n\t"
+                    "vle8.v v10, (%[tmp])\n\t"
+                    "vle8.v v11, (%[t1])\n\t"
+                    "addi %[tmp], %[tmp], 32\n\t"
+                    "addi %[t1], %[t1], 32\n\t"
+                    "vand.vi v0, v0, 0x3\n\t"
+                    "vand.vi v1, v1, 0x3\n\t"
+                    "vand.vi v2, v2, 0x3\n\t"
+                    "vle8.v v12, (%[tmp])\n\t"
+                    "vle8.v v13, (%[t1])\n\t"
+                    "addi %[tmp], %[tmp], 32\n\t"
+                    "addi %[t1], %[t1], 32\n\t"
+                    "vand.vi v3, v3, 0x3\n\t"
+                    "vand.vi v4, v4, 0x3\n\t"
+                    "vand.vi v5, v5, 0x3\n\t"
+                    "vle8.v v14, (%[tmp])\n\t"
+                    "vle8.v v15, (%[t1])\n\t"
                     "vwmul.vv v16, v0, v8\n\t"
+                    "vwmul.vv v18, v1, v9\n\t"
+                    "vwmul.vv v20, v2, v10\n\t"
+                    "vwmul.vv v22, v3, v11\n\t"
                     "vwmul.vv v24, v4, v12\n\t"
-                    "vsetivli zero, 16, e16, m2\n\t"
+                    "vwmul.vv v26, v5, v13\n\t"
+                    "vwmul.vv v28, v6, v14\n\t"
+                    "vwmul.vv v30, v7, v15\n\t"
+                    "vsetivli zero, 8, e16, m1\n\t"
                     "vmv.v.x v0, zero\n\t"
-                    "vwredsum.vs v10, v16, v0\n\t"
+                    "lbu %[tmp], 0(%[scale])\n\t"
+                    "vwredsum.vs v8, v16, v0\n\t"
                     "vwredsum.vs v9, v18, v0\n\t"
-                    "vwredsum.vs v8, v20, v0\n\t"
-                    "vwredsum.vs v7, v22, v0\n\t"
-                    "vwredsum.vs v11, v24, v0\n\t"
-                    "vwredsum.vs v12, v26, v0\n\t"
-                    "vwredsum.vs v13, v28, v0\n\t"
-                    "vwredsum.vs v14, v30, v0\n\t"
+                    "lbu %[t1], 1(%[scale])\n\t"
+                    "vwredsum.vs v10, v20, v0\n\t"
+                    "vwredsum.vs v11, v22, v0\n\t"
+                    "lbu %[t2], 2(%[scale])\n\t"
+                    "vwredsum.vs v12, v24, v0\n\t"
+                    "vwredsum.vs v13, v26, v0\n\t"
+                    "lbu %[t3], 3(%[scale])\n\t"
+                    "vwredsum.vs v14, v28, v0\n\t"
+                    "vwredsum.vs v15, v30, v0\n\t"
+                    "lbu %[t4], 4(%[scale])\n\t"
+                    "vwredsum.vs v8, v17, v8\n\t"
+                    "vwredsum.vs v9, v19, v9\n\t"
+                    "lbu %[t5], 5(%[scale])\n\t"
+                    "vwredsum.vs v10, v21, v10\n\t"
+                    "vwredsum.vs v11, v23, v11\n\t"
+                    "lbu %[t6], 6(%[scale])\n\t"
+                    "vwredsum.vs v12, v25, v12\n\t"
+                    "vwredsum.vs v13, v27, v13\n\t"
+                    "lbu %[t7], 7(%[scale])\n\t"
+                    "vwredsum.vs v14, v29, v14\n\t"
+                    "vwredsum.vs v15, v31, v15\n\t"
                     "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v10, v9, 1\n\t"
-                    "vslideup.vi v8, v7, 1\n\t"
-                    "vslideup.vi v11, v12, 1\n\t"
-                    "vslideup.vi v13, v14, 1\n\t"
-                    "vslideup.vi v10, v8, 2\n\t"
-                    "vslideup.vi v11, v13, 2\n\t"
-                    "vsetivli zero, 8, e32, m2\n\t"
-                    "vle8.v v15, (%[scale])\n\t"
-                    "vzext.vf4 v12, v15\n\t"
-                    "vmul.vv v10, v10, v12\n\t"
-                    "vredsum.vs v0, v10, v0\n\t"
+                    "vmul.vx v0, v8, %[tmp]\n\t"
+                    "vmul.vx v1, v9, %[t1]\n\t"
+                    "vmacc.vx v0, %[t2], v10\n\t"
+                    "vmacc.vx v1, %[t3], v11\n\t"
+                    "vmacc.vx v0, %[t4], v12\n\t"
+                    "vmacc.vx v1, %[t5], v13\n\t"
+                    "vmacc.vx v0, %[t6], v14\n\t"
+                    "vmacc.vx v1, %[t7], v15\n\t"
                     "vmv.x.s %[tmp], v0\n\t"
-                    "add %[isum], %[isum], %[tmp]"
-                    : [tmp] "=&r" (tmp), [isum] "+&r" (isum)
+                    "vmv.x.s %[t1], v1\n\t"
+                    "add %[isum], %[isum], %[tmp]\n\t"
+                    "add %[isum], %[isum], %[t1]"
+                    : [tmp] "=&r" (tmp), [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3)
+                    , [t4] "=&r" (t4), [t5] "=&r" (t5), [t6] "=&r" (t6), [t7] "=&r" (t7)
+                    , [isum] "+&r" (isum)
                     : [q2] "r" (q2), [scale] "r" (patmp), [q8] "r" (q8)
-                    , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
                     : "memory"
                     , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
                     , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
@@ -929,7 +975,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const  int8_t * restrict q8 = y[i].qs;
 
             int8_t * scale = (int8_t *)utmp;
-            int tmp;
+            int tmp, t1, t2, t3, t4, t5, t6, t7;
             __asm__ __volatile__(
                 "vsetivli zero, 12, e8, m1\n\t"
                 "vle8.v v0, (%[s6b])\n\t"
@@ -967,19 +1013,23 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             int isum = 0;
             for (int j = 0; j < QK_K; j += 128) {
                 __asm__ __volatile__(
+                    "lb zero, 31(%[q3])\n\t"
                     "vsetvli zero, %[vl32], e8, m2, ta, mu\n\t"
                     "vle8.v v8, (%[q3])\n\t"
                     "vsrl.vi v10, v8, 2\n\t"
                     "vsrl.vi v12, v8, 4\n\t"
                     "vsrl.vi v14, v8, 6\n\t"
+                    "lb zero, 64(%[q8])\n\t"
                     "vand.vi v8, v8, 3\n\t"
                     "vand.vi v10, v10, 3\n\t"
                     "vand.vi v12, v12, 3\n\t"
                     "vle8.v v2, (%[qh])\n\t"
+                    "lb zero, 127(%[q8])\n\t"
                     "vand.vx v4, v2, %[m]\n\t"
                     "slli %[m], %[m], 1\n\t"
                     "vmseq.vx v0, v4, zero\n\t"
                     "vadd.vi v8, v8, -4, v0.t\n\t"
+                    "lb zero, 0(%[q8])\n\t"
                     "vand.vx v4, v2, %[m]\n\t"
                     "slli %[m], %[m], 1\n\t"
                     "vmseq.vx v0, v4, zero\n\t"
@@ -994,34 +1044,43 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     "vadd.vi v14, v14, -4, v0.t\n\t"
                     "vsetvli zero, %[vl128], e8, m8\n\t"
                     "vle8.v v0, (%[q8])\n\t"
+                    "lb %[tmp], 0(%[scale])\n\t"
+                    "lb %[t1], 1(%[scale])\n\t"
+                    "lb %[t2], 2(%[scale])\n\t"
+                    "lb %[t3], 3(%[scale])\n\t"
                     "vsetvli zero, %[vl64], e8, m4\n\t"
                     "vwmul.vv v16, v0, v8\n\t"
                     "vwmul.vv v24, v4, v12\n\t"
                     "vsetivli zero, 16, e16, m2\n\t"
                     "vmv.v.x v0, zero\n\t"
-                    "vwredsum.vs v10, v16, v0\n\t"
+                    "vwredsum.vs v8, v16, v0\n\t"
+                    "lb %[t4], 4(%[scale])\n\t"
+                    "lb %[t5], 5(%[scale])\n\t"
                     "vwredsum.vs v9, v18, v0\n\t"
-                    "vwredsum.vs v8, v20, v0\n\t"
-                    "vwredsum.vs v7, v22, v0\n\t"
-                    "vwredsum.vs v11, v24, v0\n\t"
-                    "vwredsum.vs v12, v26, v0\n\t"
-                    "vwredsum.vs v13, v28, v0\n\t"
-                    "vwredsum.vs v14, v30, v0\n\t"
+                    "vwredsum.vs v10, v20, v0\n\t"
+                    "vwredsum.vs v11, v22, v0\n\t"
+                    "vwredsum.vs v12, v24, v0\n\t"
+                    "lb %[t6], 6(%[scale])\n\t"
+                    "lb %[t7], 7(%[scale])\n\t"
+                    "vwredsum.vs v13, v26, v0\n\t"
+                    "vwredsum.vs v14, v28, v0\n\t"
+                    "vwredsum.vs v15, v30, v0\n\t"
                     "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v10, v9, 1\n\t"
-                    "vslideup.vi v8, v7, 1\n\t"
-                    "vslideup.vi v11, v12, 1\n\t"
-                    "vslideup.vi v13, v14, 1\n\t"
-                    "vslideup.vi v10, v8, 2\n\t"
-                    "vslideup.vi v11, v13, 2\n\t"
-                    "vsetivli zero, 8, e32, m2\n\t"
-                    "vle8.v v15, (%[scale])\n\t"
-                    "vsext.vf4 v12, v15\n\t"
-                    "vmul.vv v10, v10, v12\n\t"
-                    "vredsum.vs v0, v10, v0\n\t"
+                    "vmul.vx v0, v8, %[tmp]\n\t"
+                    "vmul.vx v1, v9, %[t1]\n\t"
+                    "vmacc.vx v0, %[t2], v10\n\t"
+                    "vmacc.vx v1, %[t3], v11\n\t"
+                    "vmacc.vx v0, %[t4], v12\n\t"
+                    "vmacc.vx v1, %[t5], v13\n\t"
+                    "vmacc.vx v0, %[t6], v14\n\t"
+                    "vmacc.vx v1, %[t7], v15\n\t"
                     "vmv.x.s %[tmp], v0\n\t"
-                    "add %[isum], %[isum], %[tmp]"
-                    : [tmp] "=&r" (tmp), [m] "+&r" (m), [isum] "+&r" (isum)
+                    "vmv.x.s %[t1], v1\n\t"
+                    "add %[isum], %[isum], %[tmp]\n\t"
+                    "add %[isum], %[isum], %[t1]"
+                    : [tmp] "=&r" (tmp), [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3)
+                    , [t4] "=&r" (t4), [t5] "=&r" (t5), [t6] "=&r" (t6), [t7] "=&r" (t7)
+                    , [m] "+&r" (m), [isum] "+&r" (isum)
                     : [vl128] "r" (128), [vl64] "r" (64), [vl32] "r" (32)
                     , [q3] "r" (q3), [qh] "r" (qh), [scale] "r" (scale), [q8] "r" (q8)
                     : "memory"

ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp

@@ -0,0 +1,50 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__s390x__)
+#include <sys/auxv.h>
+
+// find hwcap bits in asm/elf.h
+#ifndef HWCAP_VXRS_EXT2
+#define HWCAP_VXRS_EXT2 (1 << 15)
+#endif
+
+#ifndef HWCAP_NNPA
+#define HWCAP_NNPA (1 << 20)
+#endif
+
+struct s390x_features {
+    bool has_vxe2 = false;
+    bool has_nnpa = false;
+
+    s390x_features() {
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        // NOTE: use hwcap2 with DFLT for z17 and later
+        // uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_vxe2 = !!(hwcap & HWCAP_VXRS_EXT2);
+        has_nnpa = !!(hwcap & HWCAP_NNPA);
+    }
+};
+
+static int ggml_backend_cpu_s390x_score() {
+    int score = 1;
+    s390x_features sf;
+
+// IBM z15 / LinuxONE 3
+#ifdef GGML_USE_VXE2
+    if (!sf.has_vxe2) { return 0; }
+    score += 1 << 1;
+#endif
+
+// IBM z16 / LinuxONE 4 and z17 / LinuxONE 5
+#ifdef GGML_USE_NNPA
+    if (!sf.has_nnpa) { return 0; }
+    score += 1 << 2;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_s390x_score)
+
+#endif  // __s390x__

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

@@ -68,7 +68,7 @@ struct ggml_compute_params {
 #endif  // __VXE2__
 #endif  // __s390x__ && __VEC__
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
 #include <sys/prctl.h>
 #endif
 
@@ -500,13 +500,15 @@ inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
 
 #endif
 
-#if defined(__loongarch_asx)
+#if defined(__loongarch_sx)
 /* float type data load instructions */
 static __m128 __lsx_vreplfr2vr_s(const float val) {
     v4f32 res = {val, val, val, val};
     return (__m128)res;
 }
+#endif
 
+#if defined(__loongarch_asx)
 static __m256 __lasx_xvreplfr2vr_s(const float val) {
     v8f32 res = {val, val, val, val, val, val, val, val};
     return (__m256)res;

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

@@ -689,8 +689,13 @@ bool ggml_is_numa(void) {
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__linux__)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#else
+    // TODO: add support of SVE for non-linux systems
+#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
+#endif
 #endif
 }
 
@@ -1608,13 +1613,8 @@ static void ggml_compute_forward_mul_mat_id(
             chunk_size = 64;
         }
 
-#if defined(__aarch64__)
-        // disable for ARM
-        const bool disable_chunking = true;
-#else
         // disable for NUMA
         const bool disable_chunking = ggml_is_numa();
-#endif // defined(__aarch64__)
 
         int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
         int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
@@ -1807,22 +1807,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_cont(params, tensor);
             } break;
-        case GGML_OP_RESHAPE:
-            {
-                ggml_compute_forward_reshape(params, tensor);
-            } break;
-        case GGML_OP_VIEW:
-            {
-                ggml_compute_forward_view(params, tensor);
-            } break;
-        case GGML_OP_PERMUTE:
-            {
-                ggml_compute_forward_permute(params, tensor);
-            } break;
-        case GGML_OP_TRANSPOSE:
-            {
-                ggml_compute_forward_transpose(params, tensor);
-            } break;
         case GGML_OP_GET_ROWS:
             {
                 ggml_compute_forward_get_rows(params, tensor);
@@ -2042,6 +2026,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 // nop
             } break;
+        case GGML_OP_RESHAPE:
+            {
+                // nop
+            } break;
+        case GGML_OP_PERMUTE:
+            {
+                // nop
+            } break;
+        case GGML_OP_VIEW:
+            {
+                // nop
+            } break;
+        case GGML_OP_TRANSPOSE:
+            {
+                // nop
+            } break;
         case GGML_OP_COUNT:
             {
                 GGML_ABORT("fatal error");
@@ -2179,6 +2179,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     {
                         n_tasks = 1;
                     } break;
@@ -2880,6 +2884,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
+        if (ggml_op_is_empty(node->op)) {
+            // skip NOPs
+            continue;
+        }
+
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&
@@ -3558,13 +3567,17 @@ void ggml_cpu_init(void) {
 #ifdef GGML_USE_OPENMP
             //if (!getenv("OMP_WAIT_POLICY")) {
             //    // set the wait policy to active, so that OpenMP threads don't sleep
-            //    putenv("OMP_WAIT_POLICY=active");
+            //    setenv("OMP_WAIT_POLICY", "active", 0)
             //}
 
             if (!getenv("KMP_BLOCKTIME")) {
                 // set the time to wait before sleeping a thread
                 // this is less aggressive than setting the wait policy to active, but should achieve similar results in most cases
-                putenv("KMP_BLOCKTIME=200"); // 200ms
+#ifdef _WIN32
+                _putenv_s("KMP_BLOCKTIME", "200"); // 200ms
+#else
+                setenv("KMP_BLOCKTIME", "200", 0); // 200ms
+#endif
             }
 #endif
         }

ggml/src/ggml-cpu/ops.cpp

@@ -4455,46 +4455,6 @@ void ggml_compute_forward_cont(
     ggml_compute_forward_dup(params, dst);
 }
 
-// ggml_compute_forward_reshape
-
-void ggml_compute_forward_reshape(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_view
-
-void ggml_compute_forward_view(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_permute
-
-void ggml_compute_forward_permute(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_transpose
-
-void ggml_compute_forward_transpose(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
 // ggml_compute_forward_get_rows
 
 static void ggml_compute_forward_get_rows_q(
@@ -5474,7 +5434,7 @@ static void ggml_rope_cache_init(
 }
 
 static void ggml_mrope_cache_init(
-     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool is_imrope, bool indep_sects,
      float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
@@ -5509,14 +5469,26 @@ static void ggml_mrope_cache_init(
         }
 
         float theta = theta_t;
-        if (sector >= sections[0] && sector < sec_w) {
-            theta = theta_h;
-        }
-        else if (sector >= sec_w && sector < sec_w + sections[2]) {
-            theta = theta_w;
-        }
-        else if (sector >= sec_w + sections[2]) {
-            theta = theta_e;
+        if (is_imrope) { // qwen3vl apply interleaved mrope
+            if (sector % 3 == 1 && sector < 3 * sections[1]) {
+                theta = theta_h;
+            } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
+                theta = theta_w;
+            } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                theta = theta_t;
+            } else {
+                theta = theta_e;
+            }
+        } else {
+            if (sector >= sections[0] && sector < sec_w) {
+                theta = theta_h;
+            }
+            else if (sector >= sec_w && sector < sec_w + sections[2]) {
+                theta = theta_w;
+            }
+            else if (sector >= sec_w + sections[2]) {
+                theta = theta_e;
+            }
         }
 
         rope_yarn(
@@ -5589,6 +5561,7 @@ static void ggml_compute_forward_rope_f32(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5627,7 +5600,7 @@ static void ggml_compute_forward_rope_f32(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -5775,6 +5748,7 @@ static void ggml_compute_forward_rope_f16(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5813,7 +5787,7 @@ static void ggml_compute_forward_rope_f16(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -7070,7 +7044,11 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
     const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
 
 #ifdef GGML_SIMD
-    const int64_t pkg_size = GGML_F32_EPR;
+    #if defined(__ARM_FEATURE_SVE)
+        const int64_t pkg_size = svcntw();
+    #else
+        const int64_t pkg_size = GGML_F32_EPR;
+    #endif
     const int64_t pkg_count = c / pkg_size;
     const int64_t c_pkg_end = pkg_count * pkg_size;
 #else
@@ -7493,10 +7471,17 @@ static void ggml_compute_forward_upscale_f32(
     float sf1 = (float)ne1/src0->ne[1];
     float sf2 = (float)ne2/src0->ne[2];
     float sf3 = (float)ne3/src0->ne[3];
+    float pixel_offset = 0.5f;
 
     const int32_t mode_flags = ggml_get_op_params_i32(dst, 0);
     const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
 
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        pixel_offset = 0.0f;
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+    }
+
     if (mode == GGML_SCALE_MODE_NEAREST) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -7516,13 +7501,6 @@ static void ggml_compute_forward_upscale_f32(
             }
         }
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        float pixel_offset = 0.5f;
-        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            pixel_offset = 0.0f;
-            sf0 = (float)(ne0 - 1) / (src0->ne[0] - 1);
-            sf1 = (float)(ne1 - 1) / (src0->ne[1] - 1);
-        }
-
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
             for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
@@ -7557,6 +7535,51 @@ static void ggml_compute_forward_upscale_f32(
 
                         const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy;
 
+                        float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *y_dst = val;
+                    }
+                }
+            }
+        }
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+        const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
+        auto weight1 = [a](float x) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+        auto weight2 = [a](float x) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+        auto bicubic = [=](float p0, float p1, float p2, float p3, float x) {
+            const float w0 = weight2(x + 1);
+            const float w1 = weight1(x + 0);
+            const float w2 = weight1(1 - x);
+            const float w3 = weight2(2 - x);
+            return p0*w0 + p1*w1 + p2*w2 + p3*w3;
+        };
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset;
+                    const int64_t y0 = (int64_t)floorf(y);
+                    const float dy = y - (float)y0;
+
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset;
+                        const int64_t x0 = (int64_t)floorf(x);
+                        const float dx = x - (float)x0;
+
+                        auto p = [=](int64_t x_off, int64_t y_off) -> float {
+                            int64_t i00 = std::max(int64_t(0), std::min(x0 + x_off, ne00 - 1));
+                            int64_t i01 = std::max(int64_t(0), std::min(y0 + y_off, ne01 - 1));
+                            return *(const float *)((const char *)src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                        };
+
+                        const float val = bicubic(
+                            bicubic(p(-1,-1), p(0,-1), p(1,-1), p(2,-1), dx),
+                            bicubic(p(-1, 0), p(0, 0), p(1, 0), p(2, 0), dx),
+                            bicubic(p(-1, 1), p(0, 1), p(1, 1), p(2, 1), dx),
+                            bicubic(p(-1, 2), p(0, 2), p(1, 2), p(2, 2), dx), dy);
+
                         float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
                         *y_dst = val;
                     }
@@ -7909,10 +7932,10 @@ void ggml_compute_forward_argsort(
 
 // ggml_compute_forward_flash_attn_ext
 
-static void ggml_compute_forward_flash_attn_ext_f16(
+static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
+        ggml_tensor * dst,
+        int ir0, int ir1) {
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -7928,9 +7951,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
-    const int ith = params->ith;
-    const int nth = params->nth;
-
     const int64_t DK = nek0;
     const int64_t DV = nev0;
     const int64_t N  = neq1;
@@ -7964,16 +7984,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
     // parallelize by q rows using ggml_vec_dot_f32
 
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
     float scale         = 1.0f;
     float max_bias      = 0.0f;
     float logit_softcap = 0.0f;
@@ -8000,6 +8010,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
     GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
+    int ith = params->ith;
+
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -8147,6 +8159,91 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     }
 }
 
+static void ggml_compute_forward_flash_attn_ext_f16(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // parallelize by q rows using ggml_vec_dot_f32
+
+    // total rows in q
+    const int64_t nr = neq1*neq2*neq3;
+
+    // rows per thread
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // disable for NUMA
+    const bool disable_chunking = ggml_is_numa();
+
+    // 4x chunks per thread
+    int nth_scaled = nth * 4;
+    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+    if (nth == 1 || nchunk < nth || disable_chunking) {
+        nchunk = nth;
+    }
+
+    if (ith == 0) {
+        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+        ggml_threadpool_chunk_set(params->threadpool, nth);
+    }
+
+    ggml_barrier(params->threadpool);
+
+    // The number of elements in each chunk
+    const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+    // The first chunk comes from our thread_id, the rest will get auto-assigned.
+    int current_chunk = ith;
+
+    while (current_chunk < nchunk) {
+        const int64_t ir0 = dr * current_chunk;
+        const int64_t ir1 = MIN(ir0 + dr, nr);
+
+        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+
+        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+    }
+}
+
 void ggml_compute_forward_flash_attn_ext(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8993,6 +9090,22 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_exp(params, dst);
             } break;
+        case GGML_UNARY_OP_FLOOR:
+            {
+                ggml_compute_forward_floor(params, dst);
+            } break;
+        case GGML_UNARY_OP_CEIL:
+            {
+                ggml_compute_forward_ceil(params, dst);
+            } break;
+        case GGML_UNARY_OP_ROUND:
+            {
+                ggml_compute_forward_round(params, dst);
+            } break;
+        case GGML_UNARY_OP_TRUNC:
+            {
+                ggml_compute_forward_trunc(params, dst);
+            } break;
         case GGML_UNARY_OP_XIELU:
             {
                 ggml_compute_forward_xielu(params, dst);

ggml/src/ggml-cpu/ops.h

@@ -51,10 +51,6 @@ void ggml_compute_forward_scale(const struct ggml_compute_params * params, struc
 void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/repack.cpp

@@ -1600,6 +1600,32 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         return false;
     }
 
+    void forward_mul_mat_one_chunk(ggml_compute_params * params, ggml_tensor * op, int64_t src0_start, int64_t src0_end) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const void * src1_wdata      = params->wdata;
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                    src0_end - src0_start);
+        }
+    }
+
     void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
         const ggml_tensor * src0 = op->src[0];
         const ggml_tensor * src1 = op->src[1];
@@ -1643,31 +1669,62 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
         }
 
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        int64_t nr = ggml_nrows(op->src[0]);
+        int nth_scaled = nth * 4;
+        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+        // Ensure minimum chunk size to avoid alignment issues with high thread counts
+        // Minimum chunk size should be at least NB_COLS to prevent overlapping chunks after alignment
+        const int64_t min_chunk_size = NB_COLS;
+        if (nchunk > 0 && (nr / nchunk) < min_chunk_size && nr >= min_chunk_size) {
+            nchunk = (nr + min_chunk_size - 1) / min_chunk_size;
+        }
+
+        if (nth == 1 || nchunk < nth || disable_chunking) {
+            nchunk = nth;
+        }
+
+        // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
+        // This prevents creating too many tiny chunks that could overlap after alignment
+        const int64_t max_nchunk = (nr + min_chunk_size - 1) / min_chunk_size;
+        if (nchunk > max_nchunk) {
+            nchunk = max_nchunk;
+        }
+
+        if (ith == 0) {
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
         ggml_barrier(params->threadpool);
 
-        const void * src1_wdata      = params->wdata;
-        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
-        int64_t      src0_start      = (ith * ne01) / nth;
-        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
-        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
-        if (src0_start >= src0_end) {
-            return;
-        }
+        // The first chunk comes from our thread_id, the rest will get auto-assigned.
+        int current_chunk = ith;
 
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
-            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                    src0_end - src0_start);
+        while (current_chunk < nchunk) {
+            int64_t src0_start = (current_chunk * ne01) / nchunk;
+            int64_t src0_end   = ((current_chunk + 1) * ne01) / nchunk;
+
+            // Align boundaries to NB_COLS - round up to ensure all data is included
+            // The chunk size limiting above ensures chunks are large enough to prevent overlaps
+            src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+            src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
+            if (src0_end > ne01) {
+                src0_end = ne01;
+            }
+
+            if (src0_start >= src0_end) {
+                break;
+            }
+
+            forward_mul_mat_one_chunk(params, dst, src0_start, src0_end);
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
     }
 
@@ -1772,8 +1829,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int64_t src0_cur_start = (ith * ne01) / nth;
             int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
 
+            // Align boundaries to NB_COLS - round up to ensure all data is included
             src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
             src0_cur_end   = (src0_cur_end   % NB_COLS) ? src0_cur_end   + NB_COLS - (src0_cur_end   % NB_COLS) : src0_cur_end;
+            if (src0_cur_end > ne01) {
+                src0_cur_end = ne01;
+            }
 
             if (src0_cur_start >= src0_cur_end) {
                 return;

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

@@ -956,7 +956,7 @@ do {                                                              \
 
 #define GGML_F32Cx8          __m256
 #define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
-#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
+#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
 
 static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
     __m256i a;
@@ -999,34 +999,34 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 
 #define GGML_F32x4         __m128
 #define GGML_F32x4_ZERO    (__m128)__lsx_vldi(0)
-#define GGML_F32x4_SET1(x) (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32x4_SET1(x) (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32x4_LOAD(x) (__m128)__lsx_vld((x), 0)
 #define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
-#define GGML_F32x4_REDUCE(res, x)                                                     \
-{                                                                                     \
-    int offset = GGML_F32_ARR >> 1;                                                   \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    const __m128 t0 = (__m128)__lsx_vshuf4i_w(tmp, 0x88);                                     \
-    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    res             = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
+
+#define GGML_F32x4_REDUCE(res, x)                               \
+{                                                               \
+    int offset = GGML_F32_ARR >> 1;                             \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    __m128i t0 = __lsx_vpickev_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128i t1 = __lsx_vpickod_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128 t2 = __lsx_vfadd_s((__m128)t0, (__m128)t1);          \
+    __m128i t3 = __lsx_vpickev_w((__m128i)t2, (__m128i)t2);     \
+    __m128i t4 = __lsx_vpickod_w((__m128i)t2, (__m128i)t2);     \
+    __m128 t5 = __lsx_vfadd_s((__m128)t3, (__m128)t4);          \
+    res = (ggml_float) ((v4f32)t5)[0];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -1068,7 +1068,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
 #define GGML_F32Cx4             __m128
 #define GGML_F32Cx4_ZERO        (__m128)__lsx_vldi(0)
-#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32Cx4_LOAD(x)     (__m128)__lsx_f16x4_load(x)
 #define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
 #define GGML_F32Cx4_FMA         GGML_F32x4_FMA

ggml/src/ggml-cpu/spacemit/ime.cpp

@@ -485,8 +485,9 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
             int32_t          start                  = ith * task_per_thread;
             int32_t          end                    = std::min((ith + 1) * task_per_thread, task_count);
             for (int32_t compute_idx = start; compute_idx < end; compute_idx++) {
-                int32_t                             gemm_idx = compute_idx / block_size_m;
-                int32_t                             m_idx    = compute_idx % block_size_m * block_size_m;
+                int32_t                             gemm_idx = compute_idx / per_gemm_block_count_m;
+                int32_t                             block_idx_in_gemm = compute_idx % per_gemm_block_count_m;
+                int32_t                             m_idx    = block_idx_in_gemm * block_size_m;
                 const qnbitgemm_spacemit_ime_args & data     = qnbitgemm_args[gemm_idx];
                 int32_t rows_tobe_handled = (gemm_m - m_idx) > block_size_m ? block_size_m : (gemm_m - m_idx);
 

ggml/src/ggml-cpu/unary-ops.cpp

@@ -73,6 +73,22 @@ static inline float op_log(float x) {
     return logf(x);
 }
 
+static inline float op_floor(float x) {
+    return floorf(x);
+}
+
+static inline float op_ceil(float x) {
+    return ceilf(x);
+}
+
+static inline float op_round(float x) {
+    return roundf(x);
+}
+
+static inline float op_trunc(float x) {
+    return truncf(x);
+}
+
 template <float (*op)(float), typename src0_t, typename dst_t>
 static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
     constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
@@ -274,6 +290,22 @@ void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor *
     unary_op<op_log>(params, dst);
 }
 
+void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_floor>(params, dst);
+}
+
+void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_ceil>(params, dst);
+}
+
+void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_round>(params, dst);
+}
+
+void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_trunc>(params, dst);
+}
+
 void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
     const float alpha_n = ggml_get_op_params_f32(dst, 1);
     const float alpha_p = ggml_get_op_params_f32(dst, 2);

ggml/src/ggml-cpu/unary-ops.h

@@ -22,6 +22,10 @@ void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus

ggml/src/ggml-cpu/vec.cpp

@@ -463,9 +463,9 @@ ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const floa
 #endif
     for (; i < n; ++i) {
         float val = x[i] - mean;
+        y[i] = val;
         val *= val;
         sum += (ggml_float)val;
-        y[i] = val;
     }
     return sum/n;
 }

ggml/src/ggml-cpu/vec.h

@@ -144,14 +144,14 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         for (int i = 0; i < np; i += ggml_f16_step) {
             ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0); // 8 elements
 
-            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elemnst
+            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elements
             sum_00 = GGML_F16x_VEC_FMA(sum_00, ax1, ay1);     // sum_00 = sum_00+ax1*ay1
             ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 0*ggml_f16_epr, 0); // 8 elements
             sum_10 = GGML_F16x_VEC_FMA(sum_10, ax1, ay1);
 
             ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1); // next 8 elements
 
-            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 ekements
+            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 elements
             sum_01 = GGML_F16x_VEC_FMA(sum_01, ax2, ay2);
             ax2 = GGML_F16x_VEC_LOAD(x[1] + i + 1*ggml_f16_epr, 1);
             sum_11 = GGML_F16x_VEC_FMA(sum_11, ax2, ay2);
@@ -160,7 +160,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 
             ax3 = GGML_F16x_VEC_LOAD(x[0] + i + 2*ggml_f16_epr, 2);
             sum_02 = GGML_F16x_VEC_FMA(sum_02, ax3, ay3);
-            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
+            ax3 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
             sum_12 = GGML_F16x_VEC_FMA(sum_12, ax3, ay3);
 
             ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
@@ -820,7 +820,8 @@ inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_f
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i])));
+        const float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : expm1f(v));
     }
 }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }

ggml/src/ggml-cuda/CMakeLists.txt

@@ -44,6 +44,8 @@ if (CUDAToolkit_FOUND)
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
 
     file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-tile*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/fattn-mma*.cu")
     list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/mmq*.cu")
@@ -122,6 +124,7 @@ if (CUDAToolkit_FOUND)
 
     if (GGML_CUDA_DEBUG)
         list(APPEND CUDA_FLAGS -lineinfo)
+        add_compile_definitions(GGML_CUDA_DEBUG)
     endif()
 
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")

ggml/src/ggml-cuda/argsort.cu

@@ -1,5 +1,81 @@
 #include "argsort.cuh"
 
+#ifdef GGML_CUDA_USE_CUB
+#    include <cub/cub.cuh>
+using namespace cub;
+#endif  // GGML_CUDA_USE_CUB
+
+static __global__ void init_indices(int * indices, const int ncols, const int nrows) {
+    const int col = blockIdx.x * blockDim.x + threadIdx.x;
+    const int row = blockIdx.y;
+
+    if (col < ncols && row < nrows) {
+        indices[row * ncols + col] = col;
+    }
+}
+
+static __global__ void init_offsets(int * offsets, const int ncols, const int nrows) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx <= nrows) {
+        offsets[idx] = idx * ncols;
+    }
+}
+
+#ifdef GGML_CUDA_USE_CUB
+static void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
+                                     const float *    x,
+                                     int *            dst,
+                                     const int        ncols,
+                                     const int        nrows,
+                                     ggml_sort_order  order,
+                                     cudaStream_t     stream) {
+    ggml_cuda_pool_alloc<int>   temp_indices_alloc(pool, ncols * nrows);
+    ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
+    ggml_cuda_pool_alloc<int>   offsets_alloc(pool, nrows + 1);
+
+    int *   temp_indices = temp_indices_alloc.get();
+    float * temp_keys    = temp_keys_alloc.get();
+    int *   d_offsets    = offsets_alloc.get();
+
+    static const int block_size = 256;
+    const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
+    init_indices<<<grid_size, block_size, 0, stream>>>(temp_indices, ncols, nrows);
+
+    const dim3 offset_grid((nrows + block_size - 1) / block_size);
+    init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
+
+    cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream);
+
+    size_t temp_storage_bytes = 0;
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        DeviceSegmentedRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+                                            temp_indices, dst,                                  // values (indices)
+                                            ncols * nrows, nrows,                            // num items, num segments
+                                            d_offsets, d_offsets + 1, 0, sizeof(float) * 8,  // all bits
+                                            stream);
+    } else {
+        DeviceSegmentedRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
+                                                      dst, ncols * nrows, nrows, d_offsets, d_offsets + 1, 0,
+                                                      sizeof(float) * 8, stream);
+    }
+
+    ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
+    void *                        d_temp_storage = temp_storage_alloc.get();
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
+                                            ncols * nrows, nrows, d_offsets, d_offsets + 1, 0, sizeof(float) * 8,
+                                            stream);
+    } else {
+        DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+                                                      temp_indices, dst, ncols * nrows, nrows, d_offsets, d_offsets + 1,
+                                                      0, sizeof(float) * 8, stream);
+    }
+}
+#endif  // GGML_CUDA_USE_CUB
+
+// Bitonic sort implementation
 template<typename T>
 static inline __device__ void ggml_cuda_swap(T & a, T & b) {
     T tmp = a;
@@ -11,7 +87,7 @@ template<ggml_sort_order order>
 static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
     // bitonic sort
     int col = threadIdx.x;
-    int row = blockIdx.y;
+    int row = blockIdx.x;
 
     if (col >= ncols_pad) {
         return;
@@ -65,21 +141,28 @@ static int next_power_of_2(int x) {
     return n;
 }
 
-static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
+static void argsort_f32_i32_cuda_bitonic(const float *   x,
+                                         int *           dst,
+                                         const int       ncols,
+                                         const int       nrows,
+                                         ggml_sort_order order,
+                                         cudaStream_t    stream) {
     // bitonic sort requires ncols to be power of 2
     const int ncols_pad = next_power_of_2(ncols);
 
     const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
+    const dim3 block_nums(nrows, 1, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     // FIXME: this limit could be raised by ~2-4x on Ampere or newer
     GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
 
     if (order == GGML_SORT_ORDER_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC>
+            <<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else if (order == GGML_SORT_ORDER_DESC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC>
+            <<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else {
         GGML_ABORT("fatal error");
     }
@@ -100,5 +183,18 @@ void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
-    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
+#ifdef GGML_CUDA_USE_CUB
+    const int    ncols_pad      = next_power_of_2(ncols);
+    const size_t shared_mem     = ncols_pad * sizeof(int);
+    const size_t max_shared_mem = ggml_cuda_info().devices[ggml_cuda_get_device()].smpb;
+
+    if (shared_mem > max_shared_mem || ncols > 1024) {
+        ggml_cuda_pool & pool = ctx.pool();
+        argsort_f32_i32_cuda_cub(pool, src0_d, (int *) dst_d, ncols, nrows, order, stream);
+    } else {
+        argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
+    }
+#else
+    argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
+#endif
 }

ggml/src/ggml-cuda/binbcast.cu

@@ -272,7 +272,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         const uint3 ne12 = init_fastdiv_values((uint32_t) cne1[2]);
         const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);
 
-        if (block_nums.z > 65535) {
+        if (block_nums.z > 65535 || block_nums.y > 65535) {
             int         block_num  = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
             const uint3 prod_012    = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
             const uint3 prod_01     = init_fastdiv_values((uint32_t) (ne0 * ne1));

ggml/src/ggml-cuda/common.cuh

@@ -224,6 +224,11 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
+// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#define VOLTA_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define TURING_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -245,7 +250,8 @@ static bool fp16_available(const int cc) {
 }
 
 static bool fast_fp16_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
+    return GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -277,7 +283,10 @@ static bool amd_mfma_available(const int cc) {
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
-// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
+static bool volta_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
+}
+
 static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }
@@ -571,6 +580,10 @@ static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v,
 }
 
 // Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
+// Important: do not use this function if dst and src both point at registers.
+//     Due to the strict aliasing rule the compiler can do incorrect optimizations if src and dst have different types.
+//     The function is intended for copies between registers and SRAM/VRAM to make the compiler emit the right instructions.
+//     If dst and src point at different address spaces then they are guaranteed to not be aliased.
 template <int nbytes, int alignment = 0>
 static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
     if constexpr (alignment != 0) {
@@ -620,8 +633,11 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 // and a shift:
 //
 // n/d = (mulhi(n, mp) + n) >> L;
-static const uint3 init_fastdiv_values(uint32_t d) {
-    GGML_ASSERT(d != 0);
+static const uint3 init_fastdiv_values(uint64_t d_64) {
+    GGML_ASSERT(d_64 != 0);
+    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());
+
+    uint32_t d = (uint32_t)d_64;
 
     // compute L = ceil(log2(d));
     uint32_t L = 0;
@@ -939,13 +955,6 @@ struct ggml_cuda_graph {
     bool disable_due_to_failed_graph_capture = false;
     int number_consecutive_updates = 0;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
-    bool use_cpy_indirection = false;
-    std::vector<char *> cpy_dest_ptrs;
-    char ** dest_ptrs_d;
-    int dest_ptrs_size = 0;
-    // Index to allow each cpy kernel to be aware of it's position within the graph
-    // relative to other cpy nodes.
-    int graph_cpynode_index = -1;
 #endif
 };
 
@@ -1007,3 +1016,16 @@ struct ggml_backend_cuda_context {
         return pool(device);
     }
 };
+
+struct ggml_cuda_mm_fusion_args_host {
+    const ggml_tensor * x_bias = nullptr;
+    const ggml_tensor * gate = nullptr;
+    const ggml_tensor * gate_bias = nullptr;
+    ggml_glu_op glu_op;
+};
+struct ggml_cuda_mm_fusion_args_device {
+    const void * x_bias = nullptr;
+    const void * gate = nullptr;
+    const void * gate_bias = nullptr;
+    ggml_glu_op glu_op;
+};

ggml/src/ggml-cuda/convert.cuh

@@ -1,3 +1,4 @@
+#pragma once
 #include "common.cuh"
 
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256

ggml/src/ggml-cuda/cpy.cu

@@ -7,19 +7,21 @@
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
+const int CUDA_CPY_TILE_DIM_2D = 32; // 2D tile dimension for transposed blocks
+const int CUDA_CPY_BLOCK_NM = 8;     // block size of 3rd dimension if available
+const int CUDA_CPY_BLOCK_ROWS = 8;   // block dimension for marching through rows
+
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
                                const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                               const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+                               const int nb12, const int nb13) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
         return;
     }
 
-    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
-
     // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
     // then combine those indices with the corresponding byte offsets to get the total offsets
     const int64_t i03 = i/(ne00 * ne01 * ne02);
@@ -37,6 +39,55 @@ static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
+template <typename T>
+static __global__ void cpy_flt_transpose(const char * cx, char * cdst, const int ne,
+                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                               const int nb12, const int nb13) {
+
+    const T* src = reinterpret_cast<const T*>(cx);
+    T* dst = reinterpret_cast<T*>(cdst);
+
+    const int64_t nmat = ne / (ne00 * ne01);
+    const int64_t n = ne00 * ne01;
+
+    const int x = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
+    const int y = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
+    const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+
+    __shared__ float tile[CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
+
+#pragma unroll
+    for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
+
+        const unsigned int imat = blockIdx.z * CUDA_CPY_BLOCK_NM + i;
+        if (imat >= nmat)
+            break;
+
+#pragma unroll
+        for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
+            if(x < ne01 && y + j < ne00){
+                const int row = threadIdx.y+j;
+                const int col = threadIdx.x * sizeof(float)/sizeof(T);
+                T *tile2 = reinterpret_cast<T*>(tile[row]);
+                tile2[col] = src[imat*n + (y+j)*ne01 + x];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
+            if (ty + j < ne01 && tx < ne00) {
+                const int col = (threadIdx.y+j)*sizeof(float)/sizeof(T);
+                const T *tile2 = reinterpret_cast<const T*>(tile[threadIdx.x]);
+                dst[imat*n + (ty+j)*ne00 + tx] = tile2[col];
+            }
+        }
+    }
+}
+
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
@@ -63,18 +114,16 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+                                 const int nb12, const int nb13) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
-    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
-
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -91,18 +140,16 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne,
+static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+                                 const int nb12, const int nb13) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
-    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
-
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -118,67 +165,92 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
-// Copy destination pointers to GPU to be available when pointer indirection is in use
+template<typename src_t, typename dst_t>
+static __global__ void cpy_flt_contiguous(const char * cx, char * cdst, const int64_t ne) {
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
-void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) {
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
-    if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers
-        CUDA_CHECK(cudaStreamSynchronize(stream));
-        if (cuda_graph->dest_ptrs_d != nullptr) {
-            CUDA_CHECK(cudaFree(cuda_graph->dest_ptrs_d));
-        }
-        CUDA_CHECK(cudaMalloc(&cuda_graph->dest_ptrs_d, host_dest_ptrs_size*sizeof(char *)));
-        cuda_graph->dest_ptrs_size = host_dest_ptrs_size;
+    if (i >= ne) {
+        return;
     }
-    // copy destination pointers to GPU
-    CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
-    cuda_graph->graph_cpynode_index = 0; // reset index
-#else
-    GGML_UNUSED_VARS(cuda_graph, host_dest_ptrs, host_dest_ptrs_size, stream);
-#endif
+
+    const src_t * x = (const src_t *) cx;
+    dst_t *     dst = (dst_t *) cdst;
+
+    dst[i] = ggml_cuda_cast<dst_t>(x[i]);
 }
 
 template<typename src_t, typename dst_t>
+static void ggml_cpy_flt_contiguous_cuda(
+    const char * cx, char * cdst, const int64_t ne,
+cudaStream_t stream) {
+
+    const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_flt_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne);
+}
+
+template<typename src_t, typename dst_t, bool transposed = false>
 static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+    if (transposed) {
+        GGML_ASSERT(ne == ne00*ne01*ne02);  // ne[3] is 1 assumed
+        int ne00n, ne01n, ne02n;
+        if (nb00 <= nb02) { // most likely safe to handle nb00 = nb02 case here
+            ne00n = ne00;
+            ne01n = ne01;
+            ne02n = ne02;
+        } else if (nb00 > nb02) {
+            ne00n = ne00;
+            ne01n = ne01*ne02;
+            ne02n = 1;
+        }
+
+        dim3 dimGrid( (ne01n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,
+                      (ne00n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,
+                      (ne/(ne01n*ne00n) + CUDA_CPY_BLOCK_NM - 1) / CUDA_CPY_BLOCK_NM);
+        dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
+        cpy_flt_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
+            (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    } else {
+        const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+        cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+            (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    }
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
     cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q8_0_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q4_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q4_0_f32_cuda(
@@ -187,22 +259,22 @@ static void ggml_cpy_q4_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q4_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q4_1_f32_cuda(
@@ -211,22 +283,22 @@ static void ggml_cpy_q4_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q5_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q5_0_f32_cuda(
@@ -235,22 +307,22 @@ static void ggml_cpy_q5_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_q5_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
     cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_q5_1_f32_cuda(
@@ -259,25 +331,25 @@ static void ggml_cpy_q5_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    cudaStream_t stream) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
 static void ggml_cpy_f32_iq4_nl_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
     cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -311,17 +383,10 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
 
-    char ** dest_ptrs_d = nullptr;
-    int graph_cpynode_index = -1;
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
-        dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
-        graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
-    }
-#else
-    GGML_UNUSED(disable_indirection_for_this_node);
-#endif
-    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+    const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) && src0->ne[3] == 1;
+
+    if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
 #if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
         if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
@@ -329,134 +394,106 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         } else
 #endif // GGML_USE_MUSA && GGML_MUSA_MUDNN_COPY
         {
-            if (src0->type == GGML_TYPE_F32) {
-                ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
-            } else {
-                CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
-            }
+            CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (can_be_transposed) {
+            ggml_cpy_flt_cuda<float, float, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<float, half>        (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, half>        (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (can_be_transposed) {
+            ggml_cpy_flt_cuda<half, half, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<half, half>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<half, nv_bfloat16>  (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<half, nv_bfloat16>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<half, float>        (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<half, float>          (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (can_be_transposed) {
+            ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, half>  (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<nv_bfloat16, half>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<nv_bfloat16, float>   (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        ggml_cpy_flt_cuda<float, int32_t> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<float, int32_t>     (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, int32_t>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<int32_t, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        if (contiguous_srcs) {
+            ggml_cpy_flt_contiguous_cuda<int32_t, float>     (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<int32_t, float>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
-        ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
-    }
-#else
-    GGML_UNUSED(disable_indirection_for_this_node);
-#endif
-
 }
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
-    bool disable_indirection = true;
-    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
-}
-
-void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
-    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-        // Prioritize CUDA graph compatibility over direct memory copy optimization.
-        // Using copy kernels here maintains graph indirection support, preventing performance regression from disabled CUDA graphs.
-        if (src0->type == GGML_TYPE_F32) {
-            return (void*) cpy_flt<cpy_1_flt<float, float>>;
-        } else {
-            return nullptr;
-        }
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<float, float>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_flt<cpy_1_flt<float, nv_bfloat16>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_flt<cpy_1_flt<float, half>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
-    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>;
-    } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>;
-    } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>;
-    } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        return (void*) cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        return (void*) cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>;
-    } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>;
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_flt<cpy_1_flt<half, half>>;
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_flt<cpy_1_flt<half, nv_bfloat16>>;
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<half, float>>;
-    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, half>>;
-    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
-    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        return (void*) cpy_flt<cpy_1_flt<float, int32_t>>;
-    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_flt<cpy_1_flt<int32_t, float>>;
-    } else {
-        GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
-                ggml_type_name(src0->type), ggml_type_name(src1->type));
-    }
+    ggml_cuda_cpy(ctx, src0, dst);
 }

ggml/src/ggml-cuda/cpy.cuh

@@ -2,10 +2,6 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-
-void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);
-
-void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream);

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

@@ -793,8 +793,6 @@ void launch_fattn(
     GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
         "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
 
-    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
-
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
     const int id  = ggml_cuda_get_device();
@@ -878,7 +876,7 @@ void launch_fattn(
     // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
     // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
     //     multiple sequences of possibly different lengths.
-    if (mask && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+    if (mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
         const int s31 = mask->nb[1] / sizeof(half2);
         const int s33 = mask->nb[3] / sizeof(half2);
 
@@ -897,6 +895,7 @@ void launch_fattn(
     const dim3 block_dim(warp_size, nwarps, 1);
     int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
     CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+    GGML_ASSERT(max_blocks_per_sm > 0);
     int parallel_blocks = max_blocks_per_sm;
 
     dim3 blocks_num;
@@ -916,8 +915,7 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
-        GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
-        const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+        const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 1) / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
 
         // parallel_blocks must not be larger than what the tensor size allows:
         parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
@@ -946,7 +944,7 @@ void launch_fattn(
 
         blocks_num.x = ntiles_x;
         blocks_num.y = parallel_blocks;
-        blocks_num.z = Q->ne[2]*Q->ne[3];
+        blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));

ggml/src/ggml-cuda/fattn-tile.cu

@@ -1,756 +1,49 @@
 #include "common.cuh"
-#include "fattn-common.cuh"
 #include "fattn-tile.cuh"
 #include "fattn-wmma-f16.cuh"
 
-// kq_stride == number of KQ rows to process per iteration
-// kq_nbatch == number of K columns to load in parallel for KQ calculation
-
-static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int cc, const int warp_size) {
-    if (GGML_CUDA_CC_IS_AMD(cc)) {
-        if (GGML_CUDA_CC_IS_RDNA(cc)) {
-            switch (D) {
-                case 64:
-                    return 128;
-                case 128:
-                case 256:
-                    return ncols <= 16 ? 128 : 64;
-                default:
-                    GGML_ABORT("fatal error");
-                    return -1;
-            }
-        }
-        switch (D) {
-            case 64:
-                return ncols == 32 ? 128 : 64;
-            case 128:
-                return ncols == 32 ? 64 : 32;
-            case 256:
-                return 32;
-            default:
-                GGML_ABORT("fatal error");
-                return -1;
-        }
-    }
-    if (fast_fp16_available(cc)) {
-        switch (D) {
-            case 64:
-            case 128:
-            case 256:
-                return ncols <= 16 ? 128 : 64;
-            default:
-                GGML_ABORT("fatal error");
-                return -1;
-        }
-    }
-    switch (D) {
-        case 64:
-            return ncols <= 16 ? 128 : 64;
-        case 128:
-            return ncols <= 16 ? 64 : 32;
-        case 256:
-            return 32;
-        default:
-            GGML_ABORT("fatal error");
-            return -1;
-    }
-    GGML_UNUSED(warp_size);
-}
-
-static constexpr __device__ int fattn_tile_get_kq_stride_device(int D, int ncols, int warp_size) {
-#ifdef GGML_USE_HIP
-#ifdef RDNA
-    switch (D) {
-        case 64:
-            return 128;
-        case 128:
-        case 256:
-            return ncols <= 16 ? 128 : 64;
-        default:
-            return -1;
-    }
-#else
-    switch (D) {
-        case 64:
-            return ncols == 32 ? 128 : 64;
-        case 128:
-            return ncols == 32 ? 64 : 32;
-        case 256:
-            return 32;
-        default:
-            return -1;
-    }
-#endif // RDNA
-#else
-#ifdef FAST_FP16_AVAILABLE
-    switch (D) {
-        case 64:
-        case 128:
-        case 256:
-            return ncols <= 16 ? 128 : 64;
-        default:
-            return -1;
-    }
-#else
-    switch (D) {
-        case 64:
-            return ncols <= 16 ? 128 : 64;
-        case 128:
-            return ncols <= 16 ? 64 : 32;
-        case 256:
-            return 32;
-        default:
-            return -1;
-    }
-#endif // FAST_FP16_AVAILABLE
-#endif // GGML_USE_HIP
-    GGML_UNUSED_VARS(ncols, warp_size);
-}
-
-static constexpr __device__ int fattn_tile_get_kq_nbatch_device(int D, int ncols, int warp_size) {
-#ifdef GGML_USE_HIP
-    switch (D) {
-        case 64:
-            return 64;
-        case 128:
-        case 256:
-            return 128;
-        default:
-            return -1;
-    }
-#else
-#ifdef FAST_FP16_AVAILABLE
-    switch (D) {
-        case 64:
-            return 64;
-        case 128:
-        case 256:
-            return 128;
-        default:
-            return -1;
-    }
-#else
-    switch (D) {
-        case 64:
-            return 64;
-        case 128:
-            return 128;
-        case 256:
-            return ncols <= 16 ? 128 : 64;
-        default:
-            return -1;
-    }
-#endif // FAST_FP16_AVAILABLE
-#endif // GGML_USE_HIP
-    GGML_UNUSED_VARS(ncols, warp_size);
-}
-
-static int fattn_tile_get_nthreads_host(const int cc, const int ncols) {
-    return 256;
-    GGML_UNUSED_VARS(cc, ncols);
-}
-
-static constexpr __device__ int fattn_tile_get_nthreads_device(int ncols) {
-    return 256;
-    GGML_UNUSED(ncols);
-}
-
-static constexpr __device__ int fattn_tile_get_occupancy_device(int ncols) {
-#ifdef RDNA
-    return 3;
-#else
-    return ncols <= 16 ? 3 : 2;
-#endif // RDNA
-    GGML_UNUSED(ncols);
-}
-
-template<int D, int ncols, bool use_logit_softcap> // D == head size
-__launch_bounds__(fattn_tile_get_nthreads_device(ncols), fattn_tile_get_occupancy_device(ncols))
-static __global__ void flash_attn_tile(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        const char * __restrict__ sinks,
-        const int  * __restrict__ KV_max,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
-                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
-        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
-                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
-                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
-                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
-                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#ifdef FLASH_ATTN_AVAILABLE
-
-    // Skip unused kernel variants for faster compilation:
-#ifdef GGML_USE_WMMA_FATTN
-    NO_DEVICE_CODE;
-    return;
-#endif // GGML_USE_WMMA_FATTN
-
-    if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-            max_bias, m0, m1, n_head_log2, logit_softcap,
-            ne00, ne01, ne02, ne03,
-                  nb01, nb02, nb03,
-            ne10, ne11, ne12, ne13,
-                  nb11, nb12, nb13,
-                  nb21, nb22, nb23,
-                  ne31, ne32, ne33,
-                  nb31, nb32, nb33);
-        NO_DEVICE_CODE;
-        return;
-    }
-
-    constexpr int warp_size = 32;
-    constexpr int nwarps    = fattn_tile_get_nthreads_device(ncols) / warp_size;
-    constexpr int kq_stride = fattn_tile_get_kq_stride_device(D, ncols, warp_size);
-    static_assert(kq_stride % warp_size == 0, "kq_stride not divisable by warp_size.");
-    constexpr int kq_nbatch = fattn_tile_get_kq_nbatch_device(D, ncols, warp_size);
-    static_assert(kq_nbatch % (2*warp_size) == 0, "bad kq_nbatch");
-
-    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
-
-    const int sequence = blockIdx.z / ne02;
-    const int head = blockIdx.z - sequence*ne02;
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f    = (const float *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half2 * K_h2   = (const half2 *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half2 * V_h2   = (const half2 *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
-    const float * sinksf = (const float *) (sinks);
-
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
-
-    constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
-    constexpr int cpy_ne = cpy_nb / 4;
-
-    constexpr int cpw = ncols/nwarps; // cols per warp
-
-    // softmax_iter_j == number of KQ columns for which to calculate softmax in parallel.
-    // KQ is originall 2D but uses a Z-shaped memory pattern for larger reads/writes.
-#ifdef FAST_FP16_AVAILABLE
-    constexpr int softmax_iter_j = cpw < 2*cpy_ne ? cpw : 2*cpy_ne;
-
-    __shared__ half  KQ[ncols/softmax_iter_j][kq_stride][softmax_iter_j];
-    __shared__ half2 Q_tmp[ncols][D/2];
-    __shared__ half2 KV_tmp[kq_stride * (kq_nbatch/2 + cpy_ne)]; // Padded to avoid memory bank conflicts.
-    half2 VKQ[cpw][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
-#else
-    constexpr int softmax_iter_j = cpw < 1*cpy_ne ? cpw : 1*cpy_ne;
-
-    __shared__ float KQ[ncols/softmax_iter_j][kq_stride][softmax_iter_j];
-    __shared__ float Q_tmp[ncols][D];
-    __shared__ float KV_tmp[kq_stride * (kq_nbatch + cpy_ne)]; // Padded to avoid memory bank conflicts.
-    float2 VKQ[cpw][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
-#endif // FAST_FP16_AVAILABLE
-    static_assert(cpw % softmax_iter_j == 0, "bad softmax_iter_j");
-
-    float KQ_max[cpw];
-#pragma unroll
-    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-        KQ_max[j0/nwarps] = -FLT_MAX/2.0f;
-    }
-    float KQ_sum[cpw] = {0.0f};
-
-    // Load Q data, convert to FP16 if fast.
-#pragma unroll
-    for (int j0 = 0; j0 < cpw; ++j0) {
-        const int j = j0 + threadIdx.y*cpw;
-
-        constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-
-#pragma unroll
-        for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-            float tmp_f[cpy_ne_D] = {0.0f};
-            if (ic0 + j < ne01) {
-                ggml_cuda_memcpy_1<sizeof(tmp_f)>(tmp_f, &Q_f[j*(nb01/sizeof(float)) + i0 + threadIdx.x*cpy_ne_D]);
-            }
-
-#pragma unroll
-            for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
-                tmp_f[i1] *= scale;
-            }
-
-#ifdef FAST_FP16_AVAILABLE
-            half2 tmp_h2[cpy_ne_D/2];
-#pragma unroll
-            for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
-                tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
-            }
-            ggml_cuda_memcpy_1<sizeof(tmp_h2)>(&Q_tmp[j][i0/2 + threadIdx.x*(cpy_ne_D/2)], tmp_h2);
-#else
-            ggml_cuda_memcpy_1<sizeof(tmp_f)> (&Q_tmp[j][i0   + threadIdx.x* cpy_ne_D],    tmp_f);
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
-    __syncthreads();
-
-    // Main loop over KV cache:
-    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
-    for (int k_VKQ_0 = blockIdx.y*kq_stride; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*kq_stride) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        float KQ_max_new[cpw];
-#pragma unroll
-        for (int j = 0; j < cpw; ++j) {
-            KQ_max_new[j] = KQ_max[j];
-        }
-
-        float KQ_acc[kq_stride/warp_size][cpw] = {{0.0f}}; // Accumulators for KQ matrix multiplication.
-
-        // KQ = K @ Q matrix multiplication:
-#pragma unroll
-        for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += kq_nbatch) {
-#pragma unroll
-            for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += nwarps) {
-                const int i_KQ = i_KQ_0 + threadIdx.y;
-
-#ifdef FAST_FP16_AVAILABLE
-                constexpr int cpy_ne_kqnb = cpy_ne < kq_nbatch/(2*warp_size) ? cpy_ne : kq_nbatch/(2*warp_size);
-#pragma unroll
-                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += warp_size*cpy_ne_kqnb) {
-                    ggml_cuda_memcpy_1<cpy_ne_kqnb*4>(
-                        &KV_tmp[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1 + threadIdx.x*cpy_ne_kqnb],
-                        &K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1 + threadIdx.x*cpy_ne_kqnb]);
-                }
-#else
-                constexpr int cpy_ne_kqnb = cpy_ne < kq_nbatch/warp_size ? cpy_ne : kq_nbatch/warp_size;
-#pragma unroll
-                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += warp_size*cpy_ne_kqnb) {
-                    half2 tmp_h2[cpy_ne_kqnb/2];
-                    ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
-                        tmp_h2, &K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1/2 + threadIdx.x*(cpy_ne_kqnb/2)]);
-
-                    float2 tmp_f2[cpy_ne_kqnb/2];
-#pragma unroll
-                    for (int k_KQ_2 = 0; k_KQ_2 < cpy_ne_kqnb/2; ++k_KQ_2) {
-                        tmp_f2[k_KQ_2] = __half22float2(tmp_h2[k_KQ_2]);
-                    }
-                    ggml_cuda_memcpy_1<sizeof(tmp_f2)>(
-                        &KV_tmp[i_KQ*(kq_nbatch + cpy_ne) + k_KQ_1 + threadIdx.x*cpy_ne_kqnb], tmp_f2);
-                }
-#endif // FAST_FP16_AVAILABLE
-            }
-
-            __syncthreads();
-
-#ifdef FAST_FP16_AVAILABLE
-#pragma unroll
-            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += cpy_ne) {
-                half2 K_k[kq_stride/warp_size][cpy_ne];
-                half2 Q_k[cpw][cpy_ne];
-#else
-#pragma unroll
-            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += cpy_ne) {
-                float K_k[kq_stride/warp_size][cpy_ne];
-                float Q_k[cpw][cpy_ne];
-#endif // FAST_FP16_AVAILABLE
-
-#pragma unroll
-                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
-                    const int i_KQ = i_KQ_0 + threadIdx.x;
-
-#ifdef FAST_FP16_AVAILABLE
-                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1]);
-#else
-                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp[i_KQ*(kq_nbatch   + cpy_ne) + k_KQ_1]);
-#endif // FAST_FP16_AVAILABLE
-                }
-#pragma unroll
-                for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
-                    const int j_KQ = j_KQ_0 + threadIdx.y*cpw;
-
-#ifdef FAST_FP16_AVAILABLE
-                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0], &Q_tmp[j_KQ][k_KQ_0/2 + k_KQ_1]);
-#else
-                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0], &Q_tmp[j_KQ][k_KQ_0   + k_KQ_1]);
-#endif // FAST_FP16_AVAILABLE
-                }
-
-#pragma unroll
-                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
-#pragma unroll
-                    for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
-#pragma unroll
-                        for (int k = 0; k < cpy_ne; ++k) {
-                            ggml_cuda_mad(KQ_acc[i_KQ_0/warp_size][j_KQ_0], K_k[i_KQ_0/warp_size][k], Q_k[j_KQ_0][k]);
-                        }
-                    }
-                }
-            }
-
-            if (k_KQ_0 + kq_nbatch < D) {
-                __syncthreads(); // Sync not needed on last iteration.
-            }
-        }
-
-        // Apply logit softcap, mask, update KQ_max:
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
-            const int i_KQ = i_KQ_0 + threadIdx.x;
-
-#pragma unroll
-            for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
-                const int j_KQ = j_KQ_0 + threadIdx.y*cpw;
-
-                if (use_logit_softcap) {
-                    KQ_acc[i_KQ_0/warp_size][j_KQ_0] = logit_softcap * tanhf(KQ_acc[i_KQ_0/warp_size][j_KQ_0]);
-                }
-
-                KQ_acc[i_KQ_0/warp_size][j_KQ_0] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
-
-                KQ_max_new[j_KQ_0] = fmaxf(KQ_max_new[j_KQ_0], KQ_acc[i_KQ_0/warp_size][j_KQ_0]);
-            }
-        }
-
-        __syncthreads();
-
-        // Calculate KQ softmax, write to shared KQ buffer, re-scale VKQ accumulators:
-#pragma unroll
-        for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
-#ifdef FAST_FP16_AVAILABLE
-            half  tmp[kq_stride/warp_size][softmax_iter_j];
-#else
-            float tmp[kq_stride/warp_size][softmax_iter_j];
-#endif // FAST_FP16_AVAILABLE
-
-#pragma unroll
-            for (int j1 = 0; j1 < softmax_iter_j; ++j1) {
-                KQ_max_new[j0+j1] = warp_reduce_max<warp_size>(KQ_max_new[j0+j1]);
-                const float KQ_max_scale = expf(KQ_max[j0+j1] - KQ_max_new[j0+j1]);
-                KQ_max[j0+j1] = KQ_max_new[j0+j1];
-
-                float KQ_sum_add = 0.0f;
-#pragma unroll
-                for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
-                    const float val = expf(KQ_acc[i0/warp_size][j0+j1] - KQ_max[j0+j1]);
-                    KQ_sum_add += val;
-                    tmp[i0/warp_size][j1] = val;
-                }
-                KQ_sum[j0+j1] = KQ_sum[j0+j1]*KQ_max_scale + KQ_sum_add;
-
-#ifdef FAST_FP16_AVAILABLE
-                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                    VKQ[j0+j1][i0/warp_size] *= KQ_max_scale_h2;
-                }
-#else
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                    VKQ[j0+j1][i0/warp_size].x *= KQ_max_scale;
-                    VKQ[j0+j1][i0/warp_size].y *= KQ_max_scale;
-                }
-#endif // FAST_FP16_AVAILABLE
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
-                const int i = i0 + threadIdx.x;
-
-                ggml_cuda_memcpy_1<sizeof(tmp[0])>(
-                    KQ[j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j)][i], tmp[i0/warp_size]);
-            }
-        }
-
-        // VKQ = V @ KQ matrix multiplication:
-        constexpr int V_cols_per_iter = kq_stride*kq_nbatch / D; // Number of V columns that fit in SRAM for K.
-        static_assert(kq_stride % V_cols_per_iter == 0, "bad V_cols_per_iter");
-#pragma unroll
-        for (int k0 = 0; k0 < kq_stride; k0 += V_cols_per_iter) {
-#pragma unroll
-            for (int k1 = 0; k1 < V_cols_per_iter; k1 += nwarps) {
-                const int k_tile = k1 + threadIdx.y;
-
-#ifdef FAST_FP16_AVAILABLE
-                constexpr int cpy_ne_D = cpy_ne < D/(2*warp_size) ? cpy_ne : D/(2*warp_size);
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
-                    ggml_cuda_memcpy_1<cpy_ne_D*4>(
-                        &KV_tmp[k_tile*(D/2) + i0 + threadIdx.x*cpy_ne_D],
-                        &V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i0 + threadIdx.x*cpy_ne_D]);
-                }
-#else
-                constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-#pragma unroll
-                for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-                    half2 tmp_h2[cpy_ne_D/2];
-                    ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
-                        tmp_h2, &V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i0/2 + threadIdx.x*(cpy_ne_D/2)]);
-
-                    float2 tmp_f2[cpy_ne_D/2];
-#pragma unroll
-                    for (int i1 = 0; i1 < cpy_ne_D/2; ++i1) {
-                        tmp_f2[i1] = __half22float2(tmp_h2[i1]);
-                    }
-                    ggml_cuda_memcpy_1<sizeof(tmp_f2)>(
-                        &KV_tmp[k_tile*D + i0 + threadIdx.x*cpy_ne_D], tmp_f2);
-                }
-#endif // FAST_FP16_AVAILABLE
-            }
-
-            __syncthreads();
-
-#ifdef FAST_FP16_AVAILABLE
-#pragma unroll
-            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
-                half2 V_k[(D/2)/warp_size];
-                half2 KQ_k[cpw];
-
-                constexpr int cpy_ne_D = cpy_ne/2 < (D/2)/warp_size ? cpy_ne/2 : (D/2)/warp_size;
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
-                    ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/warp_size], &KV_tmp[k1*(D/2) + i0 + threadIdx.x*cpy_ne_D]);
-                }
-#pragma unroll
-                for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
-                    const int j = j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j);
-
-                    half tmp[softmax_iter_j];
-                    ggml_cuda_memcpy_1<softmax_iter_j*sizeof(half)>(
-                        &tmp, KQ[j][k0 + k1]);
-#pragma unroll
-                    for (int j1 = 0; j1 < softmax_iter_j; ++j1) {
-                        KQ_k[j0+j1] = __half2half2(tmp[j1]);
-                    }
-                }
-
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-#pragma unroll
-                    for (int j0 = 0; j0 < cpw; ++j0) {
-                        VKQ[j0][i0/warp_size] += V_k[i0/warp_size]*KQ_k[j0];
-                    }
-                }
-            }
-#else
-#pragma unroll
-            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
-                float2 V_k[(D/2)/warp_size];
-                float  KQ_k[cpw];
-
-                constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-#pragma unroll
-                for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-                    ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/(2*warp_size)], &KV_tmp[k1*D + i0 + threadIdx.x*cpy_ne_D]);
-                }
-#pragma unroll
-                for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
-                    const int j = j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j);
-
-                    ggml_cuda_memcpy_1<softmax_iter_j*sizeof(float)>(
-                        &KQ_k[j0], KQ[j][k0 + k1]);
-                }
-
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-#pragma unroll
-                    for (int j0 = 0; j0 < cpw; ++j0) {
-                        VKQ[j0][i0/warp_size].x += V_k[i0/warp_size].x*KQ_k[j0];
-                        VKQ[j0][i0/warp_size].y += V_k[i0/warp_size].y*KQ_k[j0];
-                    }
-                }
-            }
-#endif // FAST_FP16_AVAILABLE
-
-            __syncthreads();
-        }
-    }
-
-
-    // Attention sink: adjust running max and sum once per head
-    if (sinksf && blockIdx.y == 0) {
-        const float sink = sinksf[head];
-
-#pragma unroll
-        for (int j0 = 0; j0 < cpw; ++j0) {
-            float KQ_max_new_j = fmaxf(KQ_max[j0], sink);
-            KQ_max_new_j = warp_reduce_max<warp_size>(KQ_max_new_j);
-
-            const float KQ_max_scale = expf(KQ_max[j0] - KQ_max_new_j);
-            KQ_max[j0] = KQ_max_new_j;
-
-            const float val = expf(sink - KQ_max[j0]);
-            KQ_sum[j0] = KQ_sum[j0] * KQ_max_scale;
-            if (threadIdx.x == 0) {
-                KQ_sum[j0] += val;
-            }
-
-#ifdef FAST_FP16_AVAILABLE
-            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                VKQ[j0][i0/warp_size] *= KQ_max_scale_h2;
-            }
-#else
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                VKQ[j0][i0/warp_size].x *= KQ_max_scale;
-                VKQ[j0][i0/warp_size].y *= KQ_max_scale;
-            }
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
-#pragma unroll
-    for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
-        KQ_sum[j_VKQ_0] = warp_reduce_sum<warp_size>(KQ_sum[j_VKQ_0]);
-    }
-    if (gridDim.y == 1) {
-#pragma unroll
-        for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
-#ifdef FAST_FP16_AVAILABLE
-            const half2 KQ_sum_j_inv = make_half2(1.0f/KQ_sum[j_VKQ_0], 1.0f/KQ_sum[j_VKQ_0]);
-#pragma unroll
-            for (int i = 0; i < (D/2)/warp_size; ++i) {
-                VKQ[j_VKQ_0][i] *= KQ_sum_j_inv;
-            }
-#else
-            const float KQ_sum_j_inv = 1.0f/KQ_sum[j_VKQ_0];
-#pragma unroll
-            for (int i = 0; i < (D/2)/warp_size; ++i) {
-                VKQ[j_VKQ_0][i].x *= KQ_sum_j_inv;
-                VKQ[j_VKQ_0][i].y *= KQ_sum_j_inv;
-            }
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
-    // Write back results:
-#pragma unroll
-    for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
-        const int j_VKQ = j_VKQ_0 + threadIdx.y*cpw;
-
-        if (ic0 + j_VKQ >= ne01) {
-            return;
-        }
-
-        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
-
-#ifdef FAST_FP16_AVAILABLE
-        constexpr int cpy_ne_D = cpy_ne/2 < (D/2)/warp_size ? cpy_ne/2 : (D/2)/warp_size;
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
-            float2 tmp[cpy_ne_D];
-#pragma unroll
-            for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
-                tmp[i1] = __half22float2(VKQ[j_VKQ_0][i0/warp_size + i1]);
-            }
-            ggml_cuda_memcpy_1<sizeof(tmp)>(&dst[j_dst_unrolled*D + 2*i0 + threadIdx.x*(2*cpy_ne_D)], tmp);
-        }
-#else
-        constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-#pragma unroll
-        for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-            ggml_cuda_memcpy_1<cpy_ne_D*4>(
-                &dst[j_dst_unrolled*D + i0 + threadIdx.x*cpy_ne_D], &VKQ[j_VKQ_0][i0/(2*warp_size)]);
-        }
-#endif // FAST_FP16_AVAILABLE
-
-        if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[j_dst_unrolled] = make_float2(KQ_max[j_VKQ_0], KQ_sum[j_VKQ_0]);
-        }
-    }
-#else
-    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-        max_bias, m0, m1, n_head_log2, logit_softcap,
-        ne00, ne01, ne02, ne03,
-              nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-              nb11, nb12, nb13,
-              nb21, nb22, nb23,
-              ne31, ne32, ne33,
-              nb31, nb32, nb33);
-    NO_DEVICE_CODE;
-#endif // FLASH_ATTN_AVAILABLE
-}
-
-template <int D, bool use_logit_softcap>
-static void launch_fattn_tile_switch_ncols(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-
-    const int id        = ggml_cuda_get_device();
-    const int cc        = ggml_cuda_info().devices[id].cc;
-    const int warp_size = 32;
-
-    constexpr size_t nbytes_shared = 0;
-
-#ifdef GGML_USE_HIP
-    if constexpr (D <= 128) {
-        if (Q->ne[1] > 32) {
-            constexpr int cols_per_block = 64;
-            const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
-            fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
-            const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
-            launch_fattn<D, cols_per_block, 1>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
-            return;
-        }
-    }
-#endif // GGML_USE_HIP
-
-    if (Q->ne[1] > 16) {
-        constexpr int cols_per_block = 32;
-        const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
-        fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
-        const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
-        launch_fattn<D, cols_per_block, 1>
-            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
-        return;
-    }
-
-    constexpr int cols_per_block = 16;
-    const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
-    fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
-    const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
-    launch_fattn<D, cols_per_block, 1>
-        (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
-}
-
-template <bool use_logit_softcap>
-static void launch_fattn_tile_switch_head_size(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    switch (Q->ne[0]) {
+void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+    switch (K->ne[0]) {
+        case  40: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 40,  40>(ctx, dst);
+        } break;
         case  64: {
-            launch_fattn_tile_switch_ncols< 64, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 64,  64>(ctx, dst);
+        } break;
+        case  72: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 72,  72>(ctx, dst);
+        } break;
+        case  80: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 80,  80>(ctx, dst);
+        } break;
+        case  96: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 96,  96>(ctx, dst);
+        } break;
+        case 112: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<112, 112>(ctx, dst);
         } break;
         case 128: {
-            launch_fattn_tile_switch_ncols<128, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<128, 128>(ctx, dst);
         } break;
         case 256: {
-            launch_fattn_tile_switch_ncols<256, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<256, 256>(ctx, dst);
+        } break;
+        case 576: {
+            GGML_ASSERT(V->ne[0] == 512);
+            ggml_cuda_flash_attn_ext_tile_case<576, 512>(ctx, dst);
         } break;
         default: {
             GGML_ABORT("Unsupported head size");
         } break;
     }
 }
-
-void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-
-    float logit_softcap;
-    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
-
-    if (logit_softcap == 0.0f) {
-        constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
-    } else {
-        constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
-    }
-}

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

@@ -516,8 +516,8 @@ void ggml_cuda_flash_attn_ext_vec_case_impl(ggml_backend_cuda_context & ctx, ggm
     const int nthreads = ggml_cuda_fattn_vec_get_nthreads_host(cc);
     const int nwarps   = nthreads / WARP_SIZE;
     fattn_kernel_t fattn_kernel = flash_attn_ext_vec<D, cols_per_block, type_K, type_V, use_logit_softcap>;
-    constexpr bool need_f16_K = false;
-    constexpr bool need_f16_V = false;
+    const bool need_f16_K = type_K == GGML_TYPE_F16;
+    const bool need_f16_V = type_V == GGML_TYPE_F16;
     constexpr size_t nbytes_shared = 0;
     launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
@@ -526,11 +526,6 @@ template <int D, ggml_type type_K, ggml_type type_V>
 void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * KQV = dst;
     const ggml_tensor * Q   = dst->src[0];
-    const ggml_tensor * K   = dst->src[1];
-    const ggml_tensor * V   = dst->src[2];
-
-    GGML_ASSERT(K->type == type_K);
-    GGML_ASSERT(V->type == type_V);
 
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -1,3 +1,5 @@
+#pragma once
+
 #include "common.cuh"
 
 #if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)

ggml/src/ggml-cuda/fattn.cu

@@ -116,11 +116,15 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
     }
 }
 
-#define FATTN_VEC_CASE(D, type_K, type_V)                                \
-    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) { \
-        ggml_cuda_flash_attn_ext_vec_case<D, type_K, type_V>(ctx, dst);  \
-        return;                                                          \
-    }                                                                    \
+#define FATTN_VEC_CASE(D, type_K, type_V)                                                                        \
+    {                                                                                                            \
+        const bool type_K_okay = K->type == (type_K) || (K->type == GGML_TYPE_F32 && (type_K) == GGML_TYPE_F16); \
+        const bool type_V_okay = V->type == (type_V) || (V->type == GGML_TYPE_F32 && (type_V) == GGML_TYPE_F16); \
+        if (Q->ne[0] == (D) && type_K_okay && type_V_okay) {                                                     \
+            ggml_cuda_flash_attn_ext_vec_case<D, type_K, type_V>(ctx, dst);                                      \
+            return;                                                                                              \
+        }                                                                                                        \
+    }                                                                                                            \
 
 #define FATTN_VEC_CASES_ALL_D(type_K, type_V) \
     FATTN_VEC_CASE( 64, type_K, type_V)       \
@@ -198,6 +202,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     return BEST_FATTN_KERNEL_NONE;
 #endif// FLASH_ATTN_AVAILABLE
 
+    const ggml_tensor * KQV   = dst;
     const ggml_tensor * Q     = dst->src[0];
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
@@ -206,37 +211,33 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     const int gqa_ratio = Q->ne[2] / K->ne[2];
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
 
+    float max_bias = 0.0f;
+    memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+    // The effective batch size for the kernel can be increased by gqa_ratio.
+    // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
+    const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+
     const int cc = ggml_cuda_info().devices[device].cc;
 
-    // TODO: temporary until support is extended
-    //       https://github.com/ggml-org/llama.cpp/pull/16148#issuecomment-3343525206
-    if (K->ne[1] % FATTN_KQ_STRIDE != 0) {
-        return BEST_FATTN_KERNEL_NONE;
-    }
-
     switch (K->ne[0]) {
+        case  40:
         case  64:
-        case 128:
-        case 256:
-            if (V->ne[0] != K->ne[0]) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
-            break;
+        case  72:
         case  80:
         case  96:
+        case 128:
         case 112:
+        case 256:
             if (V->ne[0] != K->ne[0]) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!ggml_cuda_should_use_wmma_fattn(cc) && !turing_mma_available(cc)) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
             break;
         case 576:
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!turing_mma_available(cc) || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;
@@ -251,6 +252,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
     switch (K->type) {
+        case GGML_TYPE_F32:
         case GGML_TYPE_F16:
             break;
         case GGML_TYPE_Q4_1:
@@ -270,47 +272,57 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_NONE;
     }
 
-    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0;
-
-    // If Turing tensor cores available, use them except for some cases with batch size 1:
-    if (turing_mma_available(cc)) {
-        best_fattn_kernel best = BEST_FATTN_KERNEL_MMA_F16;
+    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
+    // If Turing tensor cores available, use them:
+    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if (can_use_vector_kernel) {
-            if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+            if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
-                    best = BEST_FATTN_KERNEL_VEC;
+                    return BEST_FATTN_KERNEL_VEC;
                 }
             } else {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
                     if (Q->ne[1] <= 2) {
-                        best = BEST_FATTN_KERNEL_VEC;
+                        return BEST_FATTN_KERNEL_VEC;
                     }
                 } else {
                     if (Q->ne[1] == 1) {
-                        best = BEST_FATTN_KERNEL_VEC;
+                        return BEST_FATTN_KERNEL_VEC;
                     }
                 }
             }
-            if ((gqa_ratio % 2 != 0 || !mask) && Q->ne[1] == 1) {
-                best = BEST_FATTN_KERNEL_VEC; // GQA-specific optimizations in the mma kernel do not apply.
+            if (!gqa_opt_applies && Q->ne[1] == 1) {
+                return BEST_FATTN_KERNEL_VEC;
             }
         }
 
-        return best;
+        return BEST_FATTN_KERNEL_MMA_F16;
     }
 
-    // Use kernels specialized for small batch sizes if possible:
-    if (Q->ne[1] <= 8 && can_use_vector_kernel) {
-        return BEST_FATTN_KERNEL_VEC;
-    }
-
-    // For large batch sizes, use the WMMA kernel if possible:
-    if (ggml_cuda_should_use_wmma_fattn(cc)) {
+    // Use the WMMA kernel if possible:
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 576) {
+        if (can_use_vector_kernel && Q->ne[1] <= 2) {
+            return BEST_FATTN_KERNEL_VEC;
+        }
         return BEST_FATTN_KERNEL_WMMA_F16;
     }
 
-    // If there is no suitable kernel for tensor cores or small batch sizes, use the generic kernel for large batch sizes:
+    // If there are no tensor cores available, use the generic tile kernel:
+    if (can_use_vector_kernel) {
+        if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
+            if (Q->ne[1] == 1) {
+                if (!gqa_opt_applies) {
+                    return BEST_FATTN_KERNEL_VEC;
+                }
+            }
+        } else {
+            if (Q->ne[1] <= 2) {
+                return BEST_FATTN_KERNEL_VEC;
+            }
+        }
+    }
     return BEST_FATTN_KERNEL_TILE;
 }
 

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

@@ -50,6 +50,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
@@ -273,6 +274,15 @@ static ggml_cuda_device_info ggml_cuda_init() {
         } else if (device_name.substr(0, 21) == "NVIDIA GeForce GTX 16") {
             turing_devices_without_mma.push_back({ id, device_name });
         }
+
+        // Temporary performance fix:
+        // Setting device scheduling strategy for iGPUs with cc121 to "spinning" to avoid delays in cuda synchronize calls.
+        // TODO: Check for future drivers the default scheduling strategy and
+        // remove this call again when cudaDeviceScheduleSpin is default.
+        if (prop.major == 12 && prop.minor == 1) {
+            CUDA_CHECK(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
+        }
+
 #endif  // defined(GGML_USE_HIP)
     }
 
@@ -1948,8 +1958,15 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
 
         size_t src1_stride_size = sizeof(cuda_t);
 
-        dim3 block_dims(ne13, ne12);
-        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
+        const int threads_x = 16;
+        const int threads_y = 16;
+        dim3 block_dims(threads_x, threads_y);
+
+        dim3 grid_dims(
+            (ne13 + threads_x - 1) / threads_x,
+            (ne12 + threads_y - 1) / threads_y
+        );
+        k_compute_batched_ptrs<<<grid_dims, block_dims, 0, main_stream>>>(
                 src0_ptr, src1_ptr, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
@@ -1998,6 +2015,164 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     }
 }
 
+static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
+                                          const ggml_tensor * ffn_gate,
+                                          const ggml_tensor * glu,
+                                          const ggml_tensor * ffn_up_bias = nullptr,
+                                          const ggml_tensor * ffn_gate_bias = nullptr) {
+    const bool has_bias = ffn_up_bias != nullptr || ffn_gate_bias != nullptr;
+
+    if (has_bias && (!ffn_up_bias || !ffn_gate_bias)) {
+        return false;
+    }
+
+    const bool is_mul_mat     = ffn_up->op == GGML_OP_MUL_MAT     && ffn_gate->op == GGML_OP_MUL_MAT     && glu->op == GGML_OP_GLU;
+    const bool is_mul_mat_id  = ffn_up->op == GGML_OP_MUL_MAT_ID  && ffn_gate->op == GGML_OP_MUL_MAT_ID  && glu->op == GGML_OP_GLU;
+
+    GGML_ASSERT(ffn_up && ffn_gate && glu);
+
+    if (!is_mul_mat && !is_mul_mat_id) {
+        return false;
+    }
+
+    const ggml_op expected_bias_op = is_mul_mat ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+    if (has_bias) {
+        if (ffn_up_bias->op != expected_bias_op || ffn_gate_bias->op != expected_bias_op) {
+            return false;
+        }
+
+        if (glu->src[0] != ffn_gate_bias || glu->src[1] != ffn_up_bias) {
+            return false;
+        }
+
+        if (expected_bias_op == GGML_OP_ADD) {
+            const bool up_has_mul   = ffn_up_bias->src[0] == ffn_up || ffn_up_bias->src[1] == ffn_up;
+            const bool gate_has_mul = ffn_gate_bias->src[0] == ffn_gate || ffn_gate_bias->src[1] == ffn_gate;
+            if (!up_has_mul || !gate_has_mul) {
+                return false;
+            }
+        } else { // GGML_OP_ADD_ID
+            if (ffn_up_bias->src[0] != ffn_up || ffn_gate_bias->src[0] != ffn_gate) {
+                return false;
+            }
+            if (ffn_up_bias->src[2] != ffn_up->src[2] || ffn_gate_bias->src[2] != ffn_gate->src[2]) {
+                return false;
+            }
+        }
+    } else {
+        if (glu->src[0] != ffn_gate && glu->src[1] != ffn_up) {
+            return false;
+        }
+    }
+
+    if (ffn_up->src[0]->type != ffn_gate->src[0]->type || !ggml_are_same_shape(ffn_up->src[0], ffn_gate->src[0]) ||
+        !ggml_are_same_stride(ffn_up->src[0], ffn_gate->src[0])) {
+        return false;
+    }
+
+    if (ffn_up->src[1] != ffn_gate->src[1]) {
+        return false;
+    }
+
+    if (ffn_up->src[2] && (ffn_up->src[2] != ffn_gate->src[2])) {
+        return false;
+    }
+
+    static constexpr std::array<ggml_glu_op, 3> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI };
+
+    if (std::find(valid_glu_ops.begin(), valid_glu_ops.end(), ggml_get_glu_op(glu)) == valid_glu_ops.end()) {
+        return false;
+    }
+
+    if (const bool swapped = ggml_get_op_params_i32(glu, 1); swapped) {
+        return false;
+    }
+
+    const bool split = ggml_backend_buft_is_cuda_split(ffn_up->src[0]->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(ffn_gate->src[0]->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return true;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool is_mul_mat_id = tensor->op == GGML_OP_MUL_MAT_ID;
+
+    bool use_mul_mat_vec_f =
+        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) &&
+        src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    return use_mul_mat_vec_f;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE &&
+                                   ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) &&
+                                   src0->view_src;
+
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 &&
+                             dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    // fusion is not universally faster on Pascal
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    if (cc <= GGML_CUDA_CC_PASCAL) {
+        return false;
+    }
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return use_mul_mat_vec_q;
+}
+
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
@@ -2031,16 +2206,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
             const int cc            = ggml_cuda_info().devices[id].cc;
             const int warp_size     = ggml_cuda_info().devices[id].warp_size;
             use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
         const int cc            = ggml_cuda_info().devices[ctx.device].cc;
         const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
         use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
@@ -2111,7 +2286,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             return;
         }
 
-        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2], /*mul_mat_id=*/true)) {
+        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
             ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
             return;
         }
@@ -2259,6 +2434,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SET_ROWS:
             ggml_cuda_op_set_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cuda_op_set(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2337,6 +2515,18 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_XIELU:
                     ggml_cuda_op_xielu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_FLOOR:
+                    ggml_cuda_op_floor(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_CEIL:
+                    ggml_cuda_op_ceil(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ROUND:
+                    ggml_cuda_op_round(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TRUNC:
+                    ggml_cuda_op_trunc(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2633,11 +2823,10 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
 }
 
 #ifdef USE_CUDA_GRAPH
-static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
+static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
     bool use_cuda_graph) {
 
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
-    cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
     const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
     const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
@@ -2688,33 +2877,11 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_CPY) {
-
-            // Store the pointers which are updated for each token, such that these can be sent
-            // to the device and accessed using indirection from CUDA graph
-            cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data);
-
-            // store a pointer to each copy op CUDA kernel to identify it later
-            void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
-            if (!ptr) {
-                use_cuda_graph = false;
-#ifndef NDEBUG
-                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
-#endif
-            }
-        }
-
         if (!use_cuda_graph) {
             break;
         }
     }
 
-    if (use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = true;
-        // copy pointers to GPU so they can be accessed via indirection within CUDA graph
-        ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream());
-    }
-
     return use_cuda_graph;
 }
 
@@ -2733,7 +2900,6 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
 
 static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
     if (node->data != graph_node_properties->node_address &&
-          node->op != GGML_OP_CPY &&
           node->op != GGML_OP_VIEW) {
         return false;
     }
@@ -2754,14 +2920,13 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (node->src[i] &&
             node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_CPY &&
             node->op != GGML_OP_VIEW
         ) {
             return false;
         }
     }
 
-    if (node->op == GGML_OP_SCALE &&
+    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
         memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
         return false;
     }
@@ -2834,43 +2999,74 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
 #endif
 
     //TODO: remove special case once ggml_can_fuse can handle empty nodes
-    std::initializer_list<enum ggml_op> topk_moe_ops           = ggml_cuda_topk_moe_ops(false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm = ggml_cuda_topk_moe_ops(true);
+    std::initializer_list<enum ggml_op> topk_moe_ops =
+        ggml_cuda_topk_moe_ops(/*with_norm*/ false, /*delayed_softmax=*/false);
+    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm =
+        ggml_cuda_topk_moe_ops(/*with_norm=*/true, /*delayed_softmax=*/false);
+    std::initializer_list<enum ggml_op> topk_moe_ops_delayed_softmax =
+        ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
 
-    if (ops.size() == topk_moe_ops_with_norm.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops_with_norm.begin())) {
-
-        if (node_idx + topk_moe_ops_with_norm.size() > (size_t)cgraph->n_nodes) {
-            return false;
-        }
-
-        for (size_t i = 0; i < topk_moe_ops_with_norm.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops_with_norm.begin()[i]) return false;
-        }
+    if (ops.size() == topk_moe_ops_with_norm.size() &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx+8];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 9];
 
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
             return true;
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops.begin())) {
-
-        if (node_idx + topk_moe_ops.size() > (size_t)cgraph->n_nodes) {
-            return false;
-        }
-
-        for (size_t i = 0; i < topk_moe_ops.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops.begin()[i]) return false;
-        }
-
+    if (ops.size() == topk_moe_ops.size() &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx+4];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 4];
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
             return true;
         }
     }
 
+    if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
+        ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 5];
+
+        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+            return true;
+        }
+    }
+
+    std::initializer_list<enum ggml_op> mul_mat_bias_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_GLU };
+    std::initializer_list<enum ggml_op> mul_mat_id_bias_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
+
+    std::initializer_list<enum ggml_op> mul_mat_id_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_MUL_MAT_ID, GGML_OP_GLU };
+    std::initializer_list<enum ggml_op> mul_mat_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_MUL_MAT,    GGML_OP_GLU };
+
+    if (ops.size() == 5 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}) ||
+                            ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}))) {
+
+        const ggml_tensor * ffn_gate      = cgraph->nodes[node_idx];
+        const ggml_tensor * ffn_gate_bias = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * ffn_up        = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * ffn_up_bias   = cgraph->nodes[node_idx + 3];
+        const ggml_tensor * glu           = cgraph->nodes[node_idx + 4];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu, ffn_up_bias, ffn_gate_bias)) {
+            return true;
+        }
+    }
+
+    if (ops.size() == 3 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}) ||
+                            ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}))) {
+
+        const ggml_tensor * ffn_gate = cgraph->nodes[node_idx];
+        const ggml_tensor * ffn_up   = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * glu      = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu)) {
+            return true;
+        }
+    }
+
     if (!ggml_can_fuse(cgraph, node_idx, ops)) {
         return false;
     }
@@ -2901,7 +3097,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
 
         //if rms norm is the B operand, then we don't handle broadcast
-        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
 
@@ -2951,8 +3147,17 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
 
+            [[maybe_unused]] int prev_i = 0;
+
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
+#ifdef GGML_CUDA_DEBUG
+                const int nodes_fused = i - prev_i - 1;
+                prev_i = i;
+                if (nodes_fused > 0) {
+                    GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
+                }
+#endif
 
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
@@ -2962,21 +3167,35 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 if (!disable_fusion) {
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+8];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ true);
-                        i += 8;
+                        ggml_tensor * weights          = cgraph->nodes[i + 9];
+                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
+                        ggml_tensor * clamp            = cgraph->nodes[i + 7];
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ true,
+                                              /*delayed softmax*/ false, clamp);
+                        i += 9;
                         continue;
                     }
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+4];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ false);
+                        ggml_tensor * weights          = cgraph->nodes[i + 4];
+                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ false,
+                                              /*delayed softmax*/ false);
                         i += 4;
                         continue;
                     }
 
+                    if (ggml_cuda_can_fuse(cgraph, i,
+                                           ggml_cuda_topk_moe_ops(/*with norm*/ false, /*delayed softmax*/ true), {})) {
+                        ggml_tensor * weights = cgraph->nodes[i + 5];
+                        ggml_tensor * ids     = cgraph->nodes[i + 1];
+
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, ids, /*with norm*/ false,
+                                              /*delayed_softmax*/ true);
+                        i += 5;
+                        continue;
+                    }
+
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];
@@ -3008,6 +3227,195 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         }
                     }
 
+                    bool fused_mul_mat_vec = false;
+                    int fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu         = cgraph->nodes[i + 4];
+                            ggml_tensor * gate_bias_n = glu->src[0];
+                            ggml_tensor * up_bias_n   = glu->src[1];
+
+                            //we don't assume the order for {gate, up}. Instead infer it from the bias tensor
+                            ggml_tensor * gate_n      = nullptr;
+                            ggml_tensor * up_n        = nullptr;
+
+                            if (gate_bias_n->src[0] == cgraph->nodes[i] || gate_bias_n->src[1] == cgraph->nodes[i]) {
+                                gate_n = cgraph->nodes[i];
+                                up_n   = cgraph->nodes[i + 2];
+                            } else if (gate_bias_n->src[0] == cgraph->nodes[i + 2] || gate_bias_n->src[1] == cgraph->nodes[i + 2]) {
+                                gate_n = cgraph->nodes[i + 2];
+                                up_n   = cgraph->nodes[i];
+                            } else {
+                                continue;
+                            }
+
+                            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
+                                if (op_bias == GGML_OP_ADD) {
+                                    if (bias_node->src[0] == mul_node) {
+                                        return bias_node->src[1];
+                                    }
+                                    if (bias_node->src[1] == mul_node) {
+                                        return bias_node->src[0];
+                                    }
+                                    return (ggml_tensor *) nullptr;
+                                }
+                                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
+                                GGML_ASSERT(bias_node->src[0] == mul_node);
+                                return bias_node->src[1];
+                            };
+
+                            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
+                            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
+
+                            if (!up_bias_tensor || !gate_bias_tensor) {
+                                continue;
+                            }
+
+                            // we don't support repeating adds
+                            if (bias_op == GGML_OP_ADD &&
+                                (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
+                                 !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
+                                continue;
+                            }
+
+                            const ggml_tensor * src0 = up_n->src[0];
+                            const ggml_tensor * src1 = up_n->src[1];
+                            const ggml_tensor * ids  = up_n->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+                        } else if (ggml_cuda_can_fuse(cgraph, i, { op, op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu  = cgraph->nodes[i + 2];
+                            ggml_tensor * gate = glu->src[0];
+                            ggml_tensor * up   = glu->src[1];
+
+                            bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1])
+                                || (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
+
+                            if (!ok) continue;
+
+                            const ggml_tensor * src0 = up->src[0];
+                            const ggml_tensor * src1 = up->src[1];
+                            const ggml_tensor * ids  = up->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
+
+                    fused_mul_mat_vec = false;
+                    fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (!ggml_can_fuse(cgraph, i, { op, bias_op })) {
+                            continue;
+                        }
+
+                        ggml_tensor * mm_node   = cgraph->nodes[i];
+                        ggml_tensor * bias_node = cgraph->nodes[i + 1];
+
+                        ggml_tensor * bias_tensor = nullptr;
+                        if (bias_op == GGML_OP_ADD) {
+                            if (bias_node->src[0] == mm_node) {
+                                bias_tensor = bias_node->src[1];
+                            } else if (bias_node->src[1] == mm_node) {
+                                bias_tensor = bias_node->src[0];
+                            } else {
+                                continue;
+                            }
+                        } else {
+                            if (bias_node->src[0] != mm_node) {
+                                continue;
+                            }
+                            bias_tensor = bias_node->src[1];
+                        }
+
+                        const ggml_tensor * src0 = mm_node->src[0];
+                        const ggml_tensor * src1 = mm_node->src[1];
+                        const ggml_tensor * ids  = mm_node->src[2];
+
+                        if (bias_op == GGML_OP_ADD_ID && bias_node->src[2] != ids) {
+                            continue;
+                        }
+
+                        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
+                            continue;
+                        }
+
+                        ggml_cuda_mm_fusion_args_host fusion_data{};
+                        fusion_data.x_bias = bias_tensor;
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
+                            ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
+                            ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
 
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD}, {})) {
                         ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
@@ -3120,7 +3528,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     if (use_cuda_graph) {
         cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
 
-        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph);
+        use_cuda_graph = check_node_graph_compatibility(cgraph, use_cuda_graph);
 
         // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
         if (use_cuda_graph && cuda_graph_update_required) {
@@ -3147,10 +3555,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    if (!use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = false;
-    }
-
 #else
     bool use_cuda_graph = false;
     bool cuda_graph_update_required = false;
@@ -3377,6 +3781,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3491,6 +3899,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                        op->src[0]->type == GGML_TYPE_F32 &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
+        case GGML_OP_SET:
+            {
+                const ggml_type t = op->type;
+                return (t == GGML_TYPE_F32 || t == GGML_TYPE_I32) &&
+                    t == op->src[0]->type &&
+                    t == op->src[1]->type;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3645,12 +4060,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
-        case GGML_OP_SUM:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM:
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_ARGSORT:
-            // TODO: Support arbitrary column width
+#ifndef GGML_CUDA_USE_CUB
             return op->src[0]->ne[0] <= 1024;
+#else
+            return true;
+#endif
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:

ggml/src/ggml-cuda/mma.cuh

@@ -18,6 +18,10 @@
 
 #include "common.cuh"
 
+// On Volta each warp is doing 4 8x8 mma operations in parallel.
+// The basic memory layout for a 32x8 output tile is to stack 4 input tiles in I direction and to mirror the B tile.
+// However, the i indices in this file are by default permuted to simplify the index calculations.
+// #define GGML_CUDA_MMA_NO_VOLTA_PERM
 
 #if CUDART_VERSION >= 11080
 
@@ -73,6 +77,15 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / 64;
         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;
+            if (I == 32 && J == 32) return true;
+            return false;
+        }
+
         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>
                 return threadIdx.x % 16;
@@ -85,7 +98,8 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -101,22 +115,67 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return threadIdx.x % 32;
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+#else
+                return (l & 2) | (threadIdx.x & ~2);
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 32 && J == 8) {
+                return (threadIdx.x & 2) | (l & (4 + 1));
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && (J == 4 || J == 8)) {
+            if constexpr (I == 8 && J == 4) {
+                return threadIdx.x / 4;
+            } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 8 + threadIdx.x / 4;
+                return ((l / 2) * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 2) % 2) * 8 + threadIdx.x / 4;
+                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -124,13 +183,16 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return 4 * l + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return 2 * (threadIdx.x % 4) + l % 2;
+                return ((threadIdx.x % 4) * 2) | (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
+                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #endif // defined(GGML_USE_HIP)
@@ -140,32 +202,83 @@ namespace ggml_cuda_mma {
     struct tile<I_, J_, half2> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+#else
+                return threadIdx.x;
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr ((I == 8 || I == 32) && J == 8) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l & 2) * 2) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     };
 
     template <int I_, int J_>
@@ -175,27 +288,36 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / WARP_SIZE;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  4) return true;
+            if (I == 16 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
     };
@@ -263,8 +385,12 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(xs0, stride);
-        GGML_UNUSED(t);
+#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
     }
 
@@ -277,11 +403,35 @@ 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
+        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
     }
 
+    template <typename T>
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if 1
+        // TODO: more generic handling
+        static_assert(sizeof(T) == 4, "bad type size");
+        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);
+#else
+        load_generic(t, xs0, stride);
+#endif // 1
+#else
+        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
+        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
+        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+    }
+
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
@@ -546,4 +696,43 @@ namespace ggml_cuda_mma {
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
+
+    template <typename T1, typename T2, int J, int K>
+    static __device__ __forceinline__ void mma(
+            tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
+        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
+        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
+#else
+        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
+        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+    }
 }

ggml/src/ggml-cuda/mmf.cu

@@ -1,5 +1,7 @@
 #include "ggml.h"
 #include "mmf.cuh"
+#include "mmid.cuh"
+
 
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
@@ -37,6 +39,12 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     const int64_t ids_s0 = ids ? ids->nb[0] / ggml_type_size(ids->type) : 0;
     const int64_t ids_s1 = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
 
+    mmf_ids_data ids_info{};
+    mmf_ids_data * ids_info_ptr = nullptr;
+    ggml_cuda_pool_alloc<int32_t> ids_src_compact_dev;
+    ggml_cuda_pool_alloc<int32_t> ids_dst_compact_dev;
+    ggml_cuda_pool_alloc<int32_t> expert_bounds_dev;
+
     // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
     const int64_t ncols_dst          = ids ? ne2  : ne1;
     const int64_t nchannels_dst      = ids ? ne1 : ne2;
@@ -54,6 +62,33 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
         nchannels_y      = ids->ne[0];
     }
 
+    if (ids && ncols_dst > 16) {
+        const int64_t n_expert_used = ids->ne[0];
+        const int64_t n_experts     = ne02;
+        const int64_t n_tokens      = ne12;
+        const int64_t ne_get_rows   = n_tokens * n_expert_used;
+
+        ids_src_compact_dev.alloc(ctx.pool(), ne_get_rows);
+        ids_dst_compact_dev.alloc(ctx.pool(), ne_get_rows);
+        expert_bounds_dev.alloc(ctx.pool(), n_experts + 1);
+
+        const int si1  = static_cast<int>(ids_s1);
+        const int sis1 = static_cast<int>(src1->nb[2] / src1->nb[1]);
+
+        GGML_ASSERT(sis1 > 0);
+
+        ggml_cuda_launch_mm_ids_helper(ids_d, ids_src_compact_dev.get(), ids_dst_compact_dev.get(), expert_bounds_dev.get(),
+            static_cast<int>(n_experts), static_cast<int>(n_tokens), static_cast<int>(n_expert_used), static_cast<int>(ne11), si1, sis1, ctx.stream());
+        CUDA_CHECK(cudaGetLastError());
+
+        ids_info.ids_src_compact   = ids_src_compact_dev.get();
+        ids_info.ids_dst_compact   = ids_dst_compact_dev.get();
+        ids_info.expert_bounds_dev = expert_bounds_dev.get();
+        ids_info.n_experts         = static_cast<int>(n_experts);
+        ids_info.sis1              = sis1;
+        ids_info_ptr = &ids_info;
+    }
+
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
@@ -61,7 +96,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_F16: {
             const half2 * src0_d = (const half2 *) src0->data;
@@ -69,7 +104,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data;
@@ -77,31 +112,42 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
             mul_mat_f_switch_cols_per_block(
                 src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
     }
 }
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols, bool mul_mat_id) {
-
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne,
+        const size_t * src0_nb, const int src1_ncols, bool mul_mat_id) {
     if (ggml_is_quantized(type)) {
         return false;
     }
 
-    if (src0_ne[0] % (warp_size * (4/ggml_type_size(type))) != 0) {
+    const size_t ts = ggml_type_size(type);
+    if (src0_ne[0] % (warp_size * (4/ts)) != 0) {
         return false;
     }
+
+    if (src0_nb[0] != ts) {
+        return false;
+    }
+
+    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
+    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0_nb[i] % (2*ts) != 0) {
+            return false;
+        }
+    }
     if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
         return false;
     }
 
     if (mul_mat_id) {
-        if (type == GGML_TYPE_F32 && src1_ncols > 32) {
+        if (src0_ne[1] <= 1024 && src1_ncols > 512) {
             return false;
-        }
-        if ((type == GGML_TYPE_F16 || type == GGML_TYPE_BF16) && src1_ncols > 64) {
+        } else if(src0_ne[1] > 1024 && src1_ncols > 128) {
             return false;
         }
     } else {
@@ -114,7 +160,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return turing_mma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc);
         case GGML_TYPE_BF16:
             return ampere_mma_available(cc);
         default:

ggml/src/ggml-cuda/mmf.cuh

@@ -7,9 +7,17 @@ using namespace ggml_cuda_mma;
 
 #define MMF_ROWS_PER_BLOCK 32
 
+struct mmf_ids_data {
+    const int32_t * ids_src_compact = nullptr;
+    const int32_t * ids_dst_compact = nullptr;
+    const int32_t * expert_bounds_dev = nullptr;
+    int n_experts = 0;
+    int sis1 = 0;
+};
+
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols, bool mul_mat_id);
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const size_t * src0_nb, const int src1_ncols, bool mul_mat_id);
 
 template <typename T, int rows_per_block, int cols_per_block, int nwarps, bool has_ids>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
@@ -20,9 +28,19 @@ static __global__ void mul_mat_f(
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -224,6 +242,259 @@ static __global__ void mul_mat_f(
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
+//This kernel is for larger batch sizes of mul_mat_id
+template <typename T, int rows_per_block, int cols_per_block, int nwarps>
+__launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
+static __global__ void mul_mat_f_ids(
+        const T * __restrict__ x, const float * __restrict__ y,
+        const int32_t * __restrict__ ids_src_compact, const int32_t * __restrict__ ids_dst_compact,
+        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst,
+        const int ncols, const int ncols_dst_total, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const uint3 sis1_fd, const uint3 nch_fd) {
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work butr 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int ntA = rows_per_block / tile_A::I;
+    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
+
+    const int row0        = blockIdx.x * rows_per_block;
+
+    const int expert_idx = blockIdx.y;
+    const int expert_start = expert_bounds[expert_idx];
+    const int expert_end   = expert_bounds[expert_idx + 1];
+    const int ncols_expert = expert_end - expert_start;
+
+    const int tiles_for_expert = (ncols_expert + cols_per_block - 1) / cols_per_block;
+    const int tile_idx = blockIdx.z;
+    if (tile_idx >= tiles_for_expert) {
+        return;
+    }
+
+    const int col_base = tile_idx * cols_per_block;
+
+    GGML_UNUSED(channel_ratio);
+
+    const int channel_x   = expert_idx;
+    const int sample_dst  = 0;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x  + row0*stride_row;
+    y   += int64_t(sample_y)  *stride_sample_y;
+    dst += int64_t(sample_dst)*stride_sample_dst;
+
+    const int32_t * ids_src_expert = ids_src_compact + expert_start;
+    const int32_t * ids_dst_expert = ids_dst_compact + expert_start;
+
+    extern __shared__ char data_mmv[];
+    char * compute_base = data_mmv;
+
+    //const float2 * y2 = (const float2 *) y;
+
+    tile_C C[ntA][ntB];
+
+    T * tile_xy = (T *) compute_base + threadIdx.y*(tile_A::I * tile_k_padded);
+
+    for (int col = threadIdx.y*warp_size + threadIdx.x; col < ncols; col += nwarps*warp_size) {
+        tile_A A[ntA][warp_size / tile_A::J];
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int i = 0; i < tile_A::I; ++i) {
+                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
+                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
+            }
+        }
+
+        if constexpr (std::is_same_v<T, float>) {
+            float vals_buf[2][tile_B::I];
+            auto gather_tile = [&](int tile_idx_local, float *vals) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + tile_idx_local*tile_B::I;
+                    const int global_j = col_base + j;
+                    float val = 0.0f;
+                    if (j < cols_per_block && global_j < ncols_expert) {
+                        const int src_entry = ids_src_expert[global_j];
+                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
+                        const int token   = (int) qrm.x;
+                        const int channel = (int) qrm.y;
+                        if (token < ncols_dst_total) {
+                            val = y[channel*stride_channel_y + token*stride_col_y + col];
+                        }
+                    }
+                    vals[j0] = val;
+                }
+            };
+
+            gather_tile(0, vals_buf[0]);
+
+            int curr_buf = 0;
+            int next_buf = 1;
+#pragma unroll
+            for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = vals_buf[curr_buf][j0];
+                }
+
+                if (itB + 1 < ntB) {
+                    gather_tile(itB + 1, vals_buf[next_buf]);
+                }
+
+#pragma unroll
+                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                    tile_B B;
+                    load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                    for (int itA = 0; itA < ntA; ++itA) {
+                        mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                    }
+                }
+
+                if (itB + 1 < ntB) {
+                    curr_buf ^= 1;
+                    next_buf ^= 1;
+                }
+            }
+        } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+            float2 vals_buf[2][tile_B::I];
+            auto gather_tile = [&](int tile_idx_local, float2 *vals) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + tile_idx_local*tile_B::I;
+                    const int global_j = col_base + j;
+                    float2 tmp = make_float2(0.0f, 0.0f);
+                    if (j < cols_per_block && global_j < ncols_expert) {
+                        const int src_entry = ids_src_expert[global_j];
+                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
+                        const int token   = (int) qrm.x;
+                        const int channel = (int) qrm.y;
+                        if (token < ncols_dst_total) {
+                            tmp = *(const float2*) &y[channel*stride_channel_y + 2*(token*stride_col_y + col)];
+                        }
+                    }
+                    vals[j0] = tmp;
+                }
+            };
+
+            if (ntB > 0) {
+                gather_tile(0, vals_buf[0]);
+            }
+
+            int curr_buf = 0;
+            int next_buf = 1;
+#pragma unroll
+            for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const float2 tmp = vals_buf[curr_buf][j0];
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                }
+
+                if (itB + 1 < ntB) {
+                    gather_tile(itB + 1, vals_buf[next_buf]);
+                }
+
+#pragma unroll
+                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                    tile_B B;
+                    load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                    for (int itA = 0; itA < ntA; ++itA) {
+                        mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                    }
+                }
+
+                if (itB + 1 < ntB) {
+                    curr_buf ^= 1;
+                    next_buf ^= 1;
+                }
+            }
+        } else {
+            static_assert(std::is_same_v<T, void>, "unsupported type");
+        }
+    }
+
+    float * buf_iw = (float *) compute_base;
+    constexpr int kiw = nwarps*rows_per_block + 4;
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+#pragma unroll
+    for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int l = 0; l < tile_C::ne; ++l) {
+                const int i = threadIdx.y*rows_per_block + itA*tile_C::I + tile_C::get_i(l);
+                const int j = itB*tile_C::J + tile_C::get_j(l);
+                buf_iw[j*kiw + i] = C[itA][itB].x[l];
+            }
+        }
+    }
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j0 + nwarps > cols_per_block && j >= cols_per_block) {
+            return;
+        }
+
+        float sum = 0.0f;
+        static_assert(rows_per_block == warp_size, "need loop/check");
+#pragma unroll
+        for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
+            const int i = i0 + threadIdx.x;
+
+            sum += buf_iw[j*kiw + i];
+        }
+
+        const int global_j = col_base + j;
+        if (j < cols_per_block && global_j < ncols_expert && nchannels_dst > 0) {
+            const int dst_entry = ids_dst_expert[global_j];
+            const uint2 qrm = fast_div_modulo((uint32_t) dst_entry, nch_fd);
+            const int token = (int) qrm.x;
+            if (token < ncols_dst_total) {
+                const int slot = (int) qrm.y;
+                dst[slot*stride_channel_dst + token*stride_col_dst + row0 + threadIdx.x] = sum;
+            }
+        }
+    }
+#else
+    GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
+        ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd);
+    NO_DEVICE_CODE;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
 template<typename T, int cols_per_block, int nwarps>
 static inline void mul_mat_f_switch_ids(
         const T * x, const float * y, const int32_t * ids, float * dst,
@@ -232,13 +503,35 @@ static inline void mul_mat_f_switch_ids(
         const int64_t stride_col_id, const int64_t stride_row_id,
         const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
         const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream) {
-    if (ids) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream,
+        const mmf_ids_data * ids_data) {
+    const bool has_ids_data = ids_data && ids_data->ids_src_compact;
+
+    // Use the compact-ids kernel only for larger tiles; for small ncols_dst (< 16)
+    // we prefer the normal mul_mat_f path with has_ids=true.
+    if (has_ids_data && ncols_dst > 16) {
+        const int max_tiles = (int) ((ncols_dst + cols_per_block - 1) / cols_per_block);
+        if (max_tiles == 0) {
+            return;
+        }
+        dim3 block_nums_ids(block_nums.x, ids_data->n_experts, max_tiles);
+
+        const uint3 sis1_fd = ids_data->sis1 > 0 ? init_fastdiv_values((uint32_t) ids_data->sis1) : make_uint3(0, 0, 1);
+        const uint3 nch_fd  = init_fastdiv_values((uint32_t) nchannels_dst);
+
+        mul_mat_f_ids<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
+            (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->expert_bounds_dev, dst,
+            ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+            sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst,
+            sis1_fd, nch_fd);
+    } else if (ids) {
         const int64_t col_tiles = (ncols_dst + cols_per_block - 1) / cols_per_block;
         dim3 block_nums_ids = block_nums;
         block_nums_ids.y *= col_tiles;
+
         mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
-             (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
              stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
              sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     } else {
@@ -258,8 +551,9 @@ void mul_mat_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
-    typedef tile<16, 8, T>     tile_A;
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
+    typedef tile<16, 8, T>     tile_A_16;
+    typedef tile<32, 8, T>     tile_A_32;
     typedef tile< 8, 8, T>     tile_B;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
@@ -270,7 +564,8 @@ void mul_mat_f_cuda(
     const int64_t channel_ratio = nchannels_dst / nchannels_x;
     const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
 
-    const int device = ggml_cuda_get_device();
+    const int device    = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
 
     int64_t nwarps_best     = 1;
@@ -285,12 +580,12 @@ void mul_mat_f_cuda(
     }
 
     constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
+    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4;
     const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
     const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;
-    const int64_t grid_y = ids ? nchannels_x : nchannels_dst; // per expert when ids present
+    const int64_t grid_y = ids ? nchannels_x : nchannels_dst;
 
     const dim3 block_nums(nrows_x/rows_per_block, grid_y, nsamples_dst);
     const dim3 block_dims(warp_size, nwarps_best, 1);
@@ -300,49 +595,57 @@ void mul_mat_f_cuda(
             mul_mat_f_switch_ids<T, cols_per_block, 1>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 2: {
             mul_mat_f_switch_ids<T, cols_per_block, 2>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 3: {
             mul_mat_f_switch_ids<T, cols_per_block, 3>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 4: {
             mul_mat_f_switch_ids<T, cols_per_block, 4>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 5: {
             mul_mat_f_switch_ids<T, cols_per_block, 5>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 6: {
             mul_mat_f_switch_ids<T, cols_per_block, 6>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 7: {
             mul_mat_f_switch_ids<T, cols_per_block, 7>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         case 8: {
             mul_mat_f_switch_ids<T, cols_per_block, 8>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
+                ids_data);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -361,7 +664,7 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
 
     const int ncols_case = (ids && ncols_dst > 16) ? 16 : ncols_dst;
 
@@ -371,82 +674,82 @@ static void mul_mat_f_switch_cols_per_block(
         case  1: {
             mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  2: {
             mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  3: {
             mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  4: {
             mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  5: {
             mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream, ids_data);
         } break;
         case  6: {
             mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  7: {
             mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  8: {
             mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  9: {
             mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 10: {
             mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 11: {
             mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 12: {
             mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 13: {
             mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 14: {
             mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 15: {
             mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 16: {
             mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, 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, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -462,7 +765,7 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, \
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,\
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, \
-        cudaStream_t stream);
+        cudaStream_t stream, const mmf_ids_data * ids_data);
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 #define DECL_MMF_CASE_EXTERN(ncols_dst) \

ggml/src/ggml-cuda/mmid.cu

@@ -0,0 +1,164 @@
+#include "common.cuh"
+#include "mmid.cuh"
+
+// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
+struct mm_ids_helper_store {
+    uint32_t data;
+
+    __device__ mm_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
+        data = (it & 0x003FFFFF) | (iex_used << 22);
+    }
+
+    __device__ uint32_t it() const {
+        return data & 0x003FFFFF;
+    }
+
+    __device__ uint32_t iex_used() const {
+        return data >> 22;
+    }
+};
+static_assert(sizeof(mm_ids_helper_store) == 4, "unexpected size for mm_ids_helper_store");
+
+// Helper function for mul_mat_id, converts ids to a more convenient format.
+// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
+// ids_dst describes the same mapping but for the dst tensor.
+// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
+template <int n_expert_used_template>
+__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
+    const int expert = blockIdx.x;
+
+    extern __shared__ char data_mm_ids_helper[];
+    mm_ids_helper_store * store = (mm_ids_helper_store *) data_mm_ids_helper;
+
+    int nex_prev   = 0; // Number of columns for experts with a lower index.
+    int it_compact = 0; // Running index for the compact slice of this expert.
+
+    if constexpr (n_expert_used_template == 0) {
+        // Generic implementation:
+        for (int it = 0; it < n_tokens; ++it) {
+            int iex_used = -1; // The index at which the expert is used, if any.
+            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
+                const int expert_used = ids[it*si1 + iex];
+                nex_prev += expert_used < expert;
+                if (expert_used == expert) {
+                    iex_used = iex;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact] = mm_ids_helper_store(it, iex_used);
+            }
+
+            if (warp_reduce_any<warp_size>(iex_used != -1)) {
+                it_compact++;
+            }
+        }
+    } else {
+        // Implementation optimized for specific numbers of experts used:
+        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
+        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
+        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
+            const int it = it0 + threadIdx.x / neu_padded;
+
+            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
+            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
+                ids[it*si1 + iex] : INT_MAX;
+            const int iex_used = expert_used == expert ? iex : -1;
+            nex_prev += expert_used < expert;
+
+            // Whether the threads at this token position have used the expert:
+            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
+
+            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
+            int it_compact_add_lower = 0;
+#pragma unroll
+            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
+                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
+                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
+                    it_compact_add_lower += tmp;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact + it_compact_add_lower] = mm_ids_helper_store(it, iex_used);
+            }
+
+            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
+            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
+        }
+    }
+    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
+
+    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
+        const mm_ids_helper_store store_it = store[itc];
+        const int it       = store_it.it();
+        const int iex_used = store_it.iex_used();
+        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
+        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    expert_bounds[expert] = nex_prev;
+
+    if (expert < static_cast<int>(gridDim.x) - 1) {
+        return;
+    }
+
+    expert_bounds[gridDim.x] = nex_prev + it_compact;
+}
+
+template <int n_expert_used_template>
+static void launch_mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mm_ids_helper_store");
+    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mm_ids_helper_store");
+
+    const int id = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[id].warp_size;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+    CUDA_SET_SHARED_MEMORY_LIMIT(mm_ids_helper<n_expert_used_template>, smpbo);
+
+    const dim3 num_blocks(n_experts, 1, 1);
+    const dim3 block_size(warp_size, 1, 1);
+    const size_t nbytes_shared = n_tokens*sizeof(mm_ids_helper_store);
+    GGML_ASSERT(nbytes_shared <= smpbo);
+    mm_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
+        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
+}
+
+void ggml_cuda_launch_mm_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    switch (n_expert_used) {
+        case  2:
+            launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  4:
+            launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  6:
+            launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case  8:
+            launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case 16:
+            launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        case 32:
+            launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+        default:
+            launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, stream);
+            break;
+    }
+}

ggml/src/ggml-cuda/mmid.cuh

@@ -0,0 +1,5 @@
+#pragma once
+
+void ggml_cuda_launch_mm_ids_helper(
+        const int32_t * ids, int32_t * ids_src1, int32_t * ids_dst, int32_t * expert_bounds,
+        int n_experts, int n_tokens, int n_expert_used, int nchannels_y, int si1, int sis1, cudaStream_t stream);

ggml/src/ggml-cuda/mmq.cu

@@ -1,141 +1,6 @@
 #include "mmq.cuh"
 #include "quantize.cuh"
-
-#include <vector>
-
-// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
-struct mmq_ids_helper_store {
-    uint32_t data;
-
-    __device__ mmq_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
-        data = (it & 0x003FFFFF) | (iex_used << 22);
-    }
-
-    __device__ uint32_t it() const {
-        return data & 0x003FFFFF;
-    }
-
-    __device__ uint32_t iex_used() const {
-        return data >> 22;
-    }
-};
-static_assert(sizeof(mmq_ids_helper_store) == 4, "unexpected size for mmq_ids_helper_store");
-
-// Helper function for mul_mat_id, converts ids to a more convenient format.
-// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
-// ids_dst describes the same mapping but for the dst tensor.
-// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
-template <int n_expert_used_template>
-__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
-static __global__ void mmq_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
-    const int expert = blockIdx.x;
-
-    extern __shared__ char data_mmq_ids_helper[];
-    mmq_ids_helper_store * store = (mmq_ids_helper_store *) data_mmq_ids_helper;
-
-    int nex_prev   = 0; // Number of columns for experts with a lower index.
-    int it_compact = 0; // Running index for the compact slice of this expert.
-
-    if constexpr (n_expert_used_template == 0) {
-        // Generic implementation:
-        for (int it = 0; it < n_tokens; ++it) {
-            int iex_used = -1; // The index at which the expert is used, if any.
-            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
-                const int expert_used = ids[it*si1 + iex];
-                nex_prev += expert_used < expert;
-                if (expert_used == expert) {
-                    iex_used = iex;
-                }
-            }
-
-            if (iex_used != -1) {
-                store[it_compact] = mmq_ids_helper_store(it, iex_used);
-            }
-
-            if (warp_reduce_any<warp_size>(iex_used != -1)) {
-                it_compact++;
-            }
-        }
-    } else {
-        // Implementation optimized for specific numbers of experts used:
-        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
-        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
-        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
-            const int it = it0 + threadIdx.x / neu_padded;
-
-            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
-            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
-                ids[it*si1 + iex] : INT_MAX;
-            const int iex_used = expert_used == expert ? iex : -1;
-            nex_prev += expert_used < expert;
-
-            // Whether the threads at this token position have used the expert:
-            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
-
-            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
-            int it_compact_add_lower = 0;
-#pragma unroll
-            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
-                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
-                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
-                    it_compact_add_lower += tmp;
-                }
-            }
-
-            if (iex_used != -1) {
-                store[it_compact + it_compact_add_lower] = mmq_ids_helper_store(it, iex_used);
-            }
-
-            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
-            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
-        }
-    }
-    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
-
-    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
-        const mmq_ids_helper_store store_it = store[itc];
-        const int it       = store_it.it();
-        const int iex_used = store_it.iex_used();
-        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
-        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
-    }
-
-    if (threadIdx.x != 0) {
-        return;
-    }
-
-    expert_bounds[expert] = nex_prev;
-
-    if (expert < static_cast<int>(gridDim.x) - 1) {
-        return;
-    }
-
-    expert_bounds[gridDim.x] = nex_prev + it_compact;
-}
-
-template <int n_expert_used_template>
-static void launch_mmq_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
-    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mmq_ids_helper_store");
-    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mmq_ids_helper_store");
-
-    const int id = ggml_cuda_get_device();
-    const int warp_size = ggml_cuda_info().devices[id].warp_size;
-    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
-    CUDA_SET_SHARED_MEMORY_LIMIT(mmq_ids_helper<n_expert_used_template>, smpbo);
-
-    const dim3 num_blocks(n_experts, 1, 1);
-    const dim3 block_size(warp_size, 1, 1);
-    const size_t nbytes_shared = n_tokens*sizeof(mmq_ids_helper_store);
-    GGML_ASSERT(nbytes_shared <= smpbo);
-    mmq_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
-        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
-}
+#include "mmid.cuh"
 
 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) {
@@ -293,36 +158,8 @@ void ggml_cuda_mul_mat_q(
         const int si1  = ids->nb[1] / ggml_element_size(ids);
         const int sis1 = nb12 / nb11;
 
-        switch (n_expert_used) {
-            case  2:
-                launch_mmq_ids_helper< 2> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  4:
-                launch_mmq_ids_helper< 4> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  6:
-                launch_mmq_ids_helper< 6> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case  8:
-                launch_mmq_ids_helper< 8> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case 16:
-                launch_mmq_ids_helper<16> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            case 32:
-                launch_mmq_ids_helper<32> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-            default:
-                launch_mmq_ids_helper< 0> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
-                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
-                break;
-        }
+        ggml_cuda_launch_mm_ids_helper((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+            ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
         CUDA_CHECK(cudaGetLastError());
     }
 

ggml/src/ggml-cuda/mmq.cuh

@@ -3494,7 +3494,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
     const int col_diff = col_high - col_low;
 
     for (int j = threadIdx.y*warp_size + threadIdx.x; j < mmq_x; j += nwarps*warp_size) {
-        ids_dst_shared[j] = ids_dst[col_low + j];
+        ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
     }
     __syncthreads();
 

ggml/src/ggml-cuda/mmvf.cu

@@ -1,20 +1,21 @@
 #include "ggml.h"
 #include "common.cuh"
-#include "convert.cuh"
+#include "unary.cuh"
 #include "mmvf.cuh"
+#include "convert.cuh"
 
-template <typename T, typename type_acc, int ncols_dst, int block_size>
+template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false>
 static __global__ void mul_mat_vec_f(
-        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
-        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
     const int row         = blockIdx.x;
     const int channel_dst = blockIdx.y;
-    const int channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_x   = ids ? ids[channel_dst]          : fastdiv((uint32_t) channel_dst, channel_ratio);
     const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
     const int sample_dst  = blockIdx.z;
-    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_x    = fastdiv((uint32_t) sample_dst, sample_ratio);
     const int sample_y    = sample_dst;
     const int tid         = threadIdx.x;
 
@@ -24,58 +25,164 @@ static __global__ void mul_mat_vec_f(
     y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
     dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
 
+    bool use_gate = false;
+    bool use_bias = false;
+    bool use_gate_bias = false;
+    ggml_glu_op glu_op = ggml_glu_op::GGML_GLU_OP_SWIGLU;
+    const T * gate_x = nullptr;
+    const float * x_bias = nullptr;
+    const float * gate_bias = nullptr;
+
+    if constexpr (has_fusion) {
+        use_gate = fusion.gate != nullptr;
+        use_bias = fusion.x_bias != nullptr;
+        use_gate_bias = fusion.gate_bias != nullptr;
+        glu_op = fusion.glu_op;
+
+        if (use_gate) {
+            gate_x = static_cast<const T *>(fusion.gate);
+        }
+        if (use_bias) {
+            x_bias = static_cast<const float *>(fusion.x_bias);
+        }
+        if (use_gate_bias) {
+            gate_bias = static_cast<const float *>(fusion.gate_bias);
+            use_gate_bias = use_gate;
+        } else {
+            use_gate_bias = false;
+        }
+    }
+
+    if (use_gate) {
+        gate_x += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    }
+    if constexpr (has_fusion) {
+        const int channel_bias = ids ? channel_x : channel_dst;
+        if (use_bias) {
+            x_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
+        }
+        if (use_gate_bias) {
+            gate_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
+        }
+    }
+
     const float2 * y2 = (const float2 *) y;
 
     extern __shared__ char data_mmv[];
     float * buf_iw = (float *) data_mmv;
+    float * buf_iw_gate = nullptr;
+    if constexpr (has_fusion) {
+        buf_iw_gate = (float *) (data_mmv + warp_size*sizeof(float));
+    }
 
     if (block_size > warp_size) {
         if (tid < warp_size) {
             buf_iw[tid] = 0.0f;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    buf_iw_gate[tid] = 0.0f;
+                }
+            }
         }
         __syncthreads();
     }
 
     float sumf[ncols_dst] = {0.0f};
+    float sumf_gate[ncols_dst];
+    if constexpr (has_fusion) {
+#pragma unroll
+        for (int j = 0; j < ncols_dst; ++j) {
+            sumf_gate[j] = 0.0f;
+        }
+    }
 
     if constexpr (std::is_same_v<T, float>) {
         const float2 * x2 = (const float2 *) x;
+        const float2 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const float2 *) gate_x;
+            }
+        }
 
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const float2 tmpx = x2[col2];
+            float2 tmpx_gate = make_float2(0.0f, 0.0f);
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += tmpx.x*tmpy.x;
-                sumf[j] += tmpx.y*tmpy.y;
+                ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
+                ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                    }
+                }
             }
         }
     } else if constexpr (std::is_same_v<T, half>) {
         const half2 * x2 = (const half2 *) x;
+        const half2 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const half2 *) gate_x;
+            }
+        }
 
         if (std::is_same_v<type_acc, float>) {
             for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
-
+                float2 tmpx_gate = make_float2(0.0f, 0.0f);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmpx_gate = __half22float2(gate_x2[col2]);
+                    }
+                }
 #pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
                     const float2 tmpy = y2[j*stride_col_y2 + col2];
-                    sumf[j] += tmpx.x * tmpy.x;
-                    sumf[j] += tmpx.y * tmpy.y;
+                    ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
+                    ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                            ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                        }
+                    }
                 }
             }
         } else {
 #ifdef FP16_AVAILABLE
             half2 sumh2[ncols_dst] = {{0.0f, 0.0f}};
+            half2 sumh2_gate[ncols_dst] = {{0.0f, 0.0f}};
 
             for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const half2 tmpx = x2[col2];
-
+                half2 tmpx_gate = make_half2(0.0f, 0.0f);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmpx_gate = gate_x2[col2];
+                    }
+                }
 #pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
                     const float2 tmpy = y2[j*stride_col_y2 + col2];
                     sumh2[j] += tmpx * make_half2(tmpy.x, tmpy.y);
+
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            sumh2_gate[j] += tmpx_gate * make_half2(tmpy.x, tmpy.y);
+                        }
+                    }
                 }
             }
 
@@ -83,21 +190,86 @@ static __global__ void mul_mat_vec_f(
             for (int j = 0; j < ncols_dst; ++j) {
                 sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]);
             }
+
+            if constexpr (has_fusion) {
+                if (use_gate) {
+#pragma unroll
+                    for (int j = 0; j < ncols_dst; ++j) {
+                        sumf_gate[j] = __low2float(sumh2_gate[j]) + __high2float(sumh2_gate[j]);
+                    }
+                }
+            }
 #else
             NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
         }
     } else if constexpr (std::is_same_v<T, nv_bfloat16>) {
+//TODO: add support for ggml_cuda_mad for hip_bfloat162
+#if defined(GGML_USE_HIP)
         const int * x2 = (const int *) x;
+        const int * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const int *) gate_x;
+            }
+        }
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const int tmpx = x2[col2];
+            int tmpx_gate = 0;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+                const float tmpx0 = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]);
+                const float tmpx1 = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]);
+                ggml_cuda_mad(sumf[j], tmpx0, tmpy.x);
+                ggml_cuda_mad(sumf[j], tmpx1, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        const float tmpx0_gate = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx_gate)[0]);
+                        const float tmpx1_gate = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx_gate)[1]);
+                        ggml_cuda_mad(sumf_gate[j], tmpx0_gate, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx1_gate, tmpy.y);
+                    }
+                }
             }
         }
+#else
+        const nv_bfloat162 * x2 = (const nv_bfloat162 *) x;
+        const nv_bfloat162 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const nv_bfloat162 *) gate_x;
+            }
+        }
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const nv_bfloat162 tmpx = x2[col2];
+            nv_bfloat162 tmpx_gate;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
+                ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                    }
+                }
+            }
+        }
+#endif
     } else {
         static_assert(std::is_same_v<T, void>, "unsupported type");
     }
@@ -106,13 +278,31 @@ static __global__ void mul_mat_vec_f(
     for (int j = 0; j < ncols_dst; ++j) {
         sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
 
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                sumf_gate[j] = warp_reduce_sum<warp_size>(sumf_gate[j]);
+            }
+        }
+
         if (block_size > warp_size) {
             buf_iw[tid/warp_size] = sumf[j];
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    buf_iw_gate[tid/warp_size] = sumf_gate[j];
+                }
+            }
             __syncthreads();
             if (tid < warp_size) {
                 sumf[j] = buf_iw[tid];
                 sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        sumf_gate[j] = buf_iw_gate[tid];
+                        sumf_gate[j] = warp_reduce_sum<warp_size>(sumf_gate[j]);
+                    }
+                }
             }
+
             if (j < ncols_dst) {
                 __syncthreads();
             }
@@ -123,12 +313,74 @@ static __global__ void mul_mat_vec_f(
         return;
     }
 
-    dst[tid*stride_col_dst + row] = sumf[tid];
+    float value = sumf[tid];
+
+    if constexpr (has_fusion) {
+        if (use_bias) {
+            value += x_bias[tid*stride_col_dst + row];
+        }
+
+        if (use_gate) {
+            float gate_value = sumf_gate[tid];
+            if (use_gate_bias) {
+                gate_value += gate_bias[tid*stride_col_dst + row];
+            }
+            switch (glu_op) {
+                case GGML_GLU_OP_SWIGLU:
+                    value *= ggml_cuda_op_silu_single(gate_value);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    value *= ggml_cuda_op_gelu_single(gate_value);
+                    break;
+                case GGML_GLU_OP_SWIGLU_OAI: {
+                    value = ggml_cuda_op_swiglu_oai_single(gate_value, value);
+                    break;
+                }
+                default:
+                    break;
+            }
+        }
+    }
+
+    dst[tid*stride_col_dst + row] = value;
+
+    if constexpr (!has_fusion) {
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, glu_op, gate_x, x_bias, gate_bias, sumf_gate);
+    }
+}
+
+template<typename T, typename type_acc, int ncols_dst, int block_size>
+static void mul_mat_vec_f_switch_fusion(
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
+        const int64_t ncols, const int64_t nrows,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const cudaStream_t stream) {
+
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    if constexpr (ncols_dst == 1) {
+        if (has_fusion) {
+            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            return;
+       }
+    }
+
+    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
+
+    mul_mat_vec_f<T, type_acc, ncols_dst, block_size><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+
 }
 
 template <typename T, typename type_acc, int ncols_dst>
-static void launch_mul_mat_vec_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+void launch_mul_mat_vec_f_cuda(
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
@@ -140,8 +392,8 @@ static void launch_mul_mat_vec_f_cuda(
     GGML_ASSERT(stride_col_y % 2 == 0);
     GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
     GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
-    const int64_t channel_ratio = nchannels_dst / nchannels_x;
-    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
+    const uint3 channel_ratio_fd = ids ? make_uint3(0, 0, 0) : init_fastdiv_values(nchannels_dst / nchannels_x);
+    const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
     const int device = ggml_cuda_get_device();
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
@@ -160,57 +412,59 @@ static void launch_mul_mat_vec_f_cuda(
         }
     }
 
-    const int nbytes_shared = warp_size*sizeof(float);
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+
+    const int nbytes_shared = warp_size*sizeof(float) + (has_fusion ? warp_size*sizeof(float) : 0);
     const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case   64: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case   96: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  96><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  128: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  160: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 160><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  192: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 192><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  224: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 224><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  256: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
-                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -220,7 +474,7 @@ static void launch_mul_mat_vec_f_cuda(
 
 template <typename T, typename type_acc>
 static void mul_mat_vec_f_cuda_switch_ncols_dst(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
@@ -230,49 +484,49 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
     switch (ncols_dst) {
         case 1:
             launch_mul_mat_vec_f_cuda<T, type_acc, 1>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 2:
             launch_mul_mat_vec_f_cuda<T, type_acc, 2>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 3:
             launch_mul_mat_vec_f_cuda<T, type_acc, 3>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 4:
             launch_mul_mat_vec_f_cuda<T, type_acc, 4>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 5:
             launch_mul_mat_vec_f_cuda<T, type_acc, 5>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 6:
             launch_mul_mat_vec_f_cuda<T, type_acc, 6>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 7:
             launch_mul_mat_vec_f_cuda<T, type_acc, 7>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
         case 8:
             launch_mul_mat_vec_f_cuda<T, type_acc, 8>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, 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, stream);
             break;
@@ -284,29 +538,31 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
 
 template<typename T>
 static void mul_mat_vec_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
         const int64_t stride_row, const int64_t stride_col_y, const int stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         enum ggml_prec prec, cudaStream_t stream) {
+
     if constexpr(std::is_same_v<T, half>) {
         if (prec == GGML_PREC_DEFAULT) {
             mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
-                (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, 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, stream);
+                (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, 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, stream);
             return;
         }
     }
     mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
-        (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, 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, stream);
+        (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, 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, stream);
 }
 
-void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+    const ggml_cuda_mm_fusion_args_host * fusion) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
     GGML_ASSERT(         dst->type == GGML_TYPE_F32);
@@ -332,6 +588,30 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
+
+    if (fusion) {
+        GGML_ASSERT( !ids || dst->ne[2] == 1);
+        GGML_ASSERT(  ids || dst->ne[1] == 1);
+        if (fusion->x_bias) {
+            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
+            fusion_local.x_bias = fusion->x_bias->data;
+        }
+        if (fusion->gate) {
+            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            fusion_local.gate = fusion->gate->data;
+        }
+        if (fusion->gate_bias) {
+            GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
+            fusion_local.gate_bias = fusion->gate_bias->data;
+        }
+        fusion_local.glu_op = fusion->glu_op;
+    }
+
     const int64_t s01 = src0->nb[1] / ts_src0;
     const int64_t s11 = src1->nb[1] / ts_src1;
     const int64_t s1  =  dst->nb[1] / ts_dst;
@@ -354,19 +634,19 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
@@ -393,7 +673,6 @@ void ggml_cuda_op_mul_mat_vec_f(
     const int cc = ggml_cuda_info().devices[id].cc;
     const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
 
-
     // ggml_cuda_op provides single, contiguous matrices
     const int64_t stride_row         = ne00;
     const int64_t stride_col_y       = ne10;
@@ -410,22 +689,23 @@ void ggml_cuda_op_mul_mat_vec_f(
     const int64_t stride_sample_y    = 0;
     const int64_t stride_sample_dst  = 0;
 
+    ggml_cuda_mm_fusion_args_device empty{};
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, 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, prec, stream);
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, 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, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, 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, prec, stream);
         } break;
@@ -436,10 +716,23 @@ void ggml_cuda_op_mul_mat_vec_f(
     GGML_UNUSED_VARS(ctx, src1, dst, src1_ddq_i, src1_ncols, src1_padded_row_size);
 }
 
-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11) {
     if (src0_ne[0] % 2 != 0) {
         return false;
     }
+
+    const size_t ts = ggml_type_size(type);
+    if (src0_nb[0] != ts) {
+        return false;
+    }
+
+    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
+    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0_nb[i] % (2*ts) != 0) {
+            return false;
+        }
+    }
+
     switch (type) {
         case GGML_TYPE_F32:
             if (GGML_CUDA_CC_IS_NVIDIA(cc)) {

ggml/src/ggml-cuda/mmvf.cuh

@@ -1,6 +1,7 @@
 #include "common.cuh"
 
-void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+    const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 
 void ggml_cuda_op_mul_mat_vec_f(
     ggml_backend_cuda_context & ctx,
@@ -8,4 +9,4 @@ void ggml_cuda_op_mul_mat_vec_f(
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
 
-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11);
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11);

ggml/src/ggml-cuda/mmvq.cu

@@ -1,5 +1,6 @@
 #include "mmvq.cuh"
 #include "quantize.cuh"
+#include "unary.cuh"
 #include "vecdotq.cuh"
 
 #include <cstdint>
@@ -82,7 +83,7 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
-static constexpr __host__ __device__ int calc_nwarps(int ncols_dst,  mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
             case 1:
@@ -136,11 +137,11 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst>
 // tell the compiler to use as many registers as it wants, see nwarps definition below
+template <ggml_type type, int ncols_dst, bool has_fusion>
 __launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
@@ -169,8 +170,56 @@ static __global__ void mul_mat_vec_q(
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
 
+    bool use_gate = false;
+    bool use_bias = false;
+    bool use_gate_bias = false;
+    const void * vgate = nullptr;
+    const float * x_bias = nullptr;
+    const float * gate_bias = nullptr;
+    ggml_glu_op active_glu;
+
+    if constexpr (has_fusion) {
+        use_gate      = fusion.gate      != nullptr;
+        use_bias      = fusion.x_bias    != nullptr;
+        use_gate_bias = fusion.gate_bias != nullptr && use_gate;
+        vgate         = fusion.gate;
+        x_bias        = (const float *) fusion.x_bias;
+        gate_bias     = (const float *) fusion.gate_bias;
+        active_glu    = fusion.glu_op;
+    }
+
+    const uint32_t channel_bias = ids ? channel_x : channel_dst;
+
+    float x_biases[ncols_dst]    = { 0.0f };
+    float gate_biases[ncols_dst] = { 0.0f };
+    if constexpr (has_fusion) {
+        if (use_bias) {
+            x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            // 1. Hide latency by prefetching bias and gate here
+            // 2. load only on threads that won't die after partial sum calculation
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    x_biases[j] = x_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
+        }
+        if (use_gate_bias) {
+            gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    gate_biases[j] = gate_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
+        }
+    }
+
     // partial sum for each thread
     float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
+    float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
@@ -187,17 +236,35 @@ static __global__ void mul_mat_vec_q(
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp[j][i] += vec_dot_q_cuda(
                     vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_gate[j][i] += vec_dot_q_cuda(
+                            vgate, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
+                    }
+                }
             }
         }
     }
 
     __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
+    __shared__ float tmp_shared_gate[(has_fusion && (nwarps-1 > 0)) ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
+    if constexpr (!has_fusion) {
+        (void) tmp_shared_gate;
+    } else if (!use_gate) {
+        (void) tmp_shared_gate;
+    }
+
     if (threadIdx.y > 0) {
 #pragma unroll
         for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_shared_gate[threadIdx.y-1][j][i][threadIdx.x] = tmp_gate[j][i];
+                    }
+                }
             }
         }
     }
@@ -216,14 +283,55 @@ static __global__ void mul_mat_vec_q(
 #pragma unroll
             for (int l = 0; l < nwarps-1; ++l) {
                 tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_gate[j][i] += tmp_shared_gate[l][j][i][threadIdx.x];
+                    }
+                }
             }
             tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmp_gate[j][i] = warp_reduce_sum<warp_size>(tmp_gate[j][i]);
+                }
+            }
         }
 
         if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
-            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
+            float result = tmp[j][threadIdx.x];
+            if constexpr (has_fusion) {
+                if (use_bias) {
+                    result += x_biases[j];
+                }
+                if (use_gate) {
+                    float gate_value = tmp_gate[j][threadIdx.x];
+                    if (use_gate_bias) {
+                        gate_value += gate_biases[j];
+                    }
+                    switch (active_glu) {
+                        case GGML_GLU_OP_SWIGLU:
+                            result *= ggml_cuda_op_silu_single(gate_value);
+                            break;
+                        case GGML_GLU_OP_GEGLU:
+                            result *= ggml_cuda_op_gelu_single(gate_value);
+                            break;
+                        case GGML_GLU_OP_SWIGLU_OAI: {
+                            result = ggml_cuda_op_swiglu_oai_single(gate_value, result);
+                            break;
+                        }
+                        default:
+                            result = result * gate_value;
+                            break;
+                    }
+                }
+            }
+            dst[j*stride_col_dst + threadIdx.x] = result;
         }
     }
+
+    if constexpr (!has_fusion) {
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, active_glu, gate_bias, x_bias, tmp_gate);
+    }
 }
 
 static std::pair<dim3, dim3> calc_launch_params(
@@ -235,9 +343,37 @@ static std::pair<dim3, dim3> calc_launch_params(
     return {block_nums, block_dims};
 }
 
+template<ggml_type type, int c_ncols_dst>
+static void mul_mat_vec_q_switch_fusion(
+        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
+        const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
+        const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    if constexpr (c_ncols_dst == 1) {
+        if (has_fusion) {
+            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            return;
+        }
+    }
+
+    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
+
+    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_switch_ncols_dst(
-        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int ncols_x, const int nrows_x, const int ncols_dst,
         const int stride_row_x, const int stride_col_y, const int stride_col_dst,
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
@@ -256,80 +392,83 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+
     GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 3: {
             constexpr int c_ncols_dst = 3;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 4: {
             constexpr int c_ncols_dst = 4;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 5: {
             constexpr int c_ncols_dst = 5;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 6: {
             constexpr int c_ncols_dst = 6;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 7: {
             constexpr int c_ncols_dst = 7;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 8: {
             constexpr int c_ncols_dst = 8;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
-}
 
+    GGML_UNUSED(has_fusion);
+}
 static void mul_mat_vec_q_switch_type(
-        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
+        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int ncols_x, const int nrows_x, const int ncols_dst,
         const int stride_row_x, const int stride_col_y, const int stride_col_dst,
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
@@ -339,143 +478,123 @@ static void mul_mat_vec_q_switch_type(
     switch (type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q4_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q8_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_MXFP4:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q3_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q4_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ1_M:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ3_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (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,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -484,7 +603,8 @@ static void mul_mat_vec_q_switch_type(
 }
 
 void ggml_cuda_mul_mat_vec_q(
-        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+        const ggml_cuda_mm_fusion_args_host * fusion) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
     GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
@@ -508,6 +628,31 @@ void ggml_cuda_mul_mat_vec_q(
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
+
+    if (fusion) {
+        GGML_ASSERT( !ids || dst->ne[2] == 1);
+        GGML_ASSERT(  ids || dst->ne[1] == 1);
+
+        if (fusion->x_bias) {
+            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
+            fusion_local.x_bias = fusion->x_bias->data;
+        }
+        if (fusion->gate) {
+            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            fusion_local.gate = fusion->gate->data;
+        }
+        if (fusion->gate_bias) {
+            GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
+            fusion_local.gate_bias = fusion->gate_bias->data;
+        }
+        fusion_local.glu_op = fusion->glu_op;
+    }
+
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
         const size_t size_data  = ggml_nbytes(src0);
@@ -549,10 +694,10 @@ void ggml_cuda_mul_mat_vec_q(
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
     mul_mat_vec_q_switch_type(
-        src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
+        src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
         ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
         ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,                 stream);
+        ne03,              ne3,           s03, s13,              s3,               stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -578,8 +723,9 @@ void ggml_cuda_op_mul_mat_vec_q(
     const int stride_row_x = ne00 / ggml_blck_size(src0->type);
     const int stride_col_y = src1_padded_row_size / QK8_1;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
     mul_mat_vec_q_switch_type(
-        src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
+        src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
     GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);