llama: introduce support for model-embedded sampling parameters (#17120)

This commit adds a check to skip the output reordering logic when
n_outputs == 1. With a single output token, the data is trivially
sorted and the reordering code is currently doing unnecessary work
(resetting and rebuilding output_ids to the same values).

The motivation for this change is improved code clarity and avoiding
confusion when debugging. While the performance impact is probably
negligible, this unnecessary work happens on every decode call in
llama-server when processing batches with single-token outputs.

.devops/cann.Dockerfile

@@ -3,7 +3,8 @@
 # ==============================================================================
 
 # Define the CANN base image for easier version updates later
-ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.1.rc1-910b-openeuler22.03-py3.10
+ARG CHIP_TYPE=910b
+ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.3.rc1.alpha001-${CHIP_TYPE}-openeuler22.03-py3.11
 
 # ==============================================================================
 # BUILD STAGE
@@ -11,9 +12,6 @@ ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.1.rc1-910b-openeuler22.03-py3.10
 # ==============================================================================
 FROM ${CANN_BASE_IMAGE} AS build
 
-# Define the Ascend chip model for compilation. Default is Ascend910B3
-ARG ASCEND_SOC_TYPE=Ascend910B3
-
 # -- Install build dependencies --
 RUN yum install -y gcc g++ cmake make git libcurl-devel python3 python3-pip && \
     yum clean all && \
@@ -36,13 +34,14 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 # For brevity, only core variables are listed here. You can paste the original ENV list here.
 
 # -- Build llama.cpp --
-# Use the passed ASCEND_SOC_TYPE argument and add general build options
+# Use the passed CHIP_TYPE argument and add general build options
+ARG CHIP_TYPE
 RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh --force \
     && \
     cmake -B build \
         -DGGML_CANN=ON \
         -DCMAKE_BUILD_TYPE=Release \
-        -DSOC_TYPE=${ASCEND_SOC_TYPE} \
+        -DSOC_TYPE=ascend${CHIP_TYPE} \
         . && \
     cmake --build build --config Release -j$(nproc)
 

.devops/vulkan.Dockerfile

@@ -1,9 +1,7 @@
-ARG UBUNTU_VERSION=25.10
+ARG UBUNTU_VERSION=26.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
-
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget xz-utils
 

.github/copilot-instructions.md

@@ -9,7 +9,7 @@ llama.cpp is a large-scale C/C++ project for efficient LLM (Large Language Model
 - **Size**: ~200k+ lines of code across 1000+ files
 - **Architecture**: Modular design with main library (`libllama`) and 40+ executable tools/examples
 - **Core dependency**: ggml tensor library (vendored in `ggml/` directory)
-- **Backends supported**: CPU (AVX/NEON optimized), CUDA, Metal, Vulkan, SYCL, ROCm, MUSA
+- **Backends supported**: CPU (AVX/NEON/RVV optimized), CUDA, Metal, Vulkan, SYCL, ROCm, MUSA
 - **License**: MIT
 
 ## Build Instructions

.github/workflows/build-amd.yml

@@ -1,52 +0,0 @@
-name: CI (AMD)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-amd.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh',
-      '**/*.comp'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  ggml-ci-x64-amd-vulkan:
-    runs-on: [self-hosted, Linux, X64, AMD]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
-
-  ggml-ci-x64-amd-rocm:
-    runs-on: [self-hosted, Linux, X64, AMD]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          amd-smi static
-          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

.github/workflows/build.yml

@@ -69,13 +69,6 @@ jobs:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
       - name: Build
         id: cmake_build
         run: |
@@ -83,6 +76,8 @@ jobs:
           cmake -B build \
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=OFF \
             -DGGML_METAL_SHADER_DEBUG=ON \
@@ -110,13 +105,6 @@ jobs:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
       - name: Build
         id: cmake_build
         run: |
@@ -126,6 +114,8 @@ jobs:
           cmake -B build \
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
@@ -151,13 +141,6 @@ jobs:
           key: macOS-latest-cmake-arm64-webgpu
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
       - name: Dawn Dependency
         id: dawn-depends
         run: |
@@ -217,7 +200,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install -y --no-install-recommends \
             python3 python3-pip python3-dev \
-            libjpeg-dev build-essential libcurl4-openssl-dev \
+            libjpeg-dev build-essential libssl-dev \
             git-lfs
 
       - name: Python Dependencies
@@ -238,6 +221,8 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(nproc)
@@ -294,13 +279,15 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential libssl-dev
 
       - name: Build
         id: cmake_build
         if: ${{ matrix.sanitizer != 'THREAD' }}
         run: |
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
@@ -311,6 +298,8 @@ jobs:
         if: ${{ matrix.sanitizer == 'THREAD' }}
         run: |
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
@@ -335,7 +324,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential libssl-dev
 
       - name: Build
         id: cmake_build
@@ -343,6 +332,8 @@ jobs:
           mkdir build
           cd build
           cmake .. \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_LLGUIDANCE=ON
           cmake --build . --config Release -j $(nproc)
@@ -373,12 +364,14 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential libssl-dev
 
       - name: Build
         id: cmake_build
         run: |
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -405,12 +398,14 @@ jobs:
       - name: Dependencies
         id: depends
         run: |
-          sudo apt-get install -y glslc libvulkan-dev libcurl4-openssl-dev
+          sudo apt-get install -y glslc libvulkan-dev libssl-dev
 
       - name: Configure
         id: cmake_configure
         run: |
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DCMAKE_BUILD_TYPE=RelWithDebInfo \
             -DGGML_BACKEND_DL=ON \
             -DGGML_CPU_ALL_VARIANTS=ON \
@@ -440,7 +435,7 @@ jobs:
         run: |
           sudo add-apt-repository -y ppa:kisak/kisak-mesa
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -466,6 +461,8 @@ jobs:
         run: |
           source ./vulkan_sdk/setup-env.sh
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DGGML_VULKAN=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -497,7 +494,7 @@ jobs:
         run: |
           sudo add-apt-repository -y ppa:kisak/kisak-mesa
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -537,7 +534,10 @@ jobs:
         id: cmake_build
         run: |
           export Dawn_DIR=dawn/lib64/cmake/Dawn
-          cmake -B build -DGGML_WEBGPU=ON
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
+            -DGGML_WEBGPU=ON
           cmake --build build --config Release -j $(nproc)
 
       - name: Test
@@ -560,7 +560,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev rocwmma-dev
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev
 
       - name: ccache
         uses: ggml-org/[email protected]
@@ -572,6 +572,8 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build -S . \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
             -DGGML_HIP_ROCWMMA_FATTN=ON \
             -DGGML_HIP=ON
@@ -590,7 +592,7 @@ jobs:
         id: depends
         run: |
           apt-get update
-          apt-get install -y build-essential git cmake libcurl4-openssl-dev
+          apt-get install -y build-essential git cmake libssl-dev
 
       - name: ccache
         uses: ggml-org/[email protected]
@@ -602,6 +604,8 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build -S . \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DGGML_MUSA=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -626,7 +630,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
 
       - name: install oneAPI MKL library
         shell: bash
@@ -648,6 +652,8 @@ jobs:
         run: |
           source /opt/intel/oneapi/setvars.sh
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DGGML_SYCL=ON \
             -DCMAKE_C_COMPILER=icx \
             -DCMAKE_CXX_COMPILER=icpx
@@ -674,7 +680,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
 
       - name: install oneAPI MKL library
         shell: bash
@@ -696,6 +702,8 @@ jobs:
         run: |
           source /opt/intel/oneapi/setvars.sh
           cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DGGML_SYCL=ON \
             -DCMAKE_C_COMPILER=icx \
             -DCMAKE_CXX_COMPILER=icpx \
@@ -722,12 +730,6 @@ jobs:
           key: macOS-latest-cmake-ios
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
       - name: Build
         id: cmake_build
         run: |
@@ -759,12 +761,6 @@ jobs:
           key: macOS-latest-cmake-tvos
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
       - name: Build
         id: cmake_build
         run: |
@@ -790,12 +786,6 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
       - name: Build
         id: cmake_build
         run: |
@@ -838,12 +828,6 @@ jobs:
           name: llama-xcframework
           path: build-apple/llama.xcframework/
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
       - name: Build llama.cpp with CMake
         id: cmake_build
         run: |
@@ -995,21 +979,12 @@ jobs:
             -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
           cmake --build build-arm64-release --target install --config release
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           cmake -S . -B build ${{ matrix.defines }} `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
+            -DLLAMA_CURL=OFF -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
@@ -1053,7 +1028,7 @@ jobs:
             DEBIAN_FRONTEND: noninteractive
           run: |
               apt update
-              apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
+              apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
 
         - name: ccache
           uses: ggml-org/[email protected]
@@ -1064,10 +1039,12 @@ jobs:
         - name: Build with CMake
           run: |
             cmake -S . -B build -G Ninja \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
+              -DLLAMA_FATAL_WARNINGS=ON \
               -DCMAKE_BUILD_TYPE=Release \
               -DCMAKE_CUDA_ARCHITECTURES=89-real \
               -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined \
-              -DLLAMA_FATAL_WARNINGS=ON \
               -DGGML_NATIVE=OFF \
               -DGGML_CUDA=ON
             cmake --build build
@@ -1101,25 +1078,20 @@ jobs:
         run: |
           choco install ninja
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
         shell: cmd
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
+            -DLLAMA_CURL=OFF ^
+            -DLLAMA_BUILD_BORINGSSL=ON ^
             -DGGML_NATIVE=OFF ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_CPU_ALL_VARIANTS=ON ^
             -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include"
+            -DGGML_RPC=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
@@ -1151,7 +1123,7 @@ jobs:
         run:  |
           scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
-      # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
+      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
 
       - name: Build
         id: cmake_build
@@ -1208,14 +1180,8 @@ jobs:
           key: ${{ github.job }}
           evict-old-files: 1d
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
@@ -1224,11 +1190,12 @@ jobs:
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
             -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
             -DCMAKE_BUILD_TYPE=Release `
+            -DLLAMA_CURL=OFF `
+            -DLLAMA_BUILD_BORINGSSL=ON `
             -DROCM_DIR="${env:HIP_PATH}" `
             -DGGML_HIP=ON `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_RPC=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
+            -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
   ios-xcode-build:
@@ -1390,14 +1357,10 @@ jobs:
     strategy:
       matrix:
         arch: [x86, aarch64]
-        cann:
-          - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
-        device:
-          - 'ascend910b3'
-        build:
-          - 'Release'
+        chip_type: ['910b', '310p']
+        build: ['Release']
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-    container: ascendai/cann:${{ matrix.cann }}
+    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.2.rc1-310p-openeuler22.03-py3.11' }}
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -1414,7 +1377,7 @@ jobs:
           cmake -S . -B build \
               -DCMAKE_BUILD_TYPE=${{ matrix.build }} \
               -DGGML_CANN=on \
-              -DSOC_TYPE=${{ matrix.device }}
+              -DSOC_TYPE=ascend${{ matrix.chip_type }}
           cmake --build build -j $(nproc)
 
 # TODO: simplify the following workflows using a matrix
@@ -1599,6 +1562,34 @@ jobs:
         run: |
           bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
+  ggml-ci-x64-amd-vulkan:
+    runs-on: [self-hosted, Linux, X64, AMD]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-x64-amd-rocm:
+    runs-on: [self-hosted, Linux, X64, AMD]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          amd-smi static
+          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
   ggml-ci-mac-metal:
     runs-on: [self-hosted, macOS, ARM64]
 

.github/workflows/release.yml

@@ -693,6 +693,51 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
           name: llama-${{ steps.tag.outputs.name }}-xcframework
 
+  openEuler-cann:
+    strategy:
+      matrix:
+        arch: [x86, aarch64]
+        chip_type: ['910b', '310p']
+        build: ['Release']
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.2.rc1-310p-openeuler22.03-py3.11' }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Dependencies
+        run: |
+          yum update -y
+          yum install -y git gcc gcc-c++ make cmake libcurl-devel
+          git config --global --add safe.directory "$GITHUB_WORKSPACE"
+
+      - name: Build
+        run: |
+          export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+
+          cmake -S . -B build \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build }} \
+              -DGGML_CANN=on \
+              -DSOC_TYPE=ascend${{ matrix.chip_type }}
+          cmake --build build -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.zip ./build/bin/*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.zip
+          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.zip
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -714,6 +759,7 @@ jobs:
       - macOS-arm64
       - macOS-x64
       - ios-xcode-build
+      - openEuler-cann
 
     steps:
       - name: Clone

.github/workflows/server.yml

@@ -56,7 +56,7 @@ jobs:
             curl \
             wget \
             language-pack-en \
-            libcurl4-openssl-dev
+            libssl-dev
 
       - name: Clone
         id: checkout
@@ -242,7 +242,7 @@ jobs:
             curl \
             wget \
             language-pack-en \
-            libcurl4-openssl-dev
+            libssl-dev
 
       - name: Clone
         id: checkout
@@ -283,6 +283,8 @@ jobs:
         run: |
           cmake -B build \
               -DGGML_NATIVE=OFF \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
               -DLLAMA_BUILD_SERVER=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
@@ -295,6 +297,8 @@ jobs:
         run: |
           cmake -B build \
               -DGGML_NATIVE=OFF \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
               -DLLAMA_BUILD_SERVER=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
@@ -306,6 +310,8 @@ jobs:
         run: |
           cmake -B build \
               -DGGML_NATIVE=OFF \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
               -DLLAMA_BUILD_SERVER=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
@@ -345,16 +351,10 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
+          cmake -B build -DLLAMA_CURL=OFF -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
@@ -368,13 +368,6 @@ jobs:
         run: |
           pip install -r tools/server/tests/requirements.txt
 
-      - name: Copy Libcurl
-        id: prepare_libcurl
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
-        run: |
-          cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
-
       - name: Tests
         id: server_integration_tests
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}

.gitignore

@@ -20,52 +20,40 @@
 *.so
 *.swp
 *.tmp
+*.DS_Store
 
 # IDE / OS
 
-.cache/
-.ccls-cache/
-.direnv/
-.DS_Store
-.envrc
-.idea/
-.swiftpm
-.vs/
-.vscode/
-nppBackup
+/.cache/
+/.ccls-cache/
+/.direnv/
+/.envrc
+/.idea/
+/.swiftpm
+/.vs/
+/.vscode/
+/nppBackup
 
 
 # Coverage
 
-gcovr-report/
-lcov-report/
+/gcovr-report/
+/lcov-report/
 
 # Build Artifacts
 
-tags
-.build/
-build*
-release
-debug
-!build-info.cmake
-!build-info.cpp.in
-!build-info.sh
-!build.zig
-!docs/build.md
+/tags
+/.build/
+/build*
+/release
+/debug
 /libllama.so
 /llama-*
 /vulkan-shaders-gen
-android-ndk-*
-arm_neon.h
-cmake-build-*
-CMakeSettings.json
-compile_commands.json
-ggml-metal-embed.metal
-llama-batched-swift
 /rpc-server
-out/
-tmp/
-autogen-*.md
+/out/
+/tmp/
+/autogen-*.md
 
 # Deprecated
 
@@ -74,44 +62,38 @@ autogen-*.md
 
 # CI
 
-!.github/workflows/*.yml
+!/.github/workflows/*.yml
 
 # Models
 
-models/*
-models-mnt
-!models/.editorconfig
-!models/ggml-vocab-*.gguf*
-!models/templates
+/models/*
+/models-mnt
+!/models/.editorconfig
+!/models/ggml-vocab-*.gguf*
+!/models/templates
 
 # Zig
-zig-out/
-zig-cache/
-
-# Logs
-
-ppl-*.txt
-qnt-*.txt
-perf-*.txt
+/zig-out/
+/zig-cache/
 
 # Examples
 
-examples/jeopardy/results.txt
-tools/server/*.css.hpp
-tools/server/*.html.hpp
-tools/server/*.js.hpp
-tools/server/*.mjs.hpp
-tools/server/*.gz.hpp
-!build_64.sh
-!examples/*.bat
-!examples/*/*.kts
-!examples/*/*/*.kts
-!examples/sycl/*.bat
-!examples/sycl/*.sh
+/examples/jeopardy/results.txt
+/tools/server/*.css.hpp
+/tools/server/*.html.hpp
+/tools/server/*.js.hpp
+/tools/server/*.mjs.hpp
+/tools/server/*.gz.hpp
+!/build_64.sh
+!/examples/*.bat
+!/examples/*/*.kts
+!/examples/*/*/*.kts
+!/examples/sycl/*.bat
+!/examples/sycl/*.sh
 
 # Server Web UI temporary files
-node_modules
-tools/server/webui/dist
+/tools/server/webui/node_modules
+/tools/server/webui/dist
 
 # Python
 
@@ -147,8 +129,8 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
-.ccache/
+/.ccache/
 
 # IDE
-*.code-workspace
-.windsurf/
+/*.code-workspace
+/.windsurf/

README.md

@@ -61,6 +61,7 @@ range of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2, AVX512 and AMX support for x86 architectures
+- RVV, ZVFH, ZFH and ZICBOP support for RISC-V architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
 - Vulkan and SYCL backend support
@@ -241,6 +242,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 - [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with pre-built Mobile and Web platform wrappers and a model example)
+- [unslothai/unsloth](https://github.com/unslothai/unsloth) – 🦥 exports/saves fine-tuned and trained models to GGUF (Apache-2.0)
 
 </details>
 

common/CMakeLists.txt

@@ -50,6 +50,8 @@ add_library(${TARGET} STATIC
     base64.hpp
     chat-parser.cpp
     chat-parser.h
+    chat-parser-xml-toolcall.h
+    chat-parser-xml-toolcall.cpp
     chat.cpp
     chat.h
     common.cpp

common/arg.cpp

@@ -1232,6 +1232,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             const auto sampler_names = string_split<std::string>(value, ';');
             params.sampling.samplers = common_sampler_types_from_names(sampler_names, true);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1261,6 +1262,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.sampling.temp = std::stof(value);
             params.sampling.temp = std::max(params.sampling.temp, 0.0f);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1268,6 +1270,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k),
         [](common_params & params, int value) {
             params.sampling.top_k = value;
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1275,6 +1278,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
         [](common_params & params, const std::string & value) {
             params.sampling.top_p = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1282,6 +1286,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
         [](common_params & params, const std::string & value) {
             params.sampling.min_p = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1296,6 +1301,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_probability = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1303,6 +1309,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_threshold = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1321,6 +1328,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             params.sampling.penalty_last_n = value;
             params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1328,6 +1336,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_repeat = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1425,6 +1434,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sampling.mirostat),
         [](common_params & params, int value) {
             params.sampling.mirostat = value;
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1432,6 +1442,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_eta = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1439,6 +1450,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_tau = std::stof(value);
+            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU;
         }
     ).set_sparam());
     add_opt(common_arg(

common/chat-parser-xml-toolcall.cpp

@@ -0,0 +1,861 @@
+#include "chat.h"
+#include "chat-parser.h"
+#include "common.h"
+#include "json-partial.h"
+#include "json-schema-to-grammar.h"
+#include "log.h"
+#include "regex-partial.h"
+
+using json = nlohmann::ordered_json;
+
+class xml_toolcall_syntax_exception : public std::runtime_error {
+  public:
+    xml_toolcall_syntax_exception(const std::string & message) : std::runtime_error(message) {}
+};
+
+template<typename T>
+inline void sort_uniq(std::vector<T> &vec) {
+    std::sort(vec.begin(), vec.end());
+    vec.erase(std::unique(vec.begin(), vec.end()), vec.end());
+}
+
+template<typename T>
+inline bool all_space(const T &str) {
+    return std::all_of(str.begin(), str.end(), [](unsigned char ch) { return std::isspace(ch); });
+}
+
+static size_t utf8_truncate_safe(const std::string_view s) {
+    size_t len = s.size();
+    if (len == 0) return 0;
+    size_t i = len;
+    for (size_t back = 0; back < 4 && i > 0; ++back) {
+        --i;
+        unsigned char c = s[i];
+        if ((c & 0x80) == 0) {
+            return len;
+        } else if ((c & 0xC0) == 0xC0) {
+            size_t expected_len = 0;
+            if ((c & 0xE0) == 0xC0) expected_len = 2;
+            else if ((c & 0xF0) == 0xE0) expected_len = 3;
+            else if ((c & 0xF8) == 0xF0) expected_len = 4;
+            else return i;
+            if (len - i >= expected_len) {
+                return len;
+            } else {
+                return i;
+            }
+        }
+    }
+    return len - std::min(len, size_t(3));
+}
+
+inline void utf8_truncate_safe_resize(std::string &s) {
+    s.resize(utf8_truncate_safe(s));
+}
+
+inline std::string_view utf8_truncate_safe_view(const std::string_view s) {
+    return s.substr(0, utf8_truncate_safe(s));
+}
+
+static std::optional<common_chat_msg_parser::find_regex_result> try_find_2_literal_splited_by_spaces(common_chat_msg_parser & builder, const std::string & literal1, const std::string & literal2) {
+    if (literal1.size() == 0) return builder.try_find_literal(literal2);
+    const auto saved_pos = builder.pos();
+    while (auto res = builder.try_find_literal(literal1)) {
+        builder.consume_spaces();
+        const auto match_len = std::min(literal2.size(), builder.input().size() - builder.pos());
+        if (builder.input().compare(builder.pos(), match_len, literal2, 0, match_len) == 0) {
+            if (res->prelude.size() != res->groups[0].begin - saved_pos) {
+                res->prelude = builder.str({saved_pos, res->groups[0].begin});
+            }
+            builder.move_to(builder.pos() + match_len);
+            res->groups[0].end = builder.pos();
+            GGML_ASSERT(res->groups[0].begin != res->groups[0].end);
+            return res;
+        }
+        builder.move_to(res->groups[0].begin + 1);
+    }
+    builder.move_to(saved_pos);
+    return std::nullopt;
+}
+
+/**
+ * make a GBNF that accept any strings except those containing any of the forbidden strings.
+ */
+std::string make_gbnf_excluding(std::vector<std::string> forbids) {
+    constexpr auto charclass_escape = [](unsigned char c) -> std::string {
+        if (c == '\\' || c == ']' || c == '^' || c == '-') {
+            std::string s = "\\";
+            s.push_back((char)c);
+            return s;
+        }
+        if (isprint(c)) {
+            return std::string(1, (char)c);
+        }
+        char buf[16];
+        snprintf(buf, 15, "\\x%02X", c);
+        return std::string(buf);
+    };
+    constexpr auto build_expr = [charclass_escape](auto self, const std::vector<std::string>& forbids, int l, int r, int depth) -> std::string {
+        std::vector<std::pair<unsigned char, std::pair<int,int>>> children;
+        int i = l;
+        while (i < r) {
+            const std::string &s = forbids[i];
+            if ((int)s.size() == depth) {
+                ++i;
+                continue;
+            }
+            unsigned char c = (unsigned char)s[depth];
+            int j = i;
+            while (j < r && (int)forbids[j].size() > depth &&
+                   (unsigned char)forbids[j][depth] == c) {
+                ++j;
+            }
+            children.push_back({c, {i, j}});
+            i = j;
+        }
+        std::vector<std::string> alts;
+        if (!children.empty()) {
+            std::string cls;
+            for (auto &ch : children) cls += charclass_escape(ch.first);
+            alts.push_back(std::string("[^") + cls + "]");
+        }
+        for (auto &ch : children) {
+            std::string childExpr = self(self, forbids, ch.second.first, ch.second.second, depth+1);
+            if (!childExpr.empty()) {
+                std::string quoted_ch = "\"";
+                if (ch.first == '\\') quoted_ch += "\\\\";
+                else if (ch.first == '"') quoted_ch += "\\\"";
+                else if (isprint(ch.first)) quoted_ch.push_back(ch.first);
+                else {
+                    char buf[16];
+                    snprintf(buf, 15, "\\x%02X", ch.first);
+                    quoted_ch += buf;
+                }
+                quoted_ch += "\"";
+                std::string branch = quoted_ch + std::string(" ") + childExpr;
+                alts.push_back(branch);
+            }
+        }
+        if (alts.empty()) return "";
+        std::ostringstream oss;
+        oss << "( ";
+        for (size_t k = 0; k < alts.size(); ++k) {
+            if (k) oss << " | ";
+            oss << alts[k];
+        }
+        oss << " )";
+        return oss.str();
+    };
+    if (forbids.empty()) return "( . )*";
+    sort(forbids.begin(), forbids.end());
+    std::string expr = build_expr(build_expr, forbids, 0, forbids.size(), 0);
+    if (expr.empty()) {
+        std::string cls;
+        for (auto &s : forbids) if (!s.empty()) cls += charclass_escape((unsigned char)s[0]);
+        expr = std::string("( [^") + cls + "] )";
+    }
+    if (forbids.size() == 1)
+        return expr + "*";
+    else
+        return std::string("( ") + expr + " )*";
+}
+
+/**
+ * Build grammar for xml-style tool call
+ * form.scope_start and form.scope_end can be empty.
+ * Requires data.format for model-specific hacks.
+ */
+void build_grammar_xml_tool_call(common_chat_params & data, const json & tools, const struct xml_tool_call_format & form) {
+    GGML_ASSERT(!form.tool_start.empty());
+    GGML_ASSERT(!form.tool_sep.empty());
+    GGML_ASSERT(!form.key_start.empty());
+    GGML_ASSERT(!form.val_end.empty());
+    GGML_ASSERT(!form.tool_end.empty());
+
+    std::string key_val_sep = form.key_val_sep;
+    if (form.key_val_sep2) {
+        key_val_sep += "\n";
+        key_val_sep += *form.key_val_sep2;
+    }
+    GGML_ASSERT(!key_val_sep.empty());
+
+    if (tools.is_array() && !tools.empty()) {
+        data.grammar = build_grammar([&](const common_grammar_builder &builder) {
+            auto string_arg_val = form.last_val_end ?
+                    builder.add_rule("string-arg-val", make_gbnf_excluding({form.val_end, *form.last_val_end})) :
+                    builder.add_rule("string-arg-val", make_gbnf_excluding({form.val_end}));
+
+            std::vector<std::string> tool_rules;
+            for (const auto & tool : tools) {
+                if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
+                    LOG_WRN("Skipping tool without function: %s", tool.dump(2).c_str());
+                    continue;
+                }
+                const auto & function = tool.at("function");
+                if (!function.contains("name") || !function.at("name").is_string()) {
+                    LOG_WRN("Skipping invalid function (invalid name): %s", function.dump(2).c_str());
+                    continue;
+                }
+                if (!function.contains("parameters") || !function.at("parameters").is_object()) {
+                    LOG_WRN("Skipping invalid function (invalid parameters): %s", function.dump(2).c_str());
+                    continue;
+                }
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+
+                struct parameter_rule {
+                    std::string symbol_name;
+                    bool is_required;
+                };
+                std::vector<parameter_rule> arg_rules;
+                if (!parameters.contains("properties") || !parameters.at("properties").is_object()) {
+                    LOG_WRN("Skipping invalid function (invalid properties): %s", function.dump(2).c_str());
+                    continue;
+                } else {
+                    std::vector<std::string> requiredParameters;
+                    if (parameters.contains("required")) {
+                        try { parameters.at("required").get_to(requiredParameters); }
+                        catch (const std::runtime_error&) {
+                            LOG_WRN("Invalid function required parameters, ignoring: %s", function.at("required").dump(2).c_str());
+                        }
+                    }
+                    sort_uniq(requiredParameters);
+                    for (const auto & [key, value] : parameters.at("properties").items()) {
+                        std::string quoted_key = key;
+                        bool required = std::binary_search(requiredParameters.begin(), requiredParameters.end(), key);
+                        if (form.key_start.back() == '"' && key_val_sep[0] == '"') {
+                            quoted_key = gbnf_format_literal(key);
+                            quoted_key = quoted_key.substr(1, quoted_key.size() - 2);
+                        }
+                        arg_rules.push_back(parameter_rule {builder.add_rule("func-" + name + "-kv-" + key,
+                            gbnf_format_literal(form.key_start) + " " +
+                            gbnf_format_literal(quoted_key) + " " +
+                            gbnf_format_literal(key_val_sep) + " " +
+                            ((value.contains("type") && value["type"].is_string() && value["type"] == "string" && (!form.raw_argval || *form.raw_argval)) ?
+                                    (form.raw_argval ?
+                                            string_arg_val :
+                                            "( " + string_arg_val + " | " + builder.add_schema(name + "-arg-" + key, value) + " )"
+                                    ) :
+                                    builder.add_schema(name + "-arg-" + key, value)
+                            )
+                        ), required});
+                    }
+                }
+
+                auto next_arg_with_sep = builder.add_rule(name + "-last-arg-end", form.last_val_end ? gbnf_format_literal(*form.last_val_end) : gbnf_format_literal(form.val_end));
+                decltype(next_arg_with_sep) next_arg = "\"\"";
+                for (auto i = arg_rules.size() - 1; /* i >= 0 && */ i < arg_rules.size(); --i) {
+                    std::string include_this_arg = arg_rules[i].symbol_name + " " + next_arg_with_sep;
+                    next_arg = builder.add_rule(name + "-arg-after-" + std::to_string(i), arg_rules[i].is_required ?
+                            include_this_arg : "( " + include_this_arg + " ) | " + next_arg
+                    );
+                    include_this_arg = gbnf_format_literal(form.val_end) + " " + include_this_arg;
+                    next_arg_with_sep = builder.add_rule(name + "-arg-after-" + std::to_string(i) + "-with-sep", arg_rules[i].is_required ?
+                            include_this_arg : "( " + include_this_arg + " ) | " + next_arg_with_sep
+                    );
+                }
+
+                std::string quoted_name = name;
+                if (form.tool_start.back() == '"' && form.tool_sep[0] == '"') {
+                    quoted_name = gbnf_format_literal(name);
+                    quoted_name = quoted_name.substr(1, quoted_name.size() - 2);
+                }
+                quoted_name = gbnf_format_literal(quoted_name);
+                // Kimi-K2 uses functions.{{ tool_call['function']['name'] }}:{{ loop.index }} as function name
+                if (data.format == COMMON_CHAT_FORMAT_KIMI_K2) {
+                    quoted_name = "\"functions.\" " + quoted_name + " \":\" [0-9]+";
+                }
+                tool_rules.push_back(builder.add_rule(name + "-call",
+                        gbnf_format_literal(form.tool_start) + " " +
+                        quoted_name + " " +
+                        gbnf_format_literal(form.tool_sep) + " " +
+                        next_arg
+                ));
+            }
+
+            auto tool_call_once = builder.add_rule("root-tool-call-once", string_join(tool_rules, " | "));
+            auto tool_call_more = builder.add_rule("root-tool-call-more", gbnf_format_literal(form.tool_end) + " " + tool_call_once);
+            auto call_end = builder.add_rule("root-call-end", form.last_tool_end ? gbnf_format_literal(*form.last_tool_end) : gbnf_format_literal(form.tool_end));
+            auto tool_call_multiple_with_end = builder.add_rule("root-tool-call-multiple-with-end", tool_call_once + " " + tool_call_more + "* " + call_end);
+            builder.add_rule("root",
+                (form.scope_start.empty() ? "" : gbnf_format_literal(form.scope_start) + " ") +
+                tool_call_multiple_with_end  + "?" +
+                (form.scope_end.empty() ? "" : " " + gbnf_format_literal(form.scope_end))
+            );
+        });
+
+        // grammar trigger for tool call
+        data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, form.scope_start + form.tool_start });
+    }
+}
+
+/**
+ * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
+ * Throws xml_toolcall_syntax_exception if there is invalid syntax and cannot recover the original status for common_chat_msg_parser.
+ * form.scope_start, form.tool_sep and form.scope_end can be empty.
+ */
+inline bool parse_xml_tool_calls(common_chat_msg_parser & builder, const struct xml_tool_call_format & form) {
+    GGML_ASSERT(!form.tool_start.empty());
+    GGML_ASSERT(!form.key_start.empty());
+    GGML_ASSERT(!form.key_val_sep.empty());
+    GGML_ASSERT(!form.val_end.empty());
+    GGML_ASSERT(!form.tool_end.empty());
+
+    // Helper to choose return false or throw error
+    constexpr auto return_error = [](common_chat_msg_parser & builder, auto &start_pos, const bool &recovery) {
+        LOG_DBG("Failed to parse XML-Style tool call at position: %s\n", gbnf_format_literal(builder.consume_rest().substr(0, 20)).c_str());
+        if (recovery) {
+            builder.move_to(start_pos);
+            return false;
+        } else throw xml_toolcall_syntax_exception("Tool call parsing failed with unrecoverable errors. Try using a grammar to constrain the model’s output.");
+    };
+    // Drop substring from needle to end from a JSON
+    constexpr auto partial_json = [](std::string &json_str, std::string_view needle = "XML_TOOL_CALL_PARTIAL_FLAG") {
+        auto pos = json_str.rfind(needle);
+        if (pos == std::string::npos) {
+            return false;
+        }
+        for (auto i = pos + needle.size(); i < json_str.size(); ++i) {
+            unsigned char ch = static_cast<unsigned char>(json_str[i]);
+            if (ch != '\'' && ch != '"' && ch != '}' && ch != ':' && !std::isspace(ch)) {
+                return false;
+            }
+        }
+        if (pos != 0 && json_str[pos - 1] == '"') {
+            --pos;
+        }
+        json_str.resize(pos);
+        return true;
+    };
+    // Helper to generate a partial argument JSON
+    constexpr auto gen_partial_json = [partial_json](auto set_partial_arg, auto &arguments, auto &builder, auto &function_name) {
+        auto rest = builder.consume_rest();
+        utf8_truncate_safe_resize(rest);
+        set_partial_arg(rest, "XML_TOOL_CALL_PARTIAL_FLAG");
+        auto tool_str = arguments.dump();
+        if (partial_json(tool_str)) {
+            if (builder.add_tool_call(function_name, "", tool_str)) {
+                return;
+            }
+        }
+        LOG_DBG("Failed to parse partial XML-Style tool call, fallback to non-partial: %s\n", tool_str.c_str());
+    };
+    // Helper to find a close (because there may be form.last_val_end or form.last_tool_end)
+    constexpr auto try_find_close = [](
+            common_chat_msg_parser & builder,
+            const std::string & end,
+            const std::optional<std::string> & alt_end,
+            const std::string & end_next,
+            const std::optional<std::string> & alt_end_next
+    ) {
+        auto saved_pos = builder.pos();
+        auto tc = builder.try_find_literal(end);
+        auto val_end_size = end.size();
+        if (alt_end) {
+            auto pos_1 = builder.pos();
+            builder.move_to(saved_pos);
+            auto tc2 = try_find_2_literal_splited_by_spaces(builder, *alt_end, end_next);
+            if (alt_end_next) {
+                builder.move_to(saved_pos);
+                auto tc3 = try_find_2_literal_splited_by_spaces(builder, *alt_end, *alt_end_next);
+                if (tc3 && (!tc2 || tc2->prelude.size() > tc3->prelude.size())) {
+                    tc2 = tc3;
+                }
+            }
+            if (tc2 && (!tc || tc->prelude.size() > tc2->prelude.size())) {
+                tc = tc2;
+                tc->groups[0].end = std::min(builder.input().size(), tc->groups[0].begin + alt_end->size());
+                builder.move_to(tc->groups[0].end);
+                val_end_size = alt_end->size();
+            } else {
+                builder.move_to(pos_1);
+            }
+        }
+        return std::make_pair(val_end_size, tc);
+    };
+    // Helper to find a val_end or last_val_end, returns matched pattern size
+    const auto try_find_val_end = [try_find_close, &builder, &form]() {
+        return try_find_close(builder, form.val_end, form.last_val_end, form.tool_end, form.last_tool_end);
+    };
+    // Helper to find a tool_end or last_tool_end, returns matched pattern size
+    const auto try_find_tool_end = [try_find_close, &builder, &form]() {
+        return try_find_close(builder, form.tool_end, form.last_tool_end, form.scope_end, std::nullopt);
+    };
+
+    bool recovery = true;
+    const auto start_pos = builder.pos();
+    if (!all_space(form.scope_start)) {
+        if (auto tc = builder.try_find_literal(form.scope_start)) {
+            if (all_space(tc->prelude)) {
+                if (form.scope_start.size() != tc->groups[0].end - tc->groups[0].begin)
+                    throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.scope_start));
+            } else {
+                builder.move_to(start_pos);
+                return false;
+            }
+        } else return false;
+    }
+    while (auto tc = builder.try_find_literal(form.tool_start)) {
+        if (!all_space(tc->prelude)) {
+            LOG_DBG("XML-Style tool call: Expected %s, but found %s, trying to match next pattern\n",
+                    gbnf_format_literal(form.tool_start).c_str(),
+                    gbnf_format_literal(tc->prelude).c_str()
+            );
+            builder.move_to(tc->groups[0].begin - tc->prelude.size());
+            break;
+        }
+
+        // Find tool name
+        auto func_name = builder.try_find_literal(all_space(form.tool_sep) ? form.key_start : form.tool_sep);
+        if (!func_name) {
+            auto [sz, tc] = try_find_tool_end();
+            func_name = tc;
+        }
+        if (!func_name) {
+            // Partial tool name not supported
+            throw common_chat_msg_partial_exception("incomplete tool_call");
+        }
+        // If the model generate multiple tool call and the first tool call has no argument
+        if (func_name->prelude.find(form.tool_end) != std::string::npos || (form.last_tool_end ? func_name->prelude.find(*form.last_tool_end) != std::string::npos : false)) {
+            builder.move_to(func_name->groups[0].begin - func_name->prelude.size());
+            auto [sz, tc] = try_find_tool_end();
+            func_name = tc;
+        }
+
+        // Parse tool name
+        builder.move_to(all_space(form.tool_sep) ? func_name->groups[0].begin : func_name->groups[0].end);
+        std::string function_name = string_strip(func_name->prelude);
+        // Kimi-K2 uses functions.{{ tool_call['function']['name'] }}:{{ loop.index }} as function name
+        if (builder.syntax().format == COMMON_CHAT_FORMAT_KIMI_K2) {
+            if (string_starts_with(function_name, "functions.")) {
+                static const std::regex re(":\\d+$");
+                if (std::regex_search(function_name, re)) {
+                    function_name = function_name.substr(10, function_name.rfind(":") - 10);
+                }
+            }
+        }
+
+        // Argument JSON
+        json arguments = json::object();
+
+        // Helper to generate a partial argument JSON
+        const auto gen_partial_args = [&](auto set_partial_arg) {
+            gen_partial_json(set_partial_arg, arguments, builder, function_name);
+        };
+
+        // Parse all arg_key/arg_value pairs
+        while (auto tc = builder.try_find_literal(form.key_start)) {
+            if (!all_space(tc->prelude)) {
+                LOG_DBG("XML-Style tool call: Expected %s, but found %s, trying to match next pattern\n",
+                        gbnf_format_literal(form.key_start).c_str(),
+                        gbnf_format_literal(tc->prelude).c_str()
+                );
+                builder.move_to(tc->groups[0].begin - tc->prelude.size());
+                break;
+            }
+            if (tc->groups[0].end - tc->groups[0].begin != form.key_start.size()) {
+                auto tool_call_arg = arguments.dump();
+                if (tool_call_arg.size() != 0 && tool_call_arg[tool_call_arg.size() - 1] == '}') {
+                    tool_call_arg.resize(tool_call_arg.size() - 1);
+                }
+                builder.add_tool_call(function_name, "", tool_call_arg);
+                throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.key_start));
+            }
+
+            // Parse arg_key
+            auto key_res = builder.try_find_literal(form.key_val_sep);
+            if (!key_res) {
+                gen_partial_args([&](auto &rest, auto &needle) {arguments[rest + needle] = "";});
+                throw common_chat_msg_partial_exception("Expected " + gbnf_format_literal(form.key_val_sep) + " after " + gbnf_format_literal(form.key_start));
+            }
+            if (key_res->groups[0].end - key_res->groups[0].begin != form.key_val_sep.size()) {
+                gen_partial_args([&](auto &, auto &needle) {arguments[key_res->prelude + needle] = "";});
+                throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.key_val_sep));
+            }
+            auto &key = key_res->prelude;
+            recovery = false;
+
+            // Parse arg_value
+            if (form.key_val_sep2) {
+                if (auto tc = builder.try_find_literal(*form.key_val_sep2)) {
+                    if (!all_space(tc->prelude)) {
+                        LOG_DBG("Failed to parse XML-Style tool call: Unexcepted %s between %s and %s\n",
+                                gbnf_format_literal(tc->prelude).c_str(),
+                                gbnf_format_literal(form.key_val_sep).c_str(),
+                                gbnf_format_literal(*form.key_val_sep2).c_str()
+                        );
+                        return return_error(builder, start_pos, false);
+                    }
+                    if (tc->groups[0].end - tc->groups[0].begin != form.key_val_sep2->size()) {
+                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
+                        throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(*form.key_val_sep2));
+                    }
+                } else {
+                    gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
+                    throw common_chat_msg_partial_exception("Expected " + gbnf_format_literal(*form.key_val_sep2) + " after " + gbnf_format_literal(form.key_val_sep));
+                }
+            }
+            auto val_start = builder.pos();
+
+            // Test if arg_val is a partial JSON
+            std::optional<common_json> value_json = std::nullopt;
+            if (!form.raw_argval || !*form.raw_argval) {
+                try { value_json = builder.try_consume_json(); }
+                catch (const std::runtime_error&) { builder.move_to(val_start); }
+                // TODO: Delete this when json_partial adds top-level support for null/true/false
+                if (builder.pos() == val_start) {
+                    const static std::regex number_regex(R"([0-9-][0-9]*(\.\d*)?([eE][+-]?\d*)?)");
+                    builder.consume_spaces();
+                    std::string_view sv = utf8_truncate_safe_view(builder.input());
+                    sv.remove_prefix(builder.pos());
+                    std::string rest = "a";
+                    if (sv.size() < 6) rest = sv;
+                    if (string_starts_with("null", rest) || string_starts_with("true", rest) || string_starts_with("false", rest) || std::regex_match(sv.begin(), sv.end(), number_regex)) {
+                        value_json = {123, {"123", "123"}};
+                        builder.consume_rest();
+                    } else {
+                        builder.move_to(val_start);
+                    }
+                }
+            }
+
+            // If it is a JSON and followed by </arg_value>, parse as json
+            // cannot support streaming because it may be a plain text starting with JSON
+            if (value_json) {
+                auto json_end = builder.pos();
+                builder.consume_spaces();
+                if (builder.pos() == builder.input().size()) {
+                    if (form.raw_argval && !*form.raw_argval && (value_json->json.is_string() || value_json->json.is_object() || value_json->json.is_array())) {
+                        arguments[key] = value_json->json;
+                        auto json_str = arguments.dump();
+                        if (!value_json->healing_marker.json_dump_marker.empty()) {
+                            GGML_ASSERT(std::string::npos != json_str.rfind(value_json->healing_marker.json_dump_marker));
+                            json_str.resize(json_str.rfind(value_json->healing_marker.json_dump_marker));
+                        } else {
+                            GGML_ASSERT(json_str.back() == '}');
+                            json_str.resize(json_str.size() - 1);
+                        }
+                        builder.add_tool_call(function_name, "", json_str);
+                    } else {
+                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
+                    }
+                    LOG_DBG("Possible JSON arg_value: %s\n", value_json->json.dump().c_str());
+                    throw common_chat_msg_partial_exception("JSON arg_value detected. Waiting for more tokens for validations.");
+                }
+                builder.move_to(json_end);
+                auto [val_end_size, tc] = try_find_val_end();
+                if (tc && all_space(tc->prelude) && value_json->healing_marker.marker.empty()) {
+                    if (tc->groups[0].end - tc->groups[0].begin != val_end_size) {
+                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
+                        LOG_DBG("Possible terminated JSON arg_value: %s\n", value_json->json.dump().c_str());
+                        throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.val_end) + (form.last_val_end ? gbnf_format_literal(*form.last_val_end) : ""));
+                    } else arguments[key] = value_json->json;
+                } else builder.move_to(val_start);
+            }
+
+            // If not, parse as plain text
+            if (val_start == builder.pos()) {
+                if (auto [val_end_size, value_plain] = try_find_val_end(); value_plain) {
+                    auto &value_str = value_plain->prelude;
+                    if (form.trim_raw_argval) value_str = string_strip(value_str);
+                    if (value_plain->groups[0].end - value_plain->groups[0].begin != val_end_size) {
+                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = value_str + needle;});
+                        throw common_chat_msg_partial_exception(
+                                "Expected " + gbnf_format_literal(form.val_end) +
+                                " after " + gbnf_format_literal(form.key_val_sep) +
+                                (form.key_val_sep2 ? " " + gbnf_format_literal(*form.key_val_sep2) : "")
+                        );
+                    }
+                    arguments[key] = value_str;
+                } else {
+                    if (form.trim_raw_argval) {
+                        gen_partial_args([&](auto &rest, auto &needle) {arguments[key] = string_strip(rest) + needle;});
+                    } else {
+                        gen_partial_args([&](auto &rest, auto &needle) {arguments[key] = rest + needle;});
+                    }
+                    throw common_chat_msg_partial_exception(
+                            "Expected " + gbnf_format_literal(form.val_end) +
+                            " after " + gbnf_format_literal(form.key_val_sep) +
+                            (form.key_val_sep2 ? " " + gbnf_format_literal(*form.key_val_sep2) : "")
+                    );
+                }
+            }
+        }
+
+        // Consume closing tag
+        if (auto [tool_end_size, tc] = try_find_tool_end(); tc) {
+            if (!all_space(tc->prelude)) {
+                LOG_DBG("Failed to parse XML-Style tool call: Expected %s, but found %s\n",
+                        gbnf_format_literal(form.tool_end).c_str(),
+                        gbnf_format_literal(tc->prelude).c_str()
+                );
+                return return_error(builder, start_pos, recovery);
+            }
+            if (tc->groups[0].end - tc->groups[0].begin == tool_end_size) {
+                // Add the parsed tool call
+                if (!builder.add_tool_call(function_name, "", arguments.dump())) {
+                    throw common_chat_msg_partial_exception("Failed to add XML-Style tool call");
+                }
+                recovery = false;
+                continue;
+            }
+        }
+
+        auto tool_call_arg = arguments.dump();
+        if (tool_call_arg.size() != 0 && tool_call_arg[tool_call_arg.size() - 1] == '}') {
+            tool_call_arg.resize(tool_call_arg.size() - 1);
+        }
+        builder.add_tool_call(function_name, "", tool_call_arg);
+        throw common_chat_msg_partial_exception("Expected " + gbnf_format_literal(form.tool_end) + " after " + gbnf_format_literal(form.val_end));
+    }
+    if (auto tc = builder.try_find_literal(form.scope_end)) {
+        if (!all_space(tc->prelude)) {
+            LOG_DBG("Failed to parse XML-Style tool call: Expected %s, but found %s\n",
+                    gbnf_format_literal(form.scope_end).c_str(),
+                    gbnf_format_literal(tc->prelude).c_str()
+            );
+            return return_error(builder, start_pos, recovery);
+        }
+    } else {
+        if (all_space(form.scope_end)) return true;
+        builder.consume_spaces();
+        if (builder.pos() == builder.input().size())
+            throw common_chat_msg_partial_exception("incomplete tool calls");
+        LOG_DBG("Failed to parse XML-Style tool call: Expected %s, but found %s\n",
+                gbnf_format_literal(form.scope_end).c_str(),
+                gbnf_format_literal(builder.consume_rest()).c_str()
+        );
+        return return_error(builder, start_pos, recovery);
+    }
+
+    return true;
+}
+
+/**
+ * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
+ * May cause std::runtime_error if there is invalid syntax because partial valid tool call is already sent out to client.
+ * form.scope_start, form.tool_sep and form.scope_end can be empty.
+ */
+bool common_chat_msg_parser::try_consume_xml_tool_calls(const struct xml_tool_call_format & form) {
+    auto pos = pos_;
+    auto tsize = result_.tool_calls.size();
+    try { return parse_xml_tool_calls(*this, form); }
+    catch (const xml_toolcall_syntax_exception&) {}
+    move_to(pos);
+    result_.tool_calls.resize(tsize);
+    return false;
+}
+
+/**
+ * Parse content uses reasoning and XML-Style tool call
+ * TODO: Note that form.allow_toolcall_in_think is not tested yet. If anyone confirms it works, this comment can be removed.
+ */
+inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, const struct xml_tool_call_format & form, const std::string & start_think = "<think>", const std::string & end_think = "</think>") {
+    constexpr auto rstrip = [](std::string &s) {
+        s.resize(std::distance(s.begin(), std::find_if(s.rbegin(), s.rend(), [](unsigned char ch) { return !std::isspace(ch); }).base()));
+    };
+    // Erase substring from l to r, along with additional spaces nearby
+    constexpr auto erase_spaces = [](auto &str, size_t l, size_t r) {
+        while (/* l > -1 && */ --l < str.size() && std::isspace(static_cast<unsigned char>(str[l])));
+        ++l;
+        while (++r < str.size() && std::isspace(static_cast<unsigned char>(str[r])));
+        if (l < r) str[l] = '\n';
+        if (l + 1 < r) str[l + 1] = '\n';
+        if (l != 0) l += 2;
+        str.erase(l, r - l);
+        return l;
+    };
+    constexpr auto trim_suffix = [](std::string &content, std::initializer_list<std::string_view> list) {
+        auto best_match = content.size();
+        for (auto pattern: list) {
+            if (pattern.size() == 0) continue;
+            for (auto match_idx = content.size() - std::min(pattern.size(), content.size()); content.size() > match_idx; match_idx++) {
+                auto match_len = content.size() - match_idx;
+                if (content.compare(match_idx, match_len, pattern.data(), match_len) == 0 && best_match > match_idx) {
+                    best_match = match_idx;
+                }
+            }
+        }
+        if (content.size() > best_match) {
+            content.erase(best_match);
+        }
+    };
+    const auto trim_potential_partial_word = [&start_think, &end_think, &form, trim_suffix](std::string &content) {
+        return trim_suffix(content, {
+            start_think, end_think, form.scope_start, form.tool_start, form.tool_sep, form.key_start,
+            form.key_val_sep, form.key_val_sep2 ? form.key_val_sep2->c_str() : "",
+            form.val_end, form.last_val_end ? form.last_val_end->c_str() : "",
+            form.tool_end, form.last_tool_end ? form.last_tool_end->c_str() : "",
+            form.scope_end
+        });
+    };
+
+
+    // Trim leading spaces without affecting keyword matching
+    static const common_regex spaces_regex("\\s*");
+    {
+        auto tc = builder.consume_regex(spaces_regex);
+        auto spaces = builder.str(tc.groups[0]);
+        auto s1 = spaces.size();
+        trim_potential_partial_word(spaces);
+        auto s2 = spaces.size();
+        builder.move_to(builder.pos() - (s1 - s2));
+    }
+
+    // Parse content
+    bool reasoning_unclosed = builder.syntax().thinking_forced_open;
+    std::string unclosed_reasoning_content("");
+    for (;;) {
+        auto tc = try_find_2_literal_splited_by_spaces(builder, form.scope_start, form.tool_start);
+        std::string content;
+        std::string tool_call_start;
+
+        if (tc) {
+            content = std::move(tc->prelude);
+            tool_call_start = builder.str(tc->groups[0]);
+            LOG_DBG("Matched tool start: %s\n", gbnf_format_literal(tool_call_start).c_str());
+        } else {
+            content = builder.consume_rest();
+            utf8_truncate_safe_resize(content);
+        }
+
+        // Handle unclosed think block
+        if (reasoning_unclosed) {
+            if (auto pos = content.find(end_think); pos == std::string::npos && builder.pos() != builder.input().size()) {
+                unclosed_reasoning_content += content;
+                if (form.allow_toolcall_in_think) {
+                    builder.move_to(tc->groups[0].begin);
+                    if (!builder.try_consume_xml_tool_calls(form)) {
+                        unclosed_reasoning_content += tool_call_start;
+                        builder.move_to(tc->groups[0].end);
+                    }
+                } else {
+                    unclosed_reasoning_content += tool_call_start;
+                }
+                continue;
+            } else {
+                reasoning_unclosed = false;
+                std::string reasoning_content;
+                if (pos == std::string::npos) {
+                    reasoning_content = std::move(content);
+                } else {
+                    reasoning_content = content.substr(0, pos);
+                    content.erase(0, pos + end_think.size());
+                }
+                if (builder.pos() == builder.input().size() && all_space(content)) {
+                    rstrip(reasoning_content);
+                    trim_potential_partial_word(reasoning_content);
+                    rstrip(reasoning_content);
+                    if (reasoning_content.empty()) {
+                        rstrip(unclosed_reasoning_content);
+                        trim_potential_partial_word(unclosed_reasoning_content);
+                        rstrip(unclosed_reasoning_content);
+                        if (unclosed_reasoning_content.empty()) continue;
+                    }
+                }
+                if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
+                    builder.add_content(start_think);
+                    builder.add_content(unclosed_reasoning_content);
+                    builder.add_content(reasoning_content);
+                    if (builder.pos() != builder.input().size() || !all_space(content))
+                        builder.add_content(end_think);
+                } else {
+                    builder.add_reasoning_content(unclosed_reasoning_content);
+                    builder.add_reasoning_content(reasoning_content);
+                }
+                unclosed_reasoning_content.clear();
+            }
+        }
+
+        // Handle multiple think block
+        bool toolcall_in_think = false;
+        for (auto think_start = content.find(start_think); think_start != std::string::npos; think_start = content.find(start_think, think_start)) {
+            if (auto think_end = content.find(end_think, think_start + start_think.size()); think_end != std::string::npos) {
+                if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content) {
+                    auto reasoning_content = content.substr(think_start + start_think.size(), think_end - think_start - start_think.size());
+                    builder.add_reasoning_content(reasoning_content);
+                    think_start = erase_spaces(content, think_start, think_end + end_think.size() - 1);
+                } else {
+                    think_start = think_end + end_think.size() - 1;
+                }
+            } else {
+                // This <tool_call> start is in thinking block, skip this tool call
+                auto pos = think_start + start_think.size();
+                unclosed_reasoning_content = content.substr(pos) + tool_call_start;
+                reasoning_unclosed = true;
+                content.resize(think_start);
+                toolcall_in_think = true;
+            }
+        }
+
+        if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content) {
+            rstrip(content);
+            // Handle unclosed </think> token from content: delete all </think> token
+            if (auto pos = content.rfind(end_think); pos != std::string::npos) {
+                while (pos != std::string::npos) {
+                    pos = erase_spaces(content, pos, pos + end_think.size() - 1);
+                    pos = content.rfind(end_think, pos);
+                }
+            }
+            // Strip if needed
+            if (content.size() > 0 && std::isspace(static_cast<unsigned char>(content[0]))) {
+                content = string_strip(content);
+            }
+        }
+
+        // remove potential partial suffix
+        if (content.size() > 0 && builder.pos() == builder.input().size() && unclosed_reasoning_content.empty()) {
+            rstrip(content);
+            trim_potential_partial_word(content);
+            rstrip(content);
+        }
+
+        // Add content
+        if (content.size() != 0) {
+            // If there are multiple content blocks
+            if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content && builder.result().content.size() != 0) {
+                builder.add_content("\n\n");
+            }
+            builder.add_content(content);
+        }
+
+        // This <tool_call> start is in thinking block, skip this tool call
+        if (toolcall_in_think && !form.allow_toolcall_in_think) {
+            continue;
+        }
+
+        // There is no tool call and all content is parsed
+        if (!tc) {
+            GGML_ASSERT(builder.pos() == builder.input().size());
+            GGML_ASSERT(unclosed_reasoning_content.empty());
+            GGML_ASSERT(!reasoning_unclosed);
+            break;
+        }
+
+        builder.move_to(tc->groups[0].begin);
+        if (builder.try_consume_xml_tool_calls(form)) {
+            auto end_of_tool = builder.pos();
+            builder.consume_spaces();
+            if (builder.pos() != builder.input().size()) {
+                builder.move_to(end_of_tool);
+                if (!builder.result().content.empty()) {
+                    builder.add_content("\n\n");
+                }
+            }
+        } else {
+            static const common_regex next_char_regex(".");
+            auto c = builder.str(builder.consume_regex(next_char_regex).groups[0]);
+            rstrip(c);
+            builder.add_content(c);
+        }
+    }
+}
+
+/**
+ * Parse content uses reasoning and XML-Style tool call
+ * TODO: Note that form.allow_toolcall_in_think is not tested yet. If anyone confirms it works, this comment can be removed.
+ */
+void common_chat_msg_parser::consume_reasoning_with_xml_tool_calls(const struct xml_tool_call_format & form, const std::string & start_think, const std::string & end_think) {
+    parse_msg_with_xml_tool_calls(*this, form, start_think, end_think);
+}

common/chat-parser-xml-toolcall.h

@@ -0,0 +1,45 @@
+#pragma once
+
+#include "chat.h"
+
+#include <nlohmann/json.hpp>
+
+#include <optional>
+#include <string>
+#include <vector>
+
+
+// Sample config:
+// MiniMax-M2 (left): <minimax:tool_call>\n<invoke name="tool-name">\n<parameter name="key">value</parameter>\n...</invoke>\n...</minimax:tool_call>
+// GLM 4.5   (right): <tool_call>function_name\n<arg_key>key</arg_key>\n<arg_value>value</arg_value>\n</tool_call>
+struct xml_tool_call_format {
+    std::string scope_start; // <minimax:tool_call>\n  // \n                      // can be empty
+    std::string tool_start;  // <invoke name=\"        // <tool_call>
+    std::string tool_sep;    // \">\n                  // \n                      // can be empty only for parse_xml_tool_calls
+    std::string key_start;   // <parameter name=\"     // <arg_key>
+    std::string key_val_sep; // \">                    // </arg_key>\n<arg_value>
+    std::string val_end;     // </parameter>\n         // </arg_value>\n
+    std::string tool_end;    // </invoke>\n            // </tool_call>\n
+    std::string scope_end;   // </minimax:tool_call>   //                         // can be empty
+    // Set this if there can be dynamic spaces inside key_val_sep.
+    // e.g. key_val_sep=</arg_key> key_val_sep2=<arg_value> for GLM4.5
+    std::optional<std::string> key_val_sep2 = std::nullopt;
+    // Set true if argval should only be raw string. e.g. Hello "world" hi
+    // Set false if argval should only be json string. e.g. "Hello \"world\" hi"
+    // Defaults to std::nullopt, both will be allowed.
+    std::optional<bool> raw_argval = std::nullopt;
+    std::optional<std::string> last_val_end = std::nullopt;
+    std::optional<std::string> last_tool_end = std::nullopt;
+    bool trim_raw_argval = false;
+    bool allow_toolcall_in_think = false; // TODO: UNTESTED!!!
+};
+
+// make a GBNF that accept any strings except those containing any of the forbidden strings.
+std::string make_gbnf_excluding(std::vector<std::string> forbids);
+
+/**
+ * Build grammar for xml-style tool call
+ * form.scope_start and form.scope_end can be empty.
+ * Requires data.format for model-specific hacks.
+ */
+void build_grammar_xml_tool_call(common_chat_params & data, const nlohmann::ordered_json & tools, const struct xml_tool_call_format & form);

common/chat-parser.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "chat.h"
+#include "chat-parser-xml-toolcall.h"
 #include "json-partial.h"
 #include "regex-partial.h"
 
@@ -119,5 +120,14 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
 
+    /**
+     * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
+     * form.scope_start, form.tool_sep and form.scope_end can be empty.
+     */
+    bool try_consume_xml_tool_calls(const struct xml_tool_call_format & form);
+
+    // Parse content uses reasoning and XML-Style tool call
+    void consume_reasoning_with_xml_tool_calls(const struct xml_tool_call_format & form, const std::string & start_think = "<think>", const std::string & end_think = "</think>");
+
     void clear_tools();
 };

common/chat.cpp

@@ -643,6 +643,12 @@ const char * common_chat_format_name(common_chat_format format) {
         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";
+        case COMMON_CHAT_FORMAT_MINIMAX_M2: return "MiniMax-M2";
+        case COMMON_CHAT_FORMAT_GLM_4_5: return "GLM 4.5";
+        case COMMON_CHAT_FORMAT_KIMI_K2: return "Kimi K2";
+        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: return "Qwen3 Coder";
+        case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
+        case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -1807,6 +1813,278 @@ static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
     }
 }
 
+
+static common_chat_params common_chat_params_init_minimax_m2(const common_chat_template & tmpl, const struct templates_params & params) {
+    common_chat_params data;
+    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    data.prompt = apply(tmpl, params);
+    data.format = COMMON_CHAT_FORMAT_MINIMAX_M2;
+
+    // Handle thinking tags based on prompt ending
+    if (string_ends_with(data.prompt, "<think>\n")) {
+        if (!params.enable_thinking) {
+            // Close the thinking tag immediately if thinking is disabled
+            data.prompt += "</think>\n\n";
+        } else {
+            // Mark thinking as forced open (template started with <think>)
+            data.thinking_forced_open = true;
+        }
+    }
+
+    // Preserve MiniMax-M2 special tokens
+    data.preserved_tokens = {
+        "<think>",
+        "</think>",
+        "<minimax:tool_call>",
+        "</minimax:tool_call>",
+    };
+
+    // build grammar for tool call
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "<minimax:tool_call>\n",
+        /* form.tool_start  = */ "<invoke name=\"",
+        /* form.tool_sep    = */ "\">\n",
+        /* form.key_start   = */ "<parameter name=\"",
+        /* form.key_val_sep = */ "\">",
+        /* form.val_end     = */ "</parameter>\n",
+        /* form.tool_end    = */ "</invoke>\n",
+        /* form.scope_end   = */ "</minimax:tool_call>",
+    };
+    build_grammar_xml_tool_call(data, params.tools, form);
+
+    return data;
+}
+
+static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "<minimax:tool_call>",
+        /* form.tool_start  = */ "<invoke name=\"",
+        /* form.tool_sep    = */ "\">",
+        /* form.key_start   = */ "<parameter name=\"",
+        /* form.key_val_sep = */ "\">",
+        /* form.val_end     = */ "</parameter>",
+        /* form.tool_end    = */ "</invoke>",
+        /* form.scope_end   = */ "</minimax:tool_call>",
+    };
+    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
+}
+
+static common_chat_params common_chat_params_init_qwen3_coder_xml(const common_chat_template & tmpl, const struct templates_params & params) {
+    common_chat_params data;
+    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    data.prompt = apply(tmpl, params);
+    data.format = COMMON_CHAT_FORMAT_QWEN3_CODER_XML;
+
+    data.preserved_tokens = {
+        "<tool_call>",
+        "</tool_call>",
+        "<function=",
+        "</function>",
+        "<parameter=",
+        "</parameter>",
+    };
+
+    // build grammar for tool call
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "<tool_call>\n",
+        /* form.tool_start  = */ "<function=",
+        /* form.tool_sep    = */ ">\n",
+        /* form.key_start   = */ "<parameter=",
+        /* form.key_val_sep = */ ">\n",
+        /* form.val_end     = */ "\n</parameter>\n",
+        /* form.tool_end    = */ "</function>\n",
+        /* form.scope_end   = */ "</tool_call>",
+    };
+    build_grammar_xml_tool_call(data, params.tools, form);
+
+    return data;
+}
+
+static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<tool_call>";
+        form.tool_start  = "<function=";
+        form.tool_sep    = ">";
+        form.key_start   = "<parameter=";
+        form.key_val_sep = ">";
+        form.val_end     = "</parameter>";
+        form.tool_end    = "</function>";
+        form.scope_end   = "</tool_call>";
+        form.trim_raw_argval = true;
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form);
+}
+
+static common_chat_params common_chat_params_init_kimi_k2(const common_chat_template & tmpl, const struct templates_params & params) {
+    common_chat_params data;
+    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    data.prompt = apply(tmpl, params);
+    data.format = COMMON_CHAT_FORMAT_KIMI_K2;
+
+    data.preserved_tokens = {
+        "<think>",
+        "</think>",
+        "<|tool_calls_section_begin|>",
+        "<|tool_call_begin|>",
+        "<|tool_call_argument_begin|>",
+        "<|tool_call_end|>",
+        "<|tool_calls_section_end|>",
+        "<|im_end|>",
+        "<|im_system|>",
+        "<|im_middle|>",
+    };
+
+    data.additional_stops.insert(data.additional_stops.end(), {
+        "<|im_end|>",
+        "<|im_middle|>"
+    });
+    // build grammar for tool call
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<|tool_calls_section_begin|>";
+        form.tool_start  = "<|tool_call_begin|>";
+        form.tool_sep    = "<|tool_call_argument_begin|>{";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}<|tool_call_end|>";
+        form.scope_end   = "<|tool_calls_section_end|>";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        return form;
+    })();
+    build_grammar_xml_tool_call(data, params.tools, form);
+
+    return data;
+}
+
+static void common_chat_parse_kimi_k2(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<|tool_calls_section_begin|>";
+        form.tool_start  = "<|tool_call_begin|>";
+        form.tool_sep    = "<|tool_call_argument_begin|>{";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}<|tool_call_end|>";
+        form.scope_end   = "<|tool_calls_section_end|>";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
+}
+
+static common_chat_params common_chat_params_init_apriel_1_5(const common_chat_template & tmpl, const struct templates_params & params) {
+    common_chat_params data;
+    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    data.prompt = apply(tmpl, params);
+    data.format = COMMON_CHAT_FORMAT_APRIEL_1_5;
+
+    data.preserved_tokens = {
+        "<thinking>",
+        "</thinking>",
+        "<tool_calls>",
+        "</tool_calls>",
+    };
+
+    // build grammar for tool call
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<tool_calls>[";
+        form.tool_start  = "{\"name\": \"";
+        form.tool_sep    = "\", \"arguments\": {";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}, ";
+        form.scope_end   = "]</tool_calls>";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        form.last_tool_end = "}";
+        return form;
+    })();
+    build_grammar_xml_tool_call(data, params.tools, form);
+
+    return data;
+}
+
+static void common_chat_parse_apriel_1_5(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "<tool_calls>[";
+        form.tool_start  = "{\"name\": \"";
+        form.tool_sep    = "\", \"arguments\": {";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}, ";
+        form.scope_end   = "]</tool_calls>";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        form.last_tool_end = "}";
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form, "<thinking>", "</thinking>");
+}
+
+static common_chat_params common_chat_params_init_xiaomi_mimo(const common_chat_template & tmpl, const struct templates_params & params) {
+    common_chat_params data;
+    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    data.prompt = apply(tmpl, params);
+    data.format = COMMON_CHAT_FORMAT_XIAOMI_MIMO;
+
+    data.preserved_tokens = {
+        "<tool_call>",
+        "</tool_call>",
+    };
+
+    // build grammar for tool call
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "\n";
+        form.tool_start  = "<tool_call>\n{\"name\": \"";
+        form.tool_sep    = "\", \"arguments\": {";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}\n</tool_call>";
+        form.scope_end   = "";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        return form;
+    })();
+    build_grammar_xml_tool_call(data, params.tools, form);
+
+    return data;
+}
+
+static void common_chat_parse_xiaomi_mimo(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form = ([]() {
+        xml_tool_call_format form {};
+        form.scope_start = "";
+        form.tool_start  = "<tool_call>\n{\"name\": \"";
+        form.tool_sep    = "\", \"arguments\": {";
+        form.key_start   = "\"";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
+        form.tool_end    = "}\n</tool_call>";
+        form.scope_end   = "";
+        form.raw_argval  = false;
+        form.last_val_end = "";
+        return form;
+    })();
+    builder.consume_reasoning_with_xml_tool_calls(form);
+}
+
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
@@ -2041,6 +2319,100 @@ static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
     }
 }
 
+static common_chat_params common_chat_params_init_glm_4_5(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+    data.grammar_lazy = inputs.tools.is_array() && !inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    std::string prompt = apply(tmpl, inputs);
+
+    // match the existing trimming behavior
+    if (inputs.add_bos && string_starts_with(prompt, tmpl.bos_token())) {
+        prompt.erase(0, tmpl.bos_token().size());
+    }
+    if (inputs.add_eos && string_ends_with(prompt, tmpl.eos_token())) {
+        prompt.erase(prompt.size() - tmpl.eos_token().size());
+    }
+    if (string_ends_with(prompt, "<think>")) {
+        if (!inputs.enable_thinking) {
+            prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+
+    // add GLM preserved tokens
+    data.preserved_tokens = {
+        "<|endoftext|>",
+        "[MASK]",
+        "[gMASK]",
+        "[sMASK]",
+        "<sop>",
+        "<eop>",
+        "<|system|>",
+        "<|user|>",
+        "<|assistant|>",
+        "<|observation|>",
+        "<|begin_of_image|>",
+        "<|end_of_image|>",
+        "<|begin_of_video|>",
+        "<|end_of_video|>",
+        "<|begin_of_audio|>",
+        "<|end_of_audio|>",
+        "<|begin_of_transcription|>",
+        "<|end_of_transcription|>",
+        "<|code_prefix|>",
+        "<|code_middle|>",
+        "<|code_suffix|>",
+        "/nothink",
+        "<think>",
+        "</think>",
+        "<tool_call>",
+        "</tool_call>",
+        "<arg_key>",
+        "</arg_key>",
+        "<arg_value>",
+        "</arg_value>"
+    };
+
+    // extra GLM 4.5 stop word
+    data.additional_stops.insert(data.additional_stops.end(), {
+        "<|user|>",
+        "<|observation|>"
+    });
+
+    // build grammar for tool call
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "",
+        /* form.tool_start  = */ "\n<tool_call>",
+        /* form.tool_sep    = */ "\n",
+        /* form.key_start   = */ "<arg_key>",
+        /* form.key_val_sep = */ "</arg_key>\n<arg_value>",
+        /* form.val_end     = */ "</arg_value>\n",
+        /* form.tool_end    = */ "</tool_call>\n",
+        /* form.scope_end   = */ "",
+    };
+    build_grammar_xml_tool_call(data, inputs.tools, form);
+
+    data.prompt = prompt;
+    data.format = COMMON_CHAT_FORMAT_GLM_4_5;
+    return data;
+}
+
+static void common_chat_parse_glm_4_5(common_chat_msg_parser & builder) {
+    static const xml_tool_call_format form {
+        /* form.scope_start  = */ "",
+        /* form.tool_start   = */ "<tool_call>",
+        /* form.tool_sep     = */ "",
+        /* form.key_start    = */ "<arg_key>",
+        /* form.key_val_sep  = */ "</arg_key>",
+        /* form.val_end      = */ "</arg_value>",
+        /* form.tool_end     = */ "</tool_call>",
+        /* form.scope_end    = */ "",
+        /* form.key_val_sep2 = */ "<arg_value>",
+    };
+    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
+}
+
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
@@ -2704,91 +3076,17 @@ static void common_chat_parse_lfm2(common_chat_msg_parser & builder) {
 }
 
 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>");
-
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // Parse tool calls - Seed-OSS uses <seed:tool_call> format
-    static const common_regex tool_call_begin_regex("<seed:tool_call>");
-    static const common_regex tool_call_end_regex("</seed:tool_call>");
-    static const common_regex function_regex("<function=([^>]+)>");
-    static const common_regex param_regex("<parameter=([^>]+)>");
-
-    while (auto tool_res = builder.try_find_regex(tool_call_begin_regex)) {
-        builder.consume_spaces();  // Consume whitespace after <seed:tool_call>
-
-        // Look for function call inside tool call, ignore any content before it
-        if (auto func_res = builder.try_find_regex(function_regex, std::string::npos, false)) {
-            auto function_name = builder.str(func_res->groups[1]);
-
-            // Parse Seed-OSS parameters <parameter=name>value</parameter>
-            json args = json::object();
-            // Parse all parameters
-            while (auto param_res = builder.try_find_regex(param_regex, std::string::npos, false)) {
-                // again, ignore noise around parameters
-                auto param_name = builder.str(param_res->groups[1]);
-                builder.move_to(param_res->groups[0].end);
-                builder.consume_spaces();  // Consume whitespace after parameter
-                auto savedPos = builder.pos();
-                if (auto param_parse = builder.try_find_literal("</parameter>")) {
-                    auto param = param_parse->prelude;
-                    builder.move_to(savedPos);
-                    try {
-                        if (auto param_res = builder.try_consume_json()) {
-                            args[param_name] = param_res->json;
-                        } else {
-                            args[param_name] = param;
-                        }
-                    } catch (json::exception &) {
-                        args[param_name] = param;
-                    }
-                } else {
-                    throw common_chat_msg_partial_exception("Incomplete tool parameter");
-                }
-            }
-            // Look for closing function tag
-            auto end_func = builder.try_find_literal("</function>");
-            if (end_func) {
-                builder.move_to(end_func->groups[0].end);
-                builder.consume_spaces();  // Consume whitespace after </function>
-
-                // Add the tool call with parsed arguments, but only if we REALLY got the literal
-                auto eaten_fragment = builder.input().substr(end_func->groups[0].begin, end_func->groups[0].end);
-                auto funlen = std::string("</function>").length();
-                if (eaten_fragment.length() >= funlen && eaten_fragment.substr(0, funlen) == std::string("</function>")) {
-                    if (!builder.add_tool_call(function_name, "", args.dump())) {
-                        throw common_chat_msg_partial_exception("Incomplete tool call");
-                    }
-                } else {
-                    throw common_chat_msg_partial_exception("Incomplete tool call");
-                }
-            } else {
-                throw common_chat_msg_partial_exception("Incomplete tool call");
-            }
-            // Look for closing tool call tag
-            if (auto end_tool = builder.try_find_regex(tool_call_end_regex, std::string::npos, false)) {
-                builder.move_to(end_tool->groups[0].end);
-                builder.consume_spaces();  // Consume trailing whitespace after tool call
-            } else {
-                throw common_chat_msg_partial_exception("Incomplete tool call");
-            }
-        } else {
-            // No function found - don't consume content here, let it be handled at the end
-            break;
-        }
-    }
-
-    // Consume any remaining whitespace after all tool call processing
-    builder.consume_spaces();
-    auto remaining = builder.consume_rest();
-    // If there's any non-whitespace content remaining, add it as content
-    if (!string_strip(remaining).empty()) {
-        builder.add_content(remaining);
-    }
+    static const xml_tool_call_format form {
+        /* form.scope_start = */ "<seed:tool_call>",
+        /* form.tool_start  = */ "<function=",
+        /* form.tool_sep    = */ ">",
+        /* form.key_start   = */ "<parameter=",
+        /* form.key_val_sep = */ ">",
+        /* form.val_end     = */ "</parameter>",
+        /* form.tool_end    = */ "</function>",
+        /* form.scope_end   = */ "</seed:tool_call>",
+    };
+    builder.consume_reasoning_with_xml_tool_calls(form, "<seed:think>", "</seed:think>");
 }
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
@@ -2927,6 +3225,35 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_granite(tmpl, params);
     }
 
+    // GLM 4.5: detect by <arg_key> and <arg_value> tags (check before Hermes since both use <tool_call>)
+    if (src.find("[gMASK]<sop>") != std::string::npos &&
+        src.find("<arg_key>") != std::string::npos &&
+        src.find("<arg_value>") != std::string::npos &&
+        params.json_schema.is_null()) {
+        return common_chat_params_init_glm_4_5(tmpl, params);
+    }
+
+    // Qwen3-Coder XML format detection (must come before Hermes 2 Pro)
+    // Detect via explicit XML markers unique to Qwen3-Coder to avoid false positives in other templates.
+    // Require presence of <tool_call>, <function=...>, and <parameter=...> blocks.
+    if (src.find("<tool_call>") != std::string::npos &&
+        src.find("<function>") != std::string::npos &&
+        src.find("<function=") != std::string::npos &&
+        src.find("<parameters>") != std::string::npos &&
+        src.find("<parameter=") != std::string::npos) {
+        return common_chat_params_init_qwen3_coder_xml(tmpl, params);
+    }
+
+    // Xiaomi MiMo format detection (must come before Hermes 2 Pro)
+    if (src.find("<tools>") != std::string::npos &&
+        src.find("# Tools") != std::string::npos &&
+        src.find("</tools>") != std::string::npos &&
+        src.find("<tool_calls>") != std::string::npos &&
+        src.find("</tool_calls>") != std::string::npos &&
+        src.find("<tool_response>") != std::string::npos) {
+        return common_chat_params_init_xiaomi_mimo(tmpl, params);
+    }
+
     // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
     if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
         return common_chat_params_init_hermes_2_pro(tmpl, params);
@@ -2958,6 +3285,29 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_lfm2(tmpl, params);
     }
 
+    // MiniMax-M2 format detection
+    if (src.find("]~!b[") != std::string::npos && src.find("]~b]") != std::string::npos) {
+        return common_chat_params_init_minimax_m2(tmpl, params);
+    }
+
+    // Kimi K2 format detection
+    if (src.find("<|im_system|>tool_declare<|im_middle|>") != std::string::npos &&
+        src.find("<|tool_calls_section_begin|>") != std::string::npos &&
+        src.find("## Return of") != std::string::npos) {
+        return common_chat_params_init_kimi_k2(tmpl, params);
+    }
+
+    // Apriel 1.5 format detection
+    if (src.find("<thinking>") != std::string::npos &&
+        src.find("</thinking>") != std::string::npos &&
+        src.find("<available_tools>") != std::string::npos &&
+        src.find("<|assistant|>") != std::string::npos &&
+        src.find("<|tool_result|>") != std::string::npos &&
+        src.find("<tool_calls>[") != std::string::npos &&
+        src.find("]</tool_calls>") != std::string::npos) {
+        return common_chat_params_init_apriel_1_5(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())) {
@@ -3009,7 +3359,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     const struct common_chat_templates * tmpls,
     const struct common_chat_templates_inputs & inputs)
 {
-    int alloc_size = 0;
+    size_t alloc_size = 0;
     std::vector<llama_chat_message> chat;
     std::vector<std::string> contents;
 
@@ -3031,7 +3381,8 @@ static common_chat_params common_chat_templates_apply_legacy(
         const auto & msg = inputs.messages[i];
         const auto & content = contents[i];
         chat.push_back({msg.role.c_str(), content.c_str()});
-        alloc_size += (msg.role.size() + content.size()) * 1.25;
+        size_t msg_size = msg.role.size() + content.size();
+        alloc_size += msg_size + (msg_size / 4); // == msg_size * 1.25 but avoiding float ops
     }
 
     std::vector<char> buf(alloc_size);
@@ -3053,6 +3404,11 @@ static common_chat_params common_chat_templates_apply_legacy(
         res = llama_chat_apply_template(src.c_str(), chat.data(), chat.size(), inputs.add_generation_prompt, buf.data(), buf.size());
     }
 
+    // for safety, we check the result again
+    if (res < 0 || (size_t) res > buf.size()) {
+        throw std::runtime_error("failed to apply chat template, try using --jinja");
+    }
+
     common_chat_params params;
     params.prompt = std::string(buf.data(), res);
     if (!inputs.json_schema.empty()) {
@@ -3139,6 +3495,24 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS:
             common_chat_parse_lfm2(builder);
             break;
+        case COMMON_CHAT_FORMAT_MINIMAX_M2:
+            common_chat_parse_minimax_m2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_GLM_4_5:
+            common_chat_parse_glm_4_5(builder);
+            break;
+        case COMMON_CHAT_FORMAT_KIMI_K2:
+            common_chat_parse_kimi_k2(builder);
+            break;
+        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML:
+            common_chat_parse_qwen3_coder_xml(builder);
+            break;
+        case COMMON_CHAT_FORMAT_APRIEL_1_5:
+            common_chat_parse_apriel_1_5(builder);
+            break;
+        case COMMON_CHAT_FORMAT_XIAOMI_MIMO:
+            common_chat_parse_xiaomi_mimo(builder);
+            break;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }

common/chat.h

@@ -117,6 +117,12 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_NEMOTRON_V2,
     COMMON_CHAT_FORMAT_APERTUS,
     COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS,
+    COMMON_CHAT_FORMAT_GLM_4_5,
+    COMMON_CHAT_FORMAT_MINIMAX_M2,
+    COMMON_CHAT_FORMAT_KIMI_K2,
+    COMMON_CHAT_FORMAT_QWEN3_CODER_XML,
+    COMMON_CHAT_FORMAT_APRIEL_1_5,
+    COMMON_CHAT_FORMAT_XIAOMI_MIMO,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };

common/common.cpp

@@ -8,6 +8,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -26,7 +27,6 @@
 #include <sstream>
 #include <string>
 #include <thread>
-#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -60,6 +60,14 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+common_time_meas::~common_time_meas() {
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
+    }
+}
+
 //
 // CPU utils
 //
@@ -355,11 +363,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
-    llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
-        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
-            common_log_add(common_log_main(), level, "%s", text);
-        }
-    }, NULL);
+    llama_log_set(common_log_default_callback, NULL);
 
 #ifdef NDEBUG
     const char * build_type = "";
@@ -946,6 +950,58 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
 // Model utils
 //
 
+static inline void common_init_sampler_from_model(
+    const llama_model * model,
+    common_params_sampling & sparams) {
+
+    const uint64_t config = sparams.user_sampling_config;
+
+    auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[64] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            int32_t v = strtol(buf, &end, 10);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[128] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            float v = strtof(buf, &end);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    // Sampling sequence
+    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
+        char buf[512] = {0};
+        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
+            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
+            if (!sampler_names.empty()) {
+                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            }
+        }
+    }
+
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
+}
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
@@ -957,6 +1013,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
+    common_init_sampler_from_model(model, params.sampling);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     auto cparams = common_context_params_to_llama(params);

common/common.h

@@ -2,17 +2,15 @@
 
 #pragma once
 
+#include "ggml-opt.h"
+#include "llama-cpp.h"
+
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
-#include <sstream>
-#include <cmath>
-
-#include "ggml-opt.h"
-#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -30,6 +28,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct common_time_meas {
+    common_time_meas(int64_t & t_acc, bool disable = false);
+    ~common_time_meas();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
 struct common_adapter_lora_info {
     std::string path;
     float scale;
@@ -133,6 +140,22 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
+enum common_params_sampling_config : uint64_t {
+    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
+    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
+};
+
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -165,6 +188,8 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
+    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
+
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
 

common/json-partial.cpp

@@ -297,8 +297,25 @@ bool common_json_parse(
             it = temptative_end;
             return true;
         }
-        // TODO: handle unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
-        // fprintf(stderr, "Closing: TODO\n");
+        // handle unclosed top-level primitive
+        if (err_loc.position != 0 && !healing_marker.empty() && err_loc.stack.empty()) {
+            std::string str(it, temptative_end);
+            const auto & magic_seed = out.healing_marker.marker = healing_marker;
+            if (can_parse(str + "\"")) {
+                // Was inside an string
+                str += (out.healing_marker.json_dump_marker = magic_seed) + "\"";
+            } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"")) {
+                // Was inside an string after an escape
+                str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"";
+            } else {
+                // TODO: handle more unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
+                // fprintf(stderr, "Closing: TODO\n");
+                return false;
+            }
+            out.json = json::parse(str);
+            it = temptative_end;
+            return true;
+        }
         return false;
     }
     out.json = json::parse(it, end);

common/json-schema-to-grammar.cpp

@@ -303,6 +303,8 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
+std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
+
 class SchemaConverter {
 private:
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);

common/json-schema-to-grammar.h

@@ -18,4 +18,6 @@ struct common_grammar_options {
     bool dotall = false;
 };
 
+std::string gbnf_format_literal(const std::string & literal);
+
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

common/log.cpp

@@ -442,3 +442,9 @@ void common_log_set_prefix(struct common_log * log, bool prefix) {
 void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
+
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
+    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+        common_log_add(common_log_main(), level, "%s", text);
+    }
+}

common/log.h

@@ -36,6 +36,8 @@ extern int common_log_verbosity_thold;
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
+
 // the common_log uses an internal worker thread to print/write log messages
 // when the worker thread is paused, incoming log messages are discarded
 struct common_log;

common/sampling.cpp

@@ -3,9 +3,10 @@
 #include "common.h"
 #include "log.h"
 
-#include <cmath>
-#include <unordered_map>
 #include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <unordered_map>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -112,6 +113,13 @@ struct common_sampler {
 
     llama_token_data_array cur_p;
 
+    void reset() {
+        prev.clear();
+
+        llama_sampler_reset(grmr);
+        llama_sampler_reset(chain);
+    }
+
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
@@ -128,6 +136,12 @@ struct common_sampler {
 
         cur_p = { cur.data(), cur.size(), -1, false };
     }
+
+    common_time_meas tm() {
+        return common_time_meas(t_total_us, params.no_perf);
+    }
+
+    mutable int64_t t_total_us = 0;
 };
 
 std::string common_params_sampling::print() const {
@@ -298,6 +312,8 @@ void common_sampler_free(struct common_sampler * gsmpl) {
 }
 
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
+    const auto tm = gsmpl->tm();
+
     if (accept_grammar) {
         llama_sampler_accept(gsmpl->grmr, token);
     }
@@ -308,9 +324,7 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
 }
 
 void common_sampler_reset(struct common_sampler * gsmpl) {
-    llama_sampler_reset(gsmpl->grmr);
-
-    llama_sampler_reset(gsmpl->chain);
+    gsmpl->reset();
 }
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
@@ -327,16 +341,54 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
     // TODO: measure grammar performance
 
+    const double t_sampling_ms = gsmpl ? 1e-3*gsmpl->t_total_us : 0;
+
+    llama_perf_sampler_data data_smpl;
+    llama_perf_context_data data_ctx;
+
+    memset(&data_smpl, 0, sizeof(data_smpl));
+    memset(&data_ctx,  0, sizeof(data_ctx));
+
     if (gsmpl) {
-        llama_perf_sampler_print(gsmpl->chain);
+        auto & data = data_smpl;
+
+        data = llama_perf_sampler(gsmpl->chain);
+
+        // note: the sampling time includes the samplers time + extra time spent in common/sampling
+        LOG_INF("%s:    sampling time = %10.2f ms\n", __func__, t_sampling_ms);
+        LOG_INF("%s:    samplers time = %10.2f ms / %5d tokens\n", __func__, data.t_sample_ms, data.n_sample);
     }
+
     if (ctx) {
-        llama_perf_context_print(ctx);
+        auto & data = data_ctx;
+
+        data = llama_perf_context(ctx);
+
+        const double t_end_ms = 1e-3 * ggml_time_us();
+
+        const double t_total_ms = t_end_ms - data.t_start_ms;
+        const double t_unacc_ms = t_total_ms - (t_sampling_ms + data.t_p_eval_ms + data.t_eval_ms);
+        const double t_unacc_pc = 100.0 * t_unacc_ms /  t_total_ms;
+
+        LOG_INF("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
+        LOG_INF("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
+        LOG_INF("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
+        LOG_INF("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+        LOG_INF("%s: unaccounted time = %10.2f ms / %5.1f %%      (total - sampling - prompt eval - eval) / (total)\n", __func__, t_unacc_ms, t_unacc_pc);
+        LOG_INF("%s:    graphs reused = %10d\n", __func__, data.n_reused);
+
         llama_memory_breakdown_print(ctx);
     }
 }
 
 llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+    llama_synchronize(ctx);
+
+    // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
+    const auto tm = gsmpl->tm();
+
     gsmpl->set_logits(ctx, idx);
 
     auto & grmr  = gsmpl->grmr;
@@ -428,6 +480,8 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 // helpers
 
 llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    const auto tm = gsmpl->tm();
+
     auto * res = &gsmpl->cur_p;
 
     if (do_sort && !res->sorted) {

convert_hf_to_gguf.py

@@ -189,10 +189,10 @@ class ModelBase:
             return tensors
 
         prefix = "model" if not self.is_mistral_format else "consolidated"
-        part_names: list[str] = ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors")
+        part_names: set[str] = set(ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors"))
         is_safetensors: bool = len(part_names) > 0
         if not is_safetensors:
-            part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
+            part_names = set(ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin"))
 
         tensor_names_from_index: set[str] = set()
 
@@ -209,6 +209,7 @@ class ModelBase:
                     if weight_map is None or not isinstance(weight_map, dict):
                         raise ValueError(f"Can't load 'weight_map' from {index_name!r}")
                     tensor_names_from_index.update(weight_map.keys())
+                    part_names |= set(weight_map.values())
             else:
                 weight_map = {}
         else:
@@ -564,7 +565,7 @@ class ModelBase:
                             gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
                         )
                     )
-                    or not new_name.endswith(".weight")
+                    or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
                 ):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
@@ -825,6 +826,15 @@ class TextModel(ModelBase):
             self.gguf_writer.add_expert_group_used_count(n_group_used)
             logger.info(f"gguf: expert groups used count = {n_group_used}")
 
+        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func"], optional=True)) is not None:
+            if score_func == "sigmoid":
+                self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+            elif score_func == "softmax":
+                self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+            else:
+                raise ValueError(f"Unsupported expert score gating function value: {score_func}")
+            logger.info(f"gguf: expert score gating function = {score_func}")
+
         if (head_dim := self.hparams.get("head_dim")) is not None:
             self.gguf_writer.add_key_length(head_dim)
             self.gguf_writer.add_value_length(head_dim)
@@ -1124,6 +1134,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
             # ref: https://huggingface.co/JetBrains/Mellum-4b-base
             res = "mellum"
+        if chkhsh == "49fc0303c9e0d2c2c565c510f64b2d9b271276acdcdadff733249eda9f7d59df":
+            # ref: https://huggingface.co/arcee-ai/Trinity-Tokenizer
+            res = "afmoe"
         if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
             # ref: https://huggingface.co/inclusionAI/Ling-mini-base-2.0
             res = "bailingmoe2"
@@ -1660,11 +1673,9 @@ class GPTNeoXModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GPTNEOX
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
-
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_dimension_count(
             int(self.hparams["rotary_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])),
@@ -1722,7 +1733,7 @@ class BloomModel(TextModel):
         self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
         self.gguf_writer.add_embedding_length(n_embed)
         self.gguf_writer.add_feed_forward_length(4 * n_embed)
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -1785,10 +1796,9 @@ class MPTModel(TextModel):
             self.gguf_writer.add_unk_token_id(0)
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["n_layers"]
         self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["d_model"])
         self.gguf_writer.add_head_count(self.hparams["n_heads"])
         if kv_n_heads := self.hparams["attn_config"].get("kv_n_heads"):
@@ -1821,7 +1831,6 @@ class OrionModel(TextModel):
         self._set_vocab_sentencepiece()
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -1839,7 +1848,7 @@ class OrionModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
@@ -1856,7 +1865,6 @@ class BaichuanModel(TextModel):
         self._set_vocab_sentencepiece()
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -1873,7 +1881,7 @@ class BaichuanModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(head_count)
@@ -1980,7 +1988,6 @@ class XverseModel(TextModel):
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -1997,7 +2004,7 @@ class XverseModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(head_count)
@@ -2040,10 +2047,6 @@ class FalconModel(TextModel):
     model_arch = gguf.MODEL_ARCH.FALCON
 
     def set_gguf_parameters(self):
-        block_count = self.hparams.get("num_hidden_layers")
-        if block_count is None:
-            block_count = self.hparams["n_layer"]  # old name
-
         n_head = self.hparams.get("num_attention_heads")
         if n_head is None:
             n_head = self.hparams["n_head"]  # old name
@@ -2056,7 +2059,7 @@ class FalconModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("jploski")  # qkv tensor transform
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -2094,12 +2097,10 @@ class StarCoderModel(TextModel):
     model_arch = gguf.MODEL_ARCH.STARCODER
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["n_layer"]
-
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
         self.gguf_writer.add_head_count_kv(1)
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -2129,14 +2130,12 @@ class RefactModel(TextModel):
         multiple_of = 256
         ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
 
-        block_count = self.hparams["n_layer"]
-
         # refact uses Alibi. So this is from config.json which might be used by training.
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
 
         self.gguf_writer.add_feed_forward_length(ff_dim)
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
         self.gguf_writer.add_head_count_kv(1)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
@@ -2183,11 +2182,10 @@ class StableLMModel(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
@@ -2533,6 +2531,72 @@ class ArceeModel(LlamaModel):
             self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
 
+@ModelBase.register("AfmoeForCausalLM")
+class AfmoeModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.AFMOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # MoE parameters
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (n_shared_experts := self.hparams.get("num_shared_experts")) is not None:
+            self.gguf_writer.add_expert_shared_count(n_shared_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        if (n_dense_layers := self.hparams.get("num_dense_layers")) is not None:
+            self.gguf_writer.add_leading_dense_block_count(n_dense_layers)
+
+        # Route normalization and scaling
+        if (route_norm := self.hparams.get("route_norm")) is not None:
+            self.gguf_writer.add_expert_weights_norm(route_norm)
+        if (route_scale := self.hparams.get("route_scale")) is not None:
+            self.gguf_writer.add_expert_weights_scale(route_scale)
+
+        # Sliding window attention
+        if (sliding_window := self.hparams.get("sliding_window")) is not None:
+            self.gguf_writer.add_sliding_window(sliding_window)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Handle expert weights - they're already merged in the HF format
+        # process the experts separately
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["gate_proj", "up_proj", "down_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename_to_retrieve])
+                        del self._experts[bid][ename_to_retrieve]
+
+                    data_torch = torch.stack(datas, dim=0)
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+                    new_name = self.map_tensor_name(merged_name)
+                    tensors.append((new_name, data_torch))
+
+                return tensors
+            else:
+                return []
+
+        if name.endswith(".expert_bias"):
+            name = name.replace(".expert_bias", ".expert_bias.bias")
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @ModelBase.register(
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
@@ -3072,7 +3136,7 @@ class DbrxModel(TextModel):
     def set_gguf_parameters(self):
         ffn_config = self.hparams["ffn_config"]
         attn_config = self.hparams["attn_config"]
-        self.gguf_writer.add_block_count(self.hparams["n_layers"])
+        self.gguf_writer.add_block_count(self.block_count)
 
         self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
@@ -3274,7 +3338,7 @@ class QwenModel(TextModel):
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
@@ -4119,6 +4183,21 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+@ModelBase.register("RND1")
+class RND1Model(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.RND1
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # RND1 specific parameters
+        # RND1 uses bidirectional attention
+        self.gguf_writer.add_causal_attention(False)
+
+        if (mask_token_id := self.hparams.get("mask_token_id")) is not None:
+            self.gguf_writer.add_mask_token_id(mask_token_id)
+
+
 @ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
 class Qwen3VLVisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
@@ -4305,7 +4384,7 @@ class GPT2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GPT2
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_context_length(self.hparams["n_ctx"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
@@ -4337,8 +4416,6 @@ class Phi2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI2
 
     def set_gguf_parameters(self):
-        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
-
         rot_pct = self.find_hparam(["partial_rotary_factor"])
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
@@ -4347,7 +4424,7 @@ class Phi2Model(TextModel):
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
@@ -4465,8 +4542,6 @@ class Phi3MiniModel(TextModel):
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
-        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
-
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
@@ -4480,7 +4555,7 @@ class Phi3MiniModel(TextModel):
         self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
@@ -4600,12 +4675,11 @@ class PlamoModel(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(4096)  # not in config.json
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(5)  # hparams["num_key_value_heads"]) is wrong
         self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
@@ -4728,7 +4802,6 @@ class Plamo2Model(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
         self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
 
         # Which layers are Mamba layers
@@ -4740,10 +4813,10 @@ class Plamo2Model(TextModel):
         num_attention_heads = []
 
         if mamba_enabled:
-            for i in range(block_count):
-                if block_count <= (mamba_step // 2):
+            for i in range(self.block_count):
+                if self.block_count <= (mamba_step // 2):
                     # use attention in last layer
-                    is_mamba = (i != block_count - 1)
+                    is_mamba = (i != self.block_count - 1)
                 else:
                     is_mamba = (i % mamba_step) != (mamba_step // 2)
                 if is_mamba:
@@ -4761,7 +4834,7 @@ class Plamo2Model(TextModel):
         self.gguf_writer.add_embedding_length(hparams.get("hidden_size", 4096))
         self.gguf_writer.add_key_length(hparams.get("hidden_size_per_head", 128))
         self.gguf_writer.add_value_length(hparams.get("hidden_size_per_head", 128))
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 10000))
 
@@ -4818,12 +4891,10 @@ class CodeShellModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CODESHELL
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["n_layer"]
-
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_query_groups"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -4965,7 +5036,7 @@ class InternLM2Model(TextModel):
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
@@ -5586,11 +5657,10 @@ class GemmaModel(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
@@ -5626,11 +5696,10 @@ class Gemma2Model(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
@@ -5674,12 +5743,11 @@ class Gemma3Model(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
 
         # some default values are not specified in the hparams
         self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 131072))
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 8))
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6))
@@ -5955,7 +6023,6 @@ class Rwkv6Model(TextModel):
         self._set_vocab_rwkv_world()
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         head_size = self.hparams["head_size"]
         hidden_size = self.hparams["hidden_size"]
         layer_norm_eps = self.hparams["layer_norm_epsilon"]
@@ -5967,7 +6034,7 @@ class Rwkv6Model(TextModel):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
         self.gguf_writer.add_rescale_every_n_layers(rescale_every_n_layers)
         self.gguf_writer.add_wkv_head_size(head_size)
@@ -6031,7 +6098,6 @@ class RWKV6Qwen2Model(Rwkv6Model):
             self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         num_attention_heads = self.hparams["num_attention_heads"]
         num_key_value_heads = self.hparams["num_key_value_heads"]
         hidden_size = self.hparams["hidden_size"]
@@ -6044,7 +6110,7 @@ class RWKV6Qwen2Model(Rwkv6Model):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_wkv_head_size(head_size)
         self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
         self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
@@ -6085,7 +6151,6 @@ class Rwkv7Model(TextModel):
         return max(1, round(hidden_size ** exponent * multiplier / 32)) * 32
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         try:
             head_size = self.hparams["head_size"]
             layer_norm_eps = self.hparams["layer_norm_epsilon"]
@@ -6110,7 +6175,7 @@ class Rwkv7Model(TextModel):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
         self.gguf_writer.add_wkv_head_size(head_size)
         self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
@@ -6204,7 +6269,6 @@ class ARwkv7Model(Rwkv7Model):
             self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
         hidden_size = self.hparams["hidden_size"]
         head_size = self.hparams["head_size"]
         rms_norm_eps = self.hparams["rms_norm_eps"]
@@ -6221,7 +6285,7 @@ class ARwkv7Model(Rwkv7Model):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
         self.gguf_writer.add_wkv_head_size(head_size)
         self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
@@ -7104,13 +7168,6 @@ class DeepseekV2Model(TextModel):
         self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
         self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
 
-        if hparams["scoring_func"] == "sigmoid":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        elif hparams["scoring_func"] == "softmax":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
-        else:
-            raise ValueError(f"Unsupported scoring_func value: {hparams['scoring_func']}")
-
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
         rope_scaling = self.hparams.get("rope_scaling") or {}
@@ -7216,12 +7273,6 @@ class MiniMaxM2Model(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if self.hparams["scoring_func"] == "sigmoid":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        elif self.hparams["scoring_func"] == "softmax":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
-        else:
-            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
 
         self.gguf_writer.add_expert_feed_forward_length(self.find_hparam(["intermediate_size"]))
         self.gguf_writer.add_rope_dimension_count(self.find_hparam(["rotary_dim"]))
@@ -7314,11 +7365,6 @@ class Dots1Model(Qwen2MoeModel):
         self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
         self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
 
-        if self.hparams["scoring_func"] == "noaux_tc":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        else:
-            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
@@ -7463,7 +7509,7 @@ class T5Model(TextModel):
         self.gguf_writer.add_context_length(n_ctx)
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
-        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_block_count(self.block_count)
         if (dec_n_layer := self.hparams.get("num_decoder_layers")) is not None:
             self.gguf_writer.add_decoder_block_count(dec_n_layer)
         self.gguf_writer.add_head_count(self.hparams["num_heads"])
@@ -7602,7 +7648,7 @@ class T5EncoderModel(TextModel):
         self.gguf_writer.add_context_length(n_ctx)
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
-        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(self.hparams["num_heads"])
         self.gguf_writer.add_key_length(self.hparams["d_kv"])
         self.gguf_writer.add_value_length(self.hparams["d_kv"])
@@ -7665,7 +7711,7 @@ class JaisModel(TextModel):
         self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"])
@@ -7779,12 +7825,6 @@ class Glm4MoeModel(TextModel):
         special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
         special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
 
-        # Patch broken chat template
-        if isinstance(special_vocab.chat_template, str) and "visible_text(m.content).endswith" in special_vocab.chat_template:
-            special_vocab.chat_template = special_vocab.chat_template.replace(
-                """{{ visible_text(m.content) }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not visible_text(m.content).endswith("/nothink")) else '' -}}""",
-                """{% set content = visible_text(m.content) %}{{ content }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not content.endswith("/nothink")) else '' -}}""")
-
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -8013,7 +8053,7 @@ class ChatGLMModel(TextModel):
         self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
         self.gguf_writer.add_embedding_length(n_embed)
         self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", self.hparams.get("intermediate_size", 4 * n_embed)))
-        self.gguf_writer.add_block_count(self.hparams.get("num_layers", self.hparams["num_hidden_layers"]))
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon",1e-5))
@@ -8095,7 +8135,6 @@ class ExaoneModel(TextModel):
         num_kv_heads = hparams.get("num_key_value_heads", num_heads)
         layer_norm_eps = hparams["layer_norm_epsilon"]
         intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim
-        num_layers = hparams["num_layers"]
         # ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0
         # attention_dropout_rate = hparams["attention_dropout"]
         # ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0
@@ -8106,7 +8145,7 @@ class ExaoneModel(TextModel):
         self.gguf_writer.add_context_length(max_position_embeddings)
         self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps)
         self.gguf_writer.add_feed_forward_length(intermediate_size)
-        self.gguf_writer.add_block_count(num_layers)
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_file_type(self.ftype)
 
         if (rope_theta := self.hparams.get("rope_theta")) is not None:
@@ -8639,13 +8678,6 @@ class BailingMoeV2Model(TextModel):
         self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
         self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
 
-        if hparams["score_function"] == "sigmoid":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        elif hparams["score_function"] == "softmax":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
-        else:
-            raise ValueError(f"Unsupported score_function value: {hparams['score_function']}")
-
         if (nextn_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(nextn_layers)
 
@@ -9341,16 +9373,6 @@ class HunYuanModel(TextModel):
 class SmolLM3Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.SMOLLM3
 
-    def set_vocab(self):
-        super().set_vocab()
-        # remove unsupported array slicing in chat template
-        # ref: https://huggingface.co/ggml-org/SmolLM3-3B-GGUF/discussions/1
-        from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
-        if tokenizer.chat_template is not None:
-            chat_template = tokenizer.chat_template.replace("[:]", "")
-            self.gguf_writer.add_chat_template(chat_template)
-
 
 @ModelBase.register("GptOssForCausalLM")
 class GptOssModel(TextModel):

convert_hf_to_gguf_update.py

@@ -139,6 +139,7 @@ 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": "afmoe",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/arcee-ai/Trinity-Tokenizer", },
     {"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", },

convert_lora_to_gguf.py

@@ -242,7 +242,7 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f32",
         help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
@@ -277,10 +277,15 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
-def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
+def load_hparams_from_hf(hf_model_id: str) -> tuple[dict[str, Any], Path | None]:
+    from huggingface_hub import try_to_load_from_cache
+
     # normally, adapter does not come with base model config, we need to load it from AutoConfig
     config = AutoConfig.from_pretrained(hf_model_id)
-    return config.to_dict()
+    cache_dir = try_to_load_from_cache(hf_model_id, "config.json")
+    cache_dir = Path(cache_dir).parent if isinstance(cache_dir, str) else None
+
+    return config.to_dict(), cache_dir
 
 
 if __name__ == '__main__':
@@ -325,13 +330,13 @@ if __name__ == '__main__':
     # load base model
     if base_model_id is not None:
         logger.info(f"Loading base model from Hugging Face: {base_model_id}")
-        hparams = load_hparams_from_hf(base_model_id)
+        hparams, dir_base_model = load_hparams_from_hf(base_model_id)
     elif dir_base_model is None:
         if "base_model_name_or_path" in lparams:
             model_id = lparams["base_model_name_or_path"]
             logger.info(f"Loading base model from Hugging Face: {model_id}")
             try:
-                hparams = load_hparams_from_hf(model_id)
+                hparams, dir_base_model = load_hparams_from_hf(model_id)
             except OSError as e:
                 logger.error(f"Failed to load base model config: {e}")
                 logger.error("Please try downloading the base model and add its path to --base")
@@ -480,6 +485,7 @@ if __name__ == '__main__':
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
             hparams=hparams,
+            remote_hf_model_id=base_model_id,
         )
 
         logger.info("Exporting model...")

docs/ops.md

@@ -14,103 +14,108 @@ Legend:
 
 | Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                           ADD_ID | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                           CUMSUM | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ❌ |
 |                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ❌ |
+|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
 |                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                              SUM | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
-|                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

examples/batched/README.md

@@ -3,7 +3,7 @@
 The example demonstrates batched generation from a given prompt
 
 ```bash
-./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
+./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4 --kv-unified
 
 ...
 

examples/diffusion/README.md

@@ -6,8 +6,54 @@ More Info:
 - https://github.com/ggml-org/llama.cpp/pull/14644
 - https://github.com/ggml-org/llama.cpp/pull/14771
 
+## Parameters
+The diffusion CLI supports various parameters to control the generation process:
 
-Example of using Dream architechture: `llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual`
+### Core Diffusion Parameters
+- `--diffusion-steps`: Number of diffusion steps (default: 256)
+- `--diffusion-algorithm`: Algorithm for token selection
+  - `0`: ORIGIN - Token will be generated in a purely random order from https://arxiv.org/abs/2107.03006.
+  - `1`: ENTROPY_BASED - Entropy-based selection
+  - `2`: MARGIN_BASED - Margin-based selection
+  - `3`: RANDOM - Random selection
+  - `4`: CONFIDENCE_BASED - Confidence-based selection (default)
+  - More documentation here https://github.com/DreamLM/Dream
+- `--diffusion-visual`: Enable live visualization during generation
 
-Example of using LLaDA architechture: `llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual`
+### Scheduling Parameters
+Choose one of the following scheduling methods:
 
+**Timestep-based scheduling:**
+- `--diffusion-eps`: Epsilon value for timestep scheduling (e.g., 0.001)
+
+**Block-based scheduling:**
+- `--diffusion-block-length`: Block size for block-based scheduling (e.g., 32)
+
+### Sampling Parameters
+- `--temp`: Temperature for sampling (0.0 = greedy/deterministic, higher = more random)
+- `--top-k`: Top-k filtering for sampling
+- `--top-p`: Top-p (nucleus) filtering for sampling
+- `--seed`: Random seed for reproducibility
+
+### Model Parameters
+- `-m`: Path to the GGUF model file
+- `-p`: Input prompt text
+- `-ub`: Maximum sequence length (ubatch size)
+- `-c`: Context size
+- `-b`: Batch size
+
+### Examples
+#### Dream architechture:
+```
+llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual
+```
+
+#### LLaDA architechture:
+```
+llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual
+```
+
+#### RND1 architecture:
+```
+llama-diffusion-cli -m RND1-Base-0910.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-algorithm 1 --diffusion-steps 256 --diffusion-visual --temp 0.5 --diffusion-eps 0.001
+```

examples/eval-callback/eval-callback.cpp

@@ -4,10 +4,10 @@
 #include "llama.h"
 #include "ggml.h"
 
+#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
-#include <numeric>
 
 /**
  * This the arbitrary data which will be passed to each callback.
@@ -37,23 +37,23 @@ static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
     return u.f;
 }
 
-static float ggml_get_float_value(uint8_t * data, ggml_type type, const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) {
+static float ggml_get_float_value(const uint8_t * data, ggml_type type, const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) {
     size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
     float v;
     if (type == GGML_TYPE_F16) {
-        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
+        v = ggml_fp16_to_fp32(*(const ggml_fp16_t *) &data[i]);
     } else if (type == GGML_TYPE_F32) {
-        v = *(float *) &data[i];
+        v = *(const float *) &data[i];
     } else if (type == GGML_TYPE_I64) {
-        v = (float) *(int64_t *) &data[i];
+        v = (float) *(const int64_t *) &data[i];
     } else if (type == GGML_TYPE_I32) {
-        v = (float) *(int32_t *) &data[i];
+        v = (float) *(const int32_t *) &data[i];
     } else if (type == GGML_TYPE_I16) {
-        v = (float) *(int16_t *) &data[i];
+        v = (float) *(const int16_t *) &data[i];
     } else if (type == GGML_TYPE_I8) {
-        v = (float) *(int8_t *) &data[i];
+        v = (float) *(const int8_t *) &data[i];
     } else if (type == GGML_TYPE_BF16) {
-        v = ggml_compute_bf16_to_fp32(*(ggml_bf16_t *) &data[i]);
+        v = ggml_compute_bf16_to_fp32(*(const ggml_bf16_t *) &data[i]);
     } else {
         GGML_ABORT("fatal error");
     }

ggml/CMakeLists.txt

@@ -25,16 +25,17 @@ if(GIT_EXE)
     )
 endif()
 
-# Build the version string with optional dirty flag
 set(GGML_VERSION "${GGML_VERSION_BASE}")
-if(GGML_GIT_DIRTY AND NOT GGML_GIT_DIRTY EQUAL 0)
-    set(GGML_VERSION "${GGML_VERSION}-dirty")
-endif()
 
 if(NOT GGML_BUILD_COMMIT)
     set(GGML_BUILD_COMMIT "unknown")
 endif()
 
+# Build the commit string with optional dirty flag
+if(DEFINED GGML_GIT_DIRTY AND GGML_GIT_DIRTY EQUAL 1)
+    set(GGML_BUILD_COMMIT "${GGML_BUILD_COMMIT}-dirty")
+endif()
+
 include(CheckIncludeFileCXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)

ggml/include/ggml.h

@@ -475,6 +475,7 @@ extern "C" {
         GGML_OP_COS,
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
+        GGML_OP_CUMSUM,
         GGML_OP_MEAN,
         GGML_OP_ARGMAX,
         GGML_OP_COUNT_EQUAL,
@@ -530,6 +531,8 @@ extern "C" {
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
         GGML_OP_LEAKY_RELU,
+        GGML_OP_TRI,
+        GGML_OP_FILL,
 
         GGML_OP_FLASH_ATTN_EXT,
         GGML_OP_FLASH_ATTN_BACK,
@@ -542,6 +545,7 @@ extern "C" {
         GGML_OP_RWKV_WKV6,
         GGML_OP_GATED_LINEAR_ATTN,
         GGML_OP_RWKV_WKV7,
+        GGML_OP_SOLVE_TRI,
 
         GGML_OP_UNARY,
 
@@ -576,6 +580,8 @@ extern "C" {
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
+        GGML_UNARY_OP_EXPM1,
+        GGML_UNARY_OP_SOFTPLUS,
         GGML_UNARY_OP_GELU_ERF,
         GGML_UNARY_OP_XIELU,
         GGML_UNARY_OP_FLOOR,
@@ -620,6 +626,13 @@ extern "C" {
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
+    enum ggml_tri_type {
+        GGML_TRI_TYPE_UPPER_DIAG = 0,
+        GGML_TRI_TYPE_UPPER      = 1,
+        GGML_TRI_TYPE_LOWER_DIAG = 2,
+        GGML_TRI_TYPE_LOWER      = 3
+    };
+
     struct ggml_init_params {
         // memory pool
         size_t mem_size;   // bytes
@@ -957,6 +970,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_expm1(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_expm1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_softplus(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_softplus_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_sin(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -983,6 +1012,10 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_cumsum(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
     // mean along rows
     GGML_API struct ggml_tensor * ggml_mean(
             struct ggml_context * ctx,
@@ -2187,6 +2220,23 @@ extern "C" {
             int                   shift2,
             int                   shift3);
 
+    // Convert matrix into a triangular one (upper, strict upper, lower or strict lower) by writing
+    // zeroes everywhere outside the masked area
+    GGML_API struct ggml_tensor * ggml_tri(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_tri_type    type);
+
+    // Fill tensor a with constant c
+    GGML_API struct ggml_tensor * ggml_fill(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 c);
+
+    GGML_API struct ggml_tensor * ggml_fill_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 c);
 
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
@@ -2356,6 +2406,27 @@ extern "C" {
             struct ggml_tensor  * b,
             struct ggml_tensor  * state);
 
+    /* Solves a specific equation of the form Ax=B, where A is a triangular matrix
+    *  without zeroes on the diagonal (i.e. invertible).
+    *  B can have any number of columns, but must have the same number of rows as A
+    *  If A is [n, n] and B is [n, m], then the result will be [n, m] as well
+    *  Has O(n^3) complexity (unlike most matrix ops out there), so use on cases
+    *  where n > 100 sparingly, pre-chunk if necessary.
+    *
+    *  If left = false, solves xA=B instead
+    *  If lower = false, assumes upper triangular instead
+    *  If uni = true, assumes diagonal of A to be all ones (will override actual values)
+    *
+    *  TODO: currently only lower, right, non-unitriangular variant is implemented
+    */
+    GGML_API struct ggml_tensor * ggml_solve_tri(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  left,
+        bool                  lower,
+        bool                  uni);
+
     // custom operators
 
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);

ggml/src/CMakeLists.txt

@@ -328,6 +328,14 @@ function(ggml_add_cpu_backend_variant tag_name)
             set(GGML_INTERNAL_${feat} OFF)
         endforeach()
 
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        foreach (feat RVV)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
@@ -402,6 +410,13 @@ if (GGML_CPU_ALL_VARIANTS)
         else()
             message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(riscv64_0)
+            ggml_add_cpu_backend_variant(riscv64_v   RVV)
+        else()
+            message(FATAL_ERROR "Unsupported RISC-V 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()

ggml/src/ggml-backend.cpp

@@ -1698,8 +1698,6 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
     GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
-    ggml_backend_sched_reset(sched);
-
     ggml_backend_sched_synchronize(sched);
 
     ggml_backend_sched_split_graph(sched, measure_graph);

ggml/src/ggml-cann/acl_tensor.cpp

@@ -48,15 +48,14 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
         default:
             return ACL_DT_UNDEFINED;
     }
-    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) {
+acl_tensor_ptr 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];
@@ -87,10 +86,20 @@ aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
     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 * raw = 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;
+    return acl_tensor_ptr(raw);
+}
+
+acl_int_array_ptr ggml_cann_create_int_array(const int64_t * value, uint64_t size) {
+    aclIntArray * raw = aclCreateIntArray(value, size);
+    return acl_int_array_ptr(raw);
+}
+
+acl_scalar_ptr ggml_cann_create_scalar(void * value, aclDataType dataType) {
+    aclScalar * raw = aclCreateScalar(value, dataType);
+    return acl_scalar_ptr(raw);
 }
 
 bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1) {

ggml/src/ggml-cann/acl_tensor.h

@@ -23,11 +23,12 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H
 
-#include <algorithm>
-#include <cstring>
+#include "common.h"
 
 #include <aclnn/aclnn_base.h>
-#include "common.h"
+
+#include <algorithm>
+#include <cstring>
 
 /**
  * @brief	Maps a ggml_type to its corresponding aclDataType.
@@ -43,6 +44,20 @@
  */
 aclDataType ggml_cann_type_mapping(ggml_type type);
 
+// Deleter for acl objects.
+template <typename T, aclError (*DestroyFunc)(const T *)> struct acl_deleter {
+    void operator()(T * ptr) const noexcept {
+        if (ptr) {
+            ACL_CHECK(DestroyFunc(ptr));
+        }
+    }
+};
+
+using acl_tensor_ptr      = std::unique_ptr<aclTensor, acl_deleter<aclTensor, aclDestroyTensor>>;
+using acl_int_array_ptr   = std::unique_ptr<aclIntArray, acl_deleter<aclIntArray, aclDestroyIntArray>>;
+using acl_scalar_ptr      = std::unique_ptr<aclScalar, acl_deleter<aclScalar, aclDestroyScalar>>;
+using acl_tensor_list_ptr = std::unique_ptr<aclTensorList, acl_deleter<aclTensorList, aclDestroyTensorList>>;
+
 /**
  * @brief   Creates an ACL tensor from a ggml_tensor with optional shape.
  *
@@ -62,12 +77,12 @@ 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);
+acl_tensor_ptr 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
@@ -90,14 +105,14 @@ aclTensor * ggml_cann_create_tensor(const ggml_tensor * tensor,
  * @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) {
+acl_tensor_ptr 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];
 
@@ -114,10 +129,75 @@ aclTensor * ggml_cann_create_tensor(void *      data_ptr,
     std::reverse(tmp_ne, tmp_ne + dims);
     std::reverse(tmp_stride, tmp_stride + dims);
 
-    aclTensor * acl_tensor =
+    aclTensor * raw =
         aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size, format, &acl_storage_len, 1, data_ptr);
 
-    return acl_tensor;
+    return acl_tensor_ptr(raw);
+}
+
+/**
+ * @brief Create an ACL int array resource wrapped in a smart pointer.
+ *
+ * This function constructs an aclIntArray from the provided int64_t values
+ * and returns it as an acl_int_array_ptr (a std::unique_ptr with a custom
+ * deleter). The returned pointer owns the ACL resource and will automatically
+ * destroy it via aclDestroyIntArray().
+ *
+ * @param value  Pointer to the int64_t elements.
+ * @param size   Number of elements in value.
+ *
+ * @return A smart pointer managing the created ACL int array.
+ */
+acl_int_array_ptr ggml_cann_create_int_array(const int64_t * value, uint64_t size);
+
+/**
+ * @brief Create an ACL scalar resource wrapped in a smart pointer.
+ *
+ * This function constructs an aclScalar from the raw value pointer and ACL
+ * data type, then returns it as an acl_scalar_ptr (a std::unique_ptr with
+ * a custom deleter). The returned pointer owns the ACL scalar and will
+ * automatically destroy it via aclDestroyScalar().
+ *
+ * @param value     Pointer to the raw scalar memory.
+ * @param dataType  ACL data type of the scalar.
+ *
+ * @return A smart pointer managing the created ACL scalar.
+ */
+acl_scalar_ptr ggml_cann_create_scalar(void * value, aclDataType dataType);
+
+/**
+ * @brief Create an ACL tensor list from multiple tensor smart pointers.
+ *
+ * This function accepts a variadic list of acl_tensor_ptr (a unique_ptr with
+ * custom deleter) and produces an aclTensorList using aclCreateTensorList().
+ *
+ * The lifecycle management of the tensor objects changes as follows:
+ *  - aclCreateTensorList() takes ownership of the tensors
+ *  - Each input smart pointer releases ownership using release()
+ *  - As a result, the tensors will NOT be destroyed by unique_ptr
+ *  - Instead, they will be destroyed when aclDestroyTensorList() is called
+ *
+ * This ensures correct ownership transfer and prevents double-free situations.
+ *
+ * @param acl_tensor_ptr  Variadic template parameter; each argument must be
+ *                         a unique_ptr-like type supporting get() and release().
+ *
+ * @param tensors  Variadic list of acl_tensor_ptr objects. Ownership of
+ *                         each tensor is transferred away from these smart pointers.
+ *
+ * @return A smart pointer (acl_tensor_list_ptr) owning the created ACL tensor list.
+ *
+ * @note This implementation is C++11 compatible. The ownership-release process is
+ *       executed using a pack expansion inside an initializer list.
+ */
+template <typename... acl_tensor_ptr> acl_tensor_list_ptr ggml_cann_create_tensor_list(acl_tensor_ptr &&... tensors) {
+    aclTensor *     raw_tensors[] = { tensors.get()... };
+    aclTensorList * raw           = aclCreateTensorList(raw_tensors, sizeof...(tensors));
+    // aclTensor will release by aclTensorList, so release ownership without
+    // destroying the tensor
+    int             dummy[]       = { (tensors.release(), 0)... };
+    GGML_UNUSED(dummy);
+    return acl_tensor_list_ptr(raw);
 }
 
 /**

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,33 +23,35 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
-#include <unordered_set>
-#include <functional>
+#include "acl_tensor.h"
+#include "common.h"
+
 #include <aclnnop/aclnn_abs.h>
-#include <aclnnop/aclnn_neg.h>
-#include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_arange.h>
 #include <aclnnop/aclnn_argsort.h>
 #include <aclnnop/aclnn_cat.h>
 #include <aclnnop/aclnn_clamp.h>
+#include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_gelu.h>
 #include <aclnnop/aclnn_gelu_v2.h>
-#include <aclnnop/aclnn_sigmoid.h>
 #include <aclnnop/aclnn_hardsigmoid.h>
 #include <aclnnop/aclnn_hardswish.h>
 #include <aclnnop/aclnn_leaky_relu.h>
-#include <aclnnop/aclnn_relu.h>
-#include <aclnnop/aclnn_silu.h>
-#include <aclnnop/aclnn_tanh.h>
-#include <aclnnop/aclnn_sqrt.h>
-#include <aclnnop/aclnn_sin.h>
-#include <aclnnop/aclnn_cos.h>
 #include <aclnnop/aclnn_log.h>
-#include <aclnnop/aclnn_sign.h>
-#include <aclnnop/aclnn_norm.h>
 #include <aclnnop/aclnn_logsoftmax.h>
-#include "acl_tensor.h"
-#include "common.h"
+#include <aclnnop/aclnn_neg.h>
+#include <aclnnop/aclnn_norm.h>
+#include <aclnnop/aclnn_relu.h>
+#include <aclnnop/aclnn_sigmoid.h>
+#include <aclnnop/aclnn_sign.h>
+#include <aclnnop/aclnn_silu.h>
+#include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_sqrt.h>
+#include <aclnnop/aclnn_tanh.h>
+
+#include <functional>
+#include <unordered_set>
 
 /**
  * @brief   Repeats a ggml tensor along each dimension to match the dimensions
@@ -688,12 +690,12 @@ void aclnn_sin(ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor *
  * @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,
+                 acl_tensor_ptr & acl_src0,
+                 acl_tensor_ptr & acl_src1,
+                 acl_tensor_ptr & acl_dst);
 
 /**
  * @brief   Computes the 1D transposed convolution (deconvolution) of a ggml
@@ -873,83 +875,6 @@ template <typename... Args> void register_acl_resources(std::vector<any_acl_reso
     (vec.emplace_back(make_acl_resource(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_;
-};
-
-/**
- * @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); }
-
-    virtual void run_task() override { resource_.clear(); }
-  private:
-    std::vector<any_acl_resource> resource_;
-};
-
-/**
- * @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) {}
-
-    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_;
-};
-
-/**
- * @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) {}
-
-    virtual void run_task() override { ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_)); }
-  private:
-    void *      buffer_;
-    size_t      size_;
-    int32_t     value_;
-    aclrtStream stream_;
-};
-
 /**
  * @brief Launches an asynchronous task using the memory allocator.
  *
@@ -968,95 +893,20 @@ 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();                                       \
+        }                                                                                    \
+        ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));     \
     } while (0)
 
-/**
- * @brief Registers and releases multiple ACL resources, optionally deferring the release
- *        using a task.
- *
- * @tparam Args Types of the ACL resources.
- * @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) {
-    std::vector<any_acl_resource> resources;
-    register_acl_resources(resources, std::forward<Args>(args)...);
-    if (ctx.async_mode) {
-        auto task = std::make_unique<release_resource_task>(std::move(resources));
-        ctx.task_queue.submit_task(std::move(task));
-    }
-}
-
-/**
- * @brief Performs an asynchronous memory copy operation, optionally deferred via task submission.
- *
- * @param ctx Backend context containing stream and async configuration.
- * @param dst Destination memory address.
- * @param src Source memory address.
- * @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) {
-    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));
-    } else {
-        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx.stream()));
-    }
-}
-
-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));
-    } else {
-        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx->stream()));
-    }
-}
-
-/**
- * @brief Performs an asynchronous memory set operation, optionally deferred via task submission.
- *
- * @param ctx Backend context containing stream and async configuration.
- * @param buffer Memory buffer to be set.
- * @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) {
-    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));
-    } else {
-        ACL_CHECK(aclrtMemsetAsync(buffer, size, value, size, ctx.stream()));
-    }
-}
-
 /**
  * @brief   Performs sparse expert-based matrix multiplication using the CANN backend.
  *
@@ -1129,15 +979,11 @@ template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & c
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
 
-    aclTensor * acl_src0;
-    aclTensor * acl_src1;
-    aclTensor * acl_dst;
+    acl_tensor_ptr acl_src0, acl_src1, acl_dst;
 
     // Need bcast
-    bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
-    binary_op(ctx, acl_src0, acl_src1, acl_dst);
-
-    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
+    bcast_shape(src0, src1, dst, acl_src0, acl_src1, acl_dst);
+    binary_op(ctx, acl_src0.get(), acl_src1.get(), acl_dst.get());
 }
 
 /**
@@ -1147,7 +993,7 @@ template <auto binary_op> void ggml_cann_binary_op(ggml_backend_cann_context & c
  * and stores the result in the destination tensor.
  *
  * @tparam unary_op A callable with the signature:
- *         void(ggml_backend_cann_context&, aclTensor*, aclTensor*)
+ *         void(ggml_backend_cann_context&, aclTensor *, aclTensor *)
  *         where the first aclTensor is the source and the second is the destination.
  * @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.
@@ -1156,11 +1002,10 @@ 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);
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
 
-    unary_op(ctx, acl_src, acl_dst);
-    ggml_cann_release_resources(ctx, acl_src, acl_dst);
+    unary_op(ctx, acl_src.get(), acl_dst.get());
 }
 
 /**

ggml/src/ggml-cann/common.h

@@ -23,26 +23,26 @@
 #ifndef CANN_COMMON_H
 #define CANN_COMMON_H
 
-#include <acl/acl.h>
-
-#include <cstdio>
-#include <iostream>
-#include <map>
-#include <memory>
-#include <string>
-#include <vector>
-#include <atomic>
-#include <condition_variable>
-#include <mutex>
-#include <thread>
-#include <unistd.h>
-#include <functional>
-#include <optional>
-#include <list>
-
+#include "../ggml-impl.h"
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
-#include "../ggml-impl.h"
+
+#include <acl/acl.h>
+#include <unistd.h>
+
+#include <atomic>
+#include <condition_variable>
+#include <cstdio>
+#include <functional>
+#include <iostream>
+#include <list>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <optional>
+#include <string>
+#include <thread>
+#include <vector>
 
 #define MATRIX_ROW_PADDING    512
 #define GGML_CANN_MAX_STREAMS 8
@@ -214,130 +214,6 @@ struct ggml_cann_pool_alloc {
     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);
-
-/**
- * @brief Base class for all CANN tasks to be submitted to the task queue.
- *
- * Users should override the run_task() method with actual task logic.
- */
-class cann_task {
-  public:
-    virtual void run_task() {}
-};
-
-/**
- * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
- */
-class cann_task_queue {
-  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) {
-        GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
-        mask_ = capacity_ - 1;
-    }
-
-    /**
-     * @brief Attempts to enqueue a task into the queue.
-     *
-     * @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) {
-        size_t next_tail = (tail_ + 1) & mask_;
-
-        if (next_tail == head_) {
-            return false;
-        }
-
-        buffer_[tail_] = std::move(item);
-        std::atomic_thread_fence(std::memory_order_release);
-        tail_ = next_tail;
-
-        return true;
-    }
-
-    /**
-     * @brief Submits a task to the queue, and starts the worker thread if not already running.
-     *
-     * @param task Task to be submitted.
-     */
-    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);
-        }
-    }
-
-    /**
-     * @brief Waits until the queue is completely empty and no tasks are being processed.
-     */
-    void wait() {
-        while (running_ && head_ != tail_) {
-            std::this_thread::yield();
-            continue;
-        }
-    }
-
-    /**
-     * @brief Stops the task queue and joins the worker thread.
-     */
-    void stop() {
-        running_ = false;
-        if (thread_.joinable()) {
-            thread_.join();
-        }
-    }
-
-  private:
-    /**
-     * @brief Worker thread function that continuously dequeues and executes tasks.
-     */
-    void execute() {
-        ggml_cann_set_device(device_);
-
-        while (running_) {
-            if (head_ == tail_) {
-                std::this_thread::yield();
-                continue;
-            }
-
-            std::atomic_thread_fence(std::memory_order_acquire);
-            buffer_[head_]->run_task();
-            buffer_[head_].reset();
-            head_ = (head_ + 1) & mask_;
-        }
-    }
-
-    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_;
-};
-
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
     // dst tensor
@@ -474,7 +350,6 @@ struct ggml_backend_cann_context {
     ggml_cann_graph_lru_cache graph_lru_cache;
     bool                      acl_graph_mode = true;
 #endif
-    cann_task_queue        task_queue;
     bool                   async_mode;
     // Rope Cache
     ggml_cann_rope_cache   rope_cache;
@@ -488,15 +363,10 @@ struct ggml_backend_cann_context {
      * @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)) {
         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");
 #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",
@@ -509,7 +379,6 @@ struct ggml_backend_cann_context {
      */
     ~ggml_backend_cann_context() {
         ggml_cann_set_device(device);
-        task_queue.stop();
         if (copy_event != nullptr) {
             ACL_CHECK(aclrtDestroyEvent(copy_event));
         }

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

@@ -22,24 +22,24 @@
 
 #include "ggml-cann.h"
 
-#include <acl/acl.h>
-#include <stdarg.h>
-#include <aclnnop/aclnn_trans_matmul_weight.h>
+#include "ggml-backend-impl.h"
+#include "ggml-cann/aclnn_ops.h"
+#include "ggml-cann/common.h"
+#include "ggml-impl.h"
+#include "ggml.h"
 
+#include <acl/acl.h>
+#include <aclnnop/aclnn_trans_matmul_weight.h>
+#include <stdarg.h>
+
+#include <chrono>
 #include <cmath>
 #include <cstdio>
 #include <cstring>
 #include <mutex>
-#include <queue>
-#include <chrono>
-#include <unordered_set>
 #include <optional>
-
-#include "ggml-impl.h"
-#include "ggml-backend-impl.h"
-#include "ggml-cann/aclnn_ops.h"
-#include "ggml-cann/common.h"
-#include "ggml.h"
+#include <queue>
+#include <unordered_set>
 
 #define GGML_COMMON_DECL_C
 
@@ -1177,19 +1177,18 @@ static ggml_cann_nz_workspace g_nz_workspaces[GGML_CANN_MAX_DEVICES];
  *       across calls. This reduces overhead from repeated memory allocation and deallocation.
  */
 static void weight_format_to_nz(ggml_tensor * tensor, size_t offset, int device) {
-    aclTensor * weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne, tensor->nb, 2, ACL_FORMAT_ND, offset);
-    uint64_t    workspaceSize    = 0;
+    acl_tensor_ptr weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne, tensor->nb, 2, ACL_FORMAT_ND, offset);
+    uint64_t       workspaceSize    = 0;
     aclOpExecutor * executor;
 
     // TransMatmulWeight
-    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed, &workspaceSize, &executor));
+    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed.get(), &workspaceSize, &executor));
     // Avoid frequent malloc/free of the workspace.
     g_nz_workspaces[device].realloc(workspaceSize);
 
     void * g_nz_workspace = g_nz_workspaces[device].get();
 
     ACL_CHECK(aclnnTransMatmulWeight(g_nz_workspace, workspaceSize, executor, nullptr));
-    ACL_CHECK(aclDestroyTensor(weightTransposed));
 }
 
 // TODO: need handle tensor which has paddings.
@@ -1641,7 +1640,7 @@ ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
                            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
                            },
         /* .device   = */
-         ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
+        ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
         /* .context  = */ nullptr,
     };
 
@@ -1949,7 +1948,8 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
     GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(!ggml_is_quantized(tensor->type));
 
-    ggml_cann_async_memcpy(cann_ctx, (char *) tensor->data + offset, data, size, ACL_MEMCPY_HOST_TO_DEVICE);
+    ACL_CHECK(aclrtMemcpyAsync((char *) tensor->data + offset, size, data, size, ACL_MEMCPY_HOST_TO_DEVICE,
+                               cann_ctx->stream()));
 }
 
 /**
@@ -1974,7 +1974,8 @@ static void ggml_backend_cann_get_tensor_async(ggml_backend_t      backend,
     GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(!ggml_is_quantized(tensor->type));
 
-    ggml_cann_async_memcpy(cann_ctx, data, (char *) tensor->data + offset, size, ACL_MEMCPY_DEVICE_TO_HOST);
+    ACL_CHECK(aclrtMemcpyAsync(data, size, (char *) tensor->data + offset, size, ACL_MEMCPY_DEVICE_TO_HOST,
+                               cann_ctx->stream()));
 }
 
 /**
@@ -2035,7 +2036,6 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t      backend_src,
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
 
         // wait for task_queue empty to keep task order.
-        cann_ctx_src->task_queue.wait();
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size, ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
         // record event on src stream after the copy
@@ -2068,7 +2068,6 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t      backend_src,
  */
 static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ggml_backend_cann_context * cann_ctx = (ggml_backend_cann_context *) backend->context;
-    cann_ctx->task_queue.wait();
     ggml_cann_set_device(cann_ctx->device);
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
@@ -2247,8 +2246,7 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
                                             bool &                      use_cann_graph,
                                             bool &                      cann_graph_update_required) {
 #ifdef USE_ACL_GRAPH
-    ggml_cann_graph * matched_graph = cann_ctx->graph_lru_cache.cache_list.front();
-    if (use_cann_graph && cann_graph_update_required) {
+    if (use_cann_graph && cann_graph_update_required) {  // Begin CANN graph capture
         ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
     }
 #endif  // USE_ACL_GRAPH
@@ -2272,12 +2270,14 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
     }
 
 #ifdef USE_ACL_GRAPH
-    if (use_cann_graph && cann_graph_update_required) {  // End CANN graph capture
-        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &matched_graph->graph));
-    }
-
     if (use_cann_graph) {
-        // Execute graph
+        ggml_cann_graph * matched_graph = cann_ctx->graph_lru_cache.cache_list.front();
+
+        if (cann_graph_update_required) {  // End CANN graph capture
+            ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &matched_graph->graph));
+        }
+
+        // Execute CANN graph
         ACL_CHECK(aclmdlRIExecuteAsync(matched_graph->graph, cann_ctx->stream()));
     }
 #endif  // USE_ACL_GRAPH
@@ -2303,9 +2303,9 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
     // calculate rope cache for fist layer in current device.
     cann_ctx->rope_cache.cached = false;
 
+    bool cann_graph_update_required = false;
 #ifdef USE_ACL_GRAPH
     bool use_cann_graph             = true;
-    bool cann_graph_update_required = false;
 
     static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
     if (!prefill_use_graph) {
@@ -2336,7 +2336,6 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
     }
 #else
     bool use_cann_graph             = false;
-    bool cann_graph_update_required = false;
 #endif  // USE_ACL_GRAPH
     evaluate_and_capture_cann_graph(cann_ctx, cgraph, use_cann_graph, cann_graph_update_required);
 
@@ -2485,6 +2484,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 if (mode & GGML_ROPE_TYPE_VISION) {
                     return false;
                 }
+                if (op->src[0]->ne[0] > 896) {
+                    return false;
+                }
 #ifdef ASCEND_310P
                 if (!ggml_is_contiguous(op->src[0])) {
                     return false;
@@ -2521,10 +2523,11 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 // value of paddingW should be at most half of kernelW
                 return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
             }
+        case GGML_OP_SUM:
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_L2_NORM:
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_DUP:
-        case GGML_OP_SUM:
         case GGML_OP_IM2COL:
         case GGML_OP_CONCAT:
         case GGML_OP_REPEAT:

ggml/src/ggml-cpu/CMakeLists.txt

@@ -145,26 +145,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
                 include(CheckCXXSourceRuns)
 
-                function(check_arm_feature tag code)
+                macro(check_arm_feature tag feature code)
                     set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
                     set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
                     check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
                     if (GGML_MACHINE_SUPPORTS_${tag})
-                        set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}" PARENT_SCOPE)
+                        set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}")
                     else()
                         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_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}" PARENT_SCOPE)
+                            set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}")
+                            list(APPEND ARCH_FLAGS -U__ARM_FEATURE_${feature})
                         endif()
                     endif()
                     set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
-                endfunction()
+                endmacro()
 
-                check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
-                check_arm_feature(i8mm    "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
-                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; }")
+                check_arm_feature(dotprod DOTPROD     "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
+                check_arm_feature(i8mm    MATMUL_INT8 "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
+                check_arm_feature(sve     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     SME         "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
 
                 list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
             else()
@@ -216,35 +217,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 endif()
             endif()
 
-            # show enabled features
-            if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
-                set(FEAT_INPUT_FILE "NUL")
-            else()
-                set(FEAT_INPUT_FILE "/dev/null")
-            endif()
+            message(STATUS "Checking for ARM features using flags:")
+            foreach(flag IN LISTS ARCH_FLAGS)
+                message(STATUS "  ${flag}")
+            endforeach()
 
-            execute_process(
-                COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
-                INPUT_FILE ${FEAT_INPUT_FILE}
-                OUTPUT_VARIABLE ARM_FEATURE
-                RESULT_VARIABLE ARM_FEATURE_RESULT
-            )
-            if (ARM_FEATURE_RESULT)
-                message(WARNING "Failed to get ARM features")
-            else()
-                foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
-                    string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
-                    if (NOT ${feature_pos} EQUAL -1)
-                        # Special handling for MATMUL_INT8 when machine doesn't support i8mm
-                        if ("${feature}" STREQUAL "MATMUL_INT8" AND GGML_MACHINE_SUPPORTS_noi8mm)
-                            message(STATUS "ARM feature ${feature} detected but unsetting due to machine not supporting i8mm")
-                            list(APPEND ARCH_FLAGS -U__ARM_FEATURE_MATMUL_INT8)
-                        else()
-                            message(STATUS "ARM feature ${feature} enabled")
-                        endif()
-                    endif()
-                endforeach()
-            endif()
+            include(CheckCXXSourceCompiles)
+            set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS}")
+            foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
+                set(ARM_FEATURE "HAVE_${feature}")
+                check_cxx_source_compiles(
+                    "
+                    #if !defined(__ARM_FEATURE_${feature})
+                    #  error \"Feature ${feature} is not defined\"
+                    #endif
+                    int main() { return 0; }
+                    "
+                    ${ARM_FEATURE}
+                )
+            endforeach()
+            set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
@@ -399,9 +392,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
-                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+                list(APPEND ARCH_FLAGS -mcpu=power10)
             elseif (EXTRACTED_NUMBER EQUAL 9)
-                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
+                list(APPEND ARCH_FLAGS -mcpu=power9)
             elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
                 list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
             else()
@@ -459,22 +452,35 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ggml-cpu/spacemit/ime_kernels.h
             )
         endif()
-        set(MARCH_STR "rv64gc")
-        if (GGML_RV_ZFH)
-            string(APPEND MARCH_STR "_zfh")
-        endif()
-        if (GGML_XTHEADVECTOR)
-            string(APPEND MARCH_STR "_xtheadvector")
-        elseif (GGML_RVV)
-            string(APPEND MARCH_STR "_v")
-            if (GGML_RV_ZVFH)
-                string(APPEND MARCH_STR "_zvfh")
+        if(NOT GGML_CPU_ALL_VARIANTS)
+            set(MARCH_STR "rv64gc")
+            if (GGML_RV_ZFH)
+                string(APPEND MARCH_STR "_zfh")
             endif()
+            if (GGML_XTHEADVECTOR)
+                string(APPEND MARCH_STR "_xtheadvector")
+            elseif (GGML_RVV)
+                string(APPEND MARCH_STR "_v")
+                if (GGML_RV_ZVFH)
+                    string(APPEND MARCH_STR "_zvfh")
+                endif()
+            endif()
+            if (GGML_RV_ZICBOP)
+                string(APPEND MARCH_STR "_zicbop")
+            endif()
+            list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
+        else()
+            # Begin with the lowest baseline
+            set(ARCH_DEFINITIONS "")
+
+            if (GGML_INTERNAL_RVV)
+                message(STATUS "RVV enabled")
+                list(APPEND ARCH_DEFINITIONS GGML_USE_RVV)
+                list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d)
+            endif()
+
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS})
         endif()
-        if (GGML_RV_ZICBOP)
-            string(APPEND MARCH_STR "_zicbop")
-        endif()
-        list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
         list(APPEND GGML_CPU_SOURCES

ggml/src/ggml-cpu/arch-fallback.h

@@ -51,10 +51,8 @@
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -24,6 +24,29 @@
 
 #define UNUSED GGML_UNUSED
 
+static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
+                                             int16x8_t *     out_mins,
+                                             int8_t *        out_scales) {
+    constexpr uint32_t kmask1 = 0x3f3f3f3f;
+    constexpr uint32_t kmask2 = 0x0f0f0f0f;
+    constexpr uint32_t kmask3 = 0x03030303;
+    constexpr uint8_t  scales_size = 12;
+
+    uint32_t sm[3];
+    memcpy(sm, scales_in, scales_size);
+
+    const uint32_t   mins_0_3 = sm[1] & kmask1;
+    const uint32_t   mins_4_7 = ((sm[2] >> 4) & kmask2) | (((sm[1] >> 6) & kmask3) << 4);
+    const uint32x2_t mins_u32 = { mins_0_3, mins_4_7 };
+
+    *out_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins_u32)));
+
+    uint32_t scales_u32[2];
+    scales_u32[0] = sm[0] & kmask1;
+    scales_u32[1] = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
+    memcpy(out_scales, scales_u32, 8);
+}
+
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -474,6 +497,162 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_q4_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[ncols_interleaved / 4];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < ncols_interleaved / 4; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_1   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            // 2 sb each iteration
+            int32x4_t acc_lo[col_pairs];
+            int32x4_t acc_hi[col_pairs];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_pairs; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
+
+                // Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
+                // but still need the qs to use the low and hi bits from q4
+                const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
+                int8x16_t      q8_qs[8];
+                for (int i = 0; i < 8; i++) {
+                    q8_qs[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base + i * 8));
+                }
+
+                // Q4s columns iterated in pairs (01, 23, 45, 67)
+                for (int cp = 0; cp < col_pairs; cp++) {
+                    uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_base + 16 * cp);
+                    uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_base + 16 * cp + 64);
+                    uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_base + 16 * cp + 128);
+                    uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_base + 16 * cp + 192);
+
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)), q8_qs[0]);  // 0 .. 7
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)), q8_qs[1]);  // 8 ..15
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)), q8_qs[2]);  // 16..23
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)), q8_qs[3]);  // 24..31
+
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)), q8_qs[4]);  // 32..39
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)), q8_qs[5]);  // 40..47
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)), q8_qs[6]);  // 48..55
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)), q8_qs[7]);  // 56..63
+                }
+
+                // Iterates over a pair of column pairs (4 columns) to use a single 128 register
+                // p = 0 -> 0123  p2 -> 4567
+                for (int i = 0, p = 0; p < col_pairs; i++, p += 2) {
+                    int16x4_t   group_scales_lo = p == 0 ? vget_low_s16(q4sb_scales[0]) : vget_high_s16(q4sb_scales[0]);
+                    int16x4_t   group_scales_hi = p == 0 ? vget_low_s16(q4sb_scales[1]) : vget_high_s16(q4sb_scales[1]);
+                    float32x4_t sb_scale        = p == 0 ? sb_scale_0 : sb_scale_1;
+
+                    // 0123 or 4567
+                    // TODO: Single superblock mul at the end of the superblock
+                    float32x4_t sumf_0 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_lo), vpaddq_s32(acc_lo[p], acc_lo[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_0);
+
+                    float32x4_t sumf_1 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_hi), vpaddq_s32(acc_hi[p], acc_hi[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_1);
+                }
+
+                // Multiply Acc bsum + mins
+                // Each pair of subblocks share the same bsums
+                // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                // cols 0-3 bias
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                // cols 4-7 bias
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_1);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON)
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1889,3 +2068,212 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
     ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
+
+void ggml_gemm_q4_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // bsums pairs belongs to the same q8_k subblock
+                const int16x8_t bsums[4]{
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[4][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                int32x4_t sb_acc[4];    // Aux accumulators to store subblock (partial) results
+                int32x4_t acc[8];       // rows 01 stored in [0][1][2][3] rows 23 stored in [4][5][6][7]
+                int32x4_t bias_acc[8];  // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567, [2]->r1 0123 ...
+                for (int i = 0; i < 8; i++) {
+                    acc[i]      = vdupq_n_s32(0);
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int8_t    q4sb_scales[2][8];
+                    int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                    for (int i = 0; i < 2; i++) {
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
+                    }
+
+                    // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
+                    const int8_t * q8_base = q8_ptr[b].qs + sb * 256;
+
+                    int8x16_t q8_qs_01[8];
+                    int8x16_t q8_qs_23[8];
+
+                    // Load 32-byte per row pair, 1 subblock each time
+                    for (int i = 0; i < 8; i++) {
+                        const int offset = i * 32;  // 16 for row 01, 16 for row 23
+                        q8_qs_01[i]      = vld1q_s8(q8_base + offset);
+                        q8_qs_23[i]      = vld1q_s8(q8_base + offset + 16);
+                    }
+
+                    const int8x16_t q8s[2][8] = {
+                        { q8_qs_01[0], q8_qs_01[1], q8_qs_01[2], q8_qs_01[3],
+                          q8_qs_01[4], q8_qs_01[5], q8_qs_01[6], q8_qs_01[7] },
+                        { q8_qs_23[0], q8_qs_23[1], q8_qs_23[2], q8_qs_23[3],
+                          q8_qs_23[4], q8_qs_23[5], q8_qs_23[6], q8_qs_23[7] },
+                    };
+
+                    // Q4s columns iterated in pairs (01, 23, 45, 67)
+                    for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+                        for (int i = 0; i < 4; i++) {
+                            sb_acc[i] = vdupq_n_s32(0);
+                        }
+
+                        uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 0);    // 0 .. 7 & 32..39
+                        uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 64);   // 8 ..15 & 40..47
+                        uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 128);  // 16..23 & 48..55
+                        uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 192);  // 24..31 & 56..63
+                        const int8x16_t q4_nibbles[2][4] = {
+                            {
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)),
+                            },
+                            {
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)),
+                            }
+                        };
+
+                        // Calculates the Qs muladd of every row pair (rp) rows 01 and 23 of q8
+                        // for each of the internal 32 qs subblock (blk)
+                        for (int rp = 0; rp < 2; rp++) {
+                            for (int blk = 0; blk < 2; blk++) {
+                                const int8x16_t * q8  = &q8s[rp][4 * blk];
+                                const int8x16_t * q4  = q4_nibbles[blk];
+                                int32x4_t         acc = sb_acc[2 * rp + blk];
+                                // mul add for each qs in the same subblock
+                                for (int qs_offset = 0; qs_offset < 4; qs_offset++) {
+                                    acc = vmmlaq_s32(acc, q4[qs_offset], q8[qs_offset]);
+                                }
+                                sb_acc[2 * rp + blk] = acc;
+                            }
+                        }
+
+                        // Scales[i] corresponds to column i
+                        const int scale_offset = cp * 2;
+                        for (int blk = 0; blk < 2; blk++) {
+                            const int32x4_t block_scale = {
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                            };
+                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc[blk], block_scale);
+                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[blk + 2], block_scale);
+                        }
+                    }
+
+                    // Multiply Acc bsum + mins
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        // Each pair of subblocks share the same bsums
+                        // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                        int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][q8_row * 2]);
+                        int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][q8_row * 2 + 1]);
+
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                // Reorder of i8mm output with bias and output layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d    = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q4_dmin = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].dmin + j * 4)));
+                        const float32x4_t dmins   = vmulq_f32(q4_dmin, q8_d);
+
+                        float32x4_t       q4_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q4_d, q8_d);
+
+                        acc_f32[2 * i + j] = vmlsq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(bias_acc[2 * i + j]), dmins);
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // With the previous reorder, the tile is already in the correct memory layout.
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}

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

@@ -0,0 +1,35 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__riscv) && __riscv_xlen == 64
+#include <sys/auxv.h>
+
+//https://github.com/torvalds/linux/blob/master/arch/riscv/include/uapi/asm/hwcap.h#L24
+#ifndef COMPAT_HWCAP_ISA_V
+#define COMPAT_HWCAP_ISA_V (1 << ('V' - 'A'))
+#endif
+
+struct riscv64_features {
+    bool has_rvv = false;
+
+    riscv64_features() {
+        uint32_t hwcap = getauxval(AT_HWCAP);
+
+        has_rvv = !!(hwcap & COMPAT_HWCAP_ISA_V);
+    }
+};
+
+static int ggml_backend_cpu_riscv64_score() {
+    int score = 1;
+    riscv64_features rf;
+
+#ifdef GGML_USE_RVV
+    if (!rf.has_rvv) { return 0; }
+    score += 1 << 1;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_riscv64_score)
+
+#endif  // __riscv && __riscv_xlen == 64

ggml/src/ggml-cpu/arch/x86/repack.cpp

@@ -646,7 +646,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
     __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
     int64_t xstart = 0;
     int anr = nr - nr%16; // Used to align nr with boundary of 16
-#ifdef __AVX512F__
+#if defined(__AVX512BW__) && defined(__AVX512DQ__)
     int anc = nc - nc%16; // Used to align nc with boundary of 16
                           // Mask to mask out nibbles from packed bytes expanded to 512 bit length
     const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
@@ -1041,7 +1041,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
         xstart = anc/8;
         y = 0;
     }
-#endif // __AVX512F__
+#endif // __AVX512BW__ && __AVX512DQ__
 
     // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
 
@@ -1989,7 +1989,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
     int64_t xstart = 0;
     int anr = nr - nr % 16;; // Used to align nr with boundary of 16
-#ifdef __AVX512F__
+#if defined(__AVX512BW__) && defined(__AVX512DQ__)
     int anc = nc - nc % 16; // Used to align nc with boundary of 16
     // Mask to mask out nibbles from packed bytes expanded to 512 bit length
     const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
@@ -2727,7 +2727,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         xstart = anc/8;
         y = 0;
     }
-#endif //AVX512F
+#endif // __AVX512BW__ && __AVX512DQ__
 
     // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
     for (; y < anr / 4; y += 4) {
@@ -3467,7 +3467,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     __m256i scalesmask2 = _mm256_castsi128_si256(scalesmask2_sse);
     scalesmask2 = _mm256_permute2f128_si256(scalesmask2, scalesmask2, 0);
 
-#ifdef __AVX512F__
+#if defined(__AVX512BW__) && defined(__AVX512DQ__)
 
     int anc = nc - nc % 16; // Used to align nc with boundary of 16
 
@@ -4947,7 +4947,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         y = 0;
     }
 
-#endif //AVX512F
+#endif // __AVX512BW__ && __AVX512DQ__
 
     // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
     for (; y < anr / 4; y += 4) {

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

@@ -1731,6 +1731,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_sum_rows(params, tensor);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                ggml_compute_forward_cumsum(params, tensor);
+            } break;
         case GGML_OP_MEAN:
             {
                 ggml_compute_forward_mean(params, tensor);
@@ -1927,6 +1931,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
             } break;
+        case GGML_OP_TRI:
+            {
+                ggml_compute_forward_tri(params, tensor);
+            } break;
+        case GGML_OP_FILL:
+            {
+                ggml_compute_forward_fill(params, tensor);
+            } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_compute_forward_flash_attn_ext(params, tensor);
@@ -1982,6 +1994,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv7(params, tensor);
             } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                ggml_compute_forward_solve_tri(params, tensor);
+            } break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2140,6 +2156,9 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_ACC:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
+        case GGML_OP_FILL:
             {
                 n_tasks = n_threads;
             } break;
@@ -2157,6 +2176,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = 1;
             } break;
         case GGML_OP_COUNT_EQUAL:
+        case GGML_OP_SOLVE_TRI:
             {
                 n_tasks = n_threads;
             } break;
@@ -2179,6 +2199,8 @@ 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_SOFTPLUS:
+                case GGML_UNARY_OP_EXPM1:
                 case GGML_UNARY_OP_FLOOR:
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -39,7 +39,7 @@
 
 #include "kernels.h"
 
-#define NELEMS(x) sizeof(x) / sizeof(*x)
+#define NELEMS(x) (sizeof(x) / sizeof(*x))
 
 template<size_t(*Fn)(size_t,size_t,size_t)>
 static inline size_t kernel_offs_fn3(size_t a, size_t b, size_t c) {
@@ -635,6 +635,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
     },
 #endif
 #endif
+    { /* Sentinel */ }
 };
 
 static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
@@ -803,6 +804,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
         /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
+    { /* Sentinel */ }
 };
 
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
@@ -810,7 +812,7 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
 
     if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
 #if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
-        for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
+        for (size_t i = 0; i < NELEMS(gemm_gemv_kernels) - 1; ++i) {
             if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
                 gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
                 gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
@@ -820,7 +822,7 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
             }
         }
         if (!kernel) {
-            for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8); ++i) {
+            for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8) - 1; ++i) {
                 if ((cpu_features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu &&
                     gemm_gemv_kernels_q8[i].lhs_type == tensor->src[1]->type &&
                     gemm_gemv_kernels_q8[i].rhs_type == tensor->src[0]->type &&
@@ -830,6 +832,10 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
                 }
             }
         }
+#else
+    GGML_UNUSED(gemm_gemv_kernels);
+    GGML_UNUSED(gemm_gemv_kernels_q8);
+    GGML_UNUSED(cpu_features);
 #endif
     }
 
@@ -840,12 +846,14 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features)
     ggml_kleidiai_kernels * kernels = nullptr;
 
 #if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
-    for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
+    for (size_t i = 0; i < NELEMS(gemm_gemv_kernels) - 1; ++i) {
         if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
             kernels = &gemm_gemv_kernels[i];
             break;
         }
     }
+#else
+    GGML_UNUSED(features);
 #endif
 
     return kernels;
@@ -855,12 +863,14 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features)
     ggml_kleidiai_kernels * kernels = nullptr;
 
 #if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
-    for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8); ++i) {
+    for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8) - 1; ++i) {
         if ((features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu) {
             kernels = &gemm_gemv_kernels_q8[i];
             break;
         }
     }
+#else
+    GGML_UNUSED(features);
 #endif
 
     return kernels;

ggml/src/ggml-cpu/ops.cpp

@@ -9,6 +9,7 @@
 
 #include <cfloat>
 #include <algorithm>
+#include <cmath>
 #include <functional>
 
 // ggml_compute_forward_dup
@@ -1395,6 +1396,56 @@ void ggml_compute_forward_sum(
     }
 }
 
+// ggml_compute_forward_cumsum
+
+static void ggml_compute_forward_cumsum_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(dst->nb[0] == sizeof(float));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(ne0 == ne00);
+    GGML_ASSERT(ne1 == ne01);
+    GGML_ASSERT(ne2 == ne02);
+    GGML_ASSERT(ne3 == ne03);
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        float * src_row = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+        float * dst_row = (float *) ((char *) dst->data  + i01*nb1  + i02*nb2  + i03*nb3);
+
+        ggml_vec_cumsum_f32(ne00, dst_row, src_row);
+    }
+}
+
+void ggml_compute_forward_cumsum(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_cumsum_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_sum_rows
 
 static void ggml_compute_forward_sum_rows_f32(
@@ -2141,6 +2192,83 @@ static void ggml_compute_forward_gelu(
     }
 }
 
+// ggml_compute_fill
+
+static void ggml_compute_forward_fill_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const float c = ggml_get_op_params_f32(dst, 0);
+
+    GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+    GGML_TENSOR_LOCALS(size_t,  nb, dst, nb);
+
+    const auto [ir0, ir1] = get_thread_range(params, dst);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne2*ne1);
+        const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
+        const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
+
+        float * dst_ptr  = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
+
+        ggml_vec_set_f32(ne0, dst_ptr, c);
+    }
+}
+
+void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
+    ggml_compute_forward_fill_f32(params, dst);
+}
+
+// ggml_compute_tri
+
+static void ggml_compute_forward_tri_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    const ggml_tri_type ttype = (ggml_tri_type) ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+
+    bool (*bipred)(int, int);
+
+    switch (ttype) {
+        case GGML_TRI_TYPE_LOWER:      bipred = [](int i, int r) { return i <  r; }; break;
+        case GGML_TRI_TYPE_LOWER_DIAG: bipred = [](int i, int r) { return i <= r; }; break;
+        case GGML_TRI_TYPE_UPPER:      bipred = [](int i, int r) { return i >  r; }; break;
+        case GGML_TRI_TYPE_UPPER_DIAG: bipred = [](int i, int r) { return i >= r; }; break;
+        default: GGML_ABORT("invalid tri type");
+    }
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        const float * src_ptr = (const float  *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+              float * dst_ptr = (      float  *) ((      char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1);
+
+        for (int i0 = 0; i0 < ne0; ++i0) {
+            dst_ptr[i0] = bipred(i0, i01) ? src_ptr[i0] : 0.0f;
+        }
+    }
+}
+
+void ggml_compute_forward_tri(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_tri_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_gelu_erf
 
 static void ggml_compute_forward_gelu_erf_f32(
@@ -8536,7 +8664,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                 // n_head
                 for (int h = ih0; h < ih1; ++h) {
                     // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
-                    const float dt_soft_plus = ggml_softplus(dt[h]);
+                    const float dt_soft_plus = ggml_compute_softplus_f32(dt[h]);
                     const float dA = expf(dt_soft_plus * A[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
@@ -8633,7 +8761,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                 // n_head
                 for (int h = ih0; h < ih1; ++h) {
                     // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
-                    const float dt_soft_plus = ggml_softplus(dt[h]);
+                    const float dt_soft_plus = ggml_compute_softplus_f32(dt[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
                     // dim
@@ -8916,6 +9044,14 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_xielu(params, dst);
             } break;
+        case GGML_UNARY_OP_EXPM1:
+            {
+                ggml_compute_forward_expm1(params, dst);
+            } break;
+        case GGML_UNARY_OP_SOFTPLUS:
+            {
+                ggml_compute_forward_softplus(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -9512,6 +9648,75 @@ void ggml_compute_forward_gla(
     }
 }
 
+static void ggml_compute_forward_solve_tri_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];  // A (lower triangular)
+    const struct ggml_tensor * src1 = dst->src[1];  // B (RHS)
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ne00 == ne01); // A must be square
+    GGML_ASSERT(ne0  == ne10); // solution cols == B cols
+    GGML_ASSERT(ne1  == ne11); // solution rows == B rows
+
+    GGML_ASSERT(ne02 == ne12 && ne12 == ne2);
+    GGML_ASSERT(ne03 == ne13 && ne13 == ne3);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t k = ne10;   // number of RHS columns
+    const int64_t n = ne11;   // A is n×n
+    const int64_t nr = ne02 * ne03 * k; // we're parallelizing on columns here, so seq x token x column will be the unit
+
+    // chunks per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // chunk range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    const float * A = (const float *) src0->data;  // [n, n, B1, B2]
+    const float * B = (const float *) src1->data;  // [n, k, B1, B2]
+          float * X = (      float *) dst->data;   // [n, k, B1, B2]
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*k);
+        const int64_t i02 = (ir - i03*ne02*k)/k;
+        const int64_t i01 = (ir - i03*ne02*k - i02*k);
+
+        const float * A_batch = A + i02 * nb02 / sizeof(float) + i03 * nb03 / sizeof(float);
+        const float * B_batch = B + i02 * nb12 / sizeof(float) + i03 * nb13 / sizeof(float);
+
+        float * X_batch = X + i02 * nb2 / sizeof(float) + i03 * nb3 / sizeof(float);
+
+        for (int64_t i00 = 0; i00 < n; ++i00) {
+            float sum = 0.0f;
+            for (int64_t t = 0; t < i00; ++t) {
+                sum += A_batch[i00 * n + t] * X_batch[t * k + i01];
+            }
+
+            const float diag = A_batch[i00 * n + i00];
+            GGML_ASSERT(diag != 0.0f && "Zero diagonal in triangular matrix");
+            X_batch[i00 * k + i01] = (B_batch[i00 * k + i01] - sum) / diag;
+        }
+    }
+}
+
+void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        ggml_compute_forward_solve_tri_f32(params, dst);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
 // ggml_compute_forward_rwkv_wkv7
 
 static void ggml_compute_forward_rwkv_wkv7_f32(

ggml/src/ggml-cpu/ops.h

@@ -34,6 +34,7 @@ void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_cumsum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -81,6 +82,8 @@ void ggml_compute_forward_arange(const struct ggml_compute_params * params, stru
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_fill(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_flash_attn_back(
         const struct ggml_compute_params * params,
@@ -96,6 +99,7 @@ void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params,
 void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/repack.cpp

@@ -1689,6 +1689,9 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 
         const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
 
+        // INFO: Quantization is done in planes to avoid extra complexity in chunking.
+        // Flattening dimensions not multiple of INTER_SIZE would require extra handling depending on how
+        // the planes are broadcast.
         for (int64_t i12 = 0; i12 < ne12; i12++) {
             char * data_ptr  = (char *) src1->data + i12 * nb12;
             char * wdata_ptr = wdata + i12 * nbw2;
@@ -1709,14 +1712,17 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 
         // 4x chunks per thread
         const int64_t nr0 = ggml_nrows(op->src[0]);
-        const int64_t nr1 = ne1 * ne2 * ne3;
 
         int     nth_scaled  = nth * 4;
         int64_t chunk_size0 = (nr0 + nth_scaled - 1) / nth_scaled;
-        // avoid too small chunks for narrow src1
-        int64_t chunk_size1 = MAX(16, (nr1 + nth - 1) / nth);
         int64_t nchunk0     = (nr0 + chunk_size0 - 1) / chunk_size0;
-        int64_t nchunk1     = (nr1 + chunk_size1 - 1) / chunk_size1;
+
+        // src1 is chunked only by full planes.
+        // When we flatten we need to address dimensions not multiple of the q8 INTER_SIZE
+        // to route them thorugh GEMV.
+        // nchunk1 = ne12 also avoids messing the chunking for models with no 3d tensors
+        // to avoid affecting their performance
+        int64_t nchunk1 = ne12;
 
         // 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
@@ -1725,13 +1731,11 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             nchunk0 = (nr0 + min_chunk_size - 1) / min_chunk_size;
         }
 
-        if (nth == 1 || nchunk0 * nchunk1 < nth || disable_chunking) {
-            nchunk0 = nr0 > nr1 ? nth : 1;
-            nchunk1 = nr0 > nr1 ? 1 : nth;
+        if (nth == 1 || nchunk0 < nth || disable_chunking) {
+            nchunk0 = nth;
         }
 
         const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
-        const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
 
         // 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
@@ -1755,8 +1759,9 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int64_t src0_start = dr0 * ith0;
             int64_t src0_end   = MIN(src0_start + dr0, nr0);
 
-            int64_t src1_start = dr1 * ith1;
-            int64_t src1_end   = MIN(src1_start + dr1, nr1);
+            // full-plane range for src1
+            int64_t src1_start = ith1 * ne11;
+            int64_t src1_end = (ith1 + 1) * ne11;
 
             // 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
@@ -1956,6 +1961,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_K_8x8_q8_K;
             }
         }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x8_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_Q2_K) {
         if (ggml_cpu_has_avx512()) {
             if (cur->ne[1] % 8 == 0) {

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

@@ -160,18 +160,18 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32xt                        svfloat32_t
 #define GGML_F32xt_ZERO                   svdup_n_f32(0.0f)
 #define GGML_F32xt_SET1(x)                svdup_n_f32(x)
-#define GGML_F32xt_LOAD_IMPL(pg, a, ...)  svld1_f32(pg, a)
-#define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
-#define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_LOAD_IMPL(pg, a)       svld1_f32(pg, a)
+#define GGML_F32xt_LOAD(a)                GGML_F32xt_LOAD_IMPL(DEFAULT_PG, a)
+#define GGML_F32xt_STORE_IMPL(pg, a, b)   svst1_f32(pg, a, b)
+#define GGML_F32xt_STORE(a, b)            GGML_F32xt_STORE_IMPL(DEFAULT_PG, a, b)
 #define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, b, c, a)
-#define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_FMA(a, b, c)           GGML_F32xt_FMA_IMPL(DEFAULT_PG, a, b, c)
 #define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
-#define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_ADD(a, b)              GGML_F32xt_ADD_IMPL(DEFAULT_PG, a, b)
 #define GGML_F32xt_MUL_IMPL(pg, a, b)     svmul_f32_m(pg, a, b)
-#define GGML_F32xt_MUL(...)               GGML_F32xt_MUL_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_MUL(a, b)              GGML_F32xt_MUL_IMPL(DEFAULT_PG, a, b)
 #define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
-#define GGML_F32xt_REDUCE_ONE(...)        GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE_ONE(a)          GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, a)
 #define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
 {                                                      \
     sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2);        \
@@ -183,7 +183,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5);        \
     (res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1);  \
 }
-#define GGML_F32xt_REDUCE(...) GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE(res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
+        GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)
 
 #define GGML_F32_VEC        GGML_F32xt
 #define GGML_F32_VEC_ZERO   GGML_F32xt_ZERO
@@ -206,11 +207,11 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32Cxt_STORE(dst_ptr, src_vec) svst1_f16(DEFAULT_PG16, (__fp16 *)(dst_ptr), (src_vec))
 
 #define GGML_F32Cxt_FMA_IMPL(pg, a, b, c)   svmad_f16_x(pg, b, c, a)
-#define GGML_F32Cxt_FMA(...)                GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_FMA(a, b, c)            GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, a, b, c)
 #define GGML_F32Cxt_ADD_IMPL(pg, a, b)      svadd_f16_x(pg, a, b)
-#define GGML_F32Cxt_ADD(...)                GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_ADD(a, b)               GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, a, b)
 #define GGML_F32Cxt_MUL_IMPL(pg, a, b)      svmul_f16_x(pg, a, b)
-#define GGML_F32Cxt_MUL(...)                GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_MUL(a, b)               GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, a, b)
 #define GGML_F32Cxt_REDUCE                  GGML_F16xt_REDUCE_MIXED
 
 #define GGML_F16x_VEC                GGML_F32Cxt
@@ -224,7 +225,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F16x_VEC_REDUCE         GGML_F32Cxt_REDUCE
 
 #define GGML_F16xt_REDUCE_ONE_IMPL(pg, a) svaddv_f16(pg, a)
-#define GGML_F16xt_REDUCE_ONE(...)        GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F16xt_REDUCE_ONE(a)          GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, a)
 
 #define GGML_F16xt_REDUCE_MIXED_IMPL(pg16, res, sum1, sum2, sum3, sum4)  \
 {                                                      \
@@ -234,7 +235,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     __fp16 sum_f16 = svaddv_f16(pg16, sum1);           \
     (res) = (ggml_float) sum_f16;                      \
 }
-#define GGML_F16xt_REDUCE_MIXED(...) GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F16xt_REDUCE_MIXED(res, sum1, sum2, sum3, sum4)  \
+        GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, res, sum1, sum2, sum3, sum4)
 
 // F16 NEON
 

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

@@ -73,6 +73,14 @@ static inline float op_log(float x) {
     return logf(x);
 }
 
+static inline float op_expm1(float x) {
+    return expf(x) - 1.0f;
+}
+
+static inline float op_softplus(float x) {
+    return (x > 20.0f) ? x : logf(1.0f + expf(x));
+}
+
 static inline float op_floor(float x) {
     return floorf(x);
 }
@@ -290,6 +298,14 @@ void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor *
     unary_op<op_log>(params, dst);
 }
 
+void ggml_compute_forward_expm1(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_expm1>(params, dst);
+}
+
+void ggml_compute_forward_softplus(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_softplus>(params, dst);
+}
+
 void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
     unary_op<op_floor>(params, dst);
 }

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

@@ -22,6 +22,8 @@ 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_expm1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_softplus(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);

ggml/src/ggml-cpu/vec.cpp

@@ -360,6 +360,13 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
     for (; i + 3 < n; i += 4) {
         vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
     }
+#elif defined(__riscv_v_intrinsic)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat32m2_t vy = ggml_v_silu_m2(vx, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
 #endif
     for (; i < n; ++i) {
         y[i] = ggml_silu_f32(x[i]);
@@ -460,6 +467,16 @@ ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const floa
         val = vec_mul(val, val);
         sum += (ggml_float)vec_hsum_f32x4(val);
     }
+#elif defined(__riscv_v_intrinsic)
+    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t val = __riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], vl), mean, vl);
+        __riscv_vse32_v_f32m2(&y[i], val, vl);
+        val = __riscv_vfmul_vv_f32m2(val, val, vl);
+        vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, vl);
+    }
+    sum = (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
 #endif
     for (; i < n; ++i) {
         float val = x[i] - mean;

ggml/src/ggml-cpu/vec.h

@@ -698,60 +698,61 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 }
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16;
-        const int ggml_f16_step = 2 * ggml_f16_epr;
+#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
+    const int sve_register_length = svcntb() * 8;
+    const int ggml_f16_epr = sve_register_length / 16;
+    const int ggml_f16_step = 2 * ggml_f16_epr;
 
-        GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
-        const int np = (n & ~(ggml_f16_step - 1));
-        svfloat16_t ay1, ay2;
+    GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
+    const int np = (n & ~(ggml_f16_step - 1));
+    svfloat16_t ay1, ay2;
 
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_MUL(ay1, vx);
-            GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
+    for (int i = 0; i < np; i += ggml_f16_step) {
+        ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_MUL(ay1, vx);
+        GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
 
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_MUL(ay2, vx);
-            GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
+        ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_MUL(ay2, vx);
+        GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
+    }
+    // leftovers
+    // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
+    if (np < n) {
+        svbool_t pg = svwhilelt_b16(np, n);
+        svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
+        svfloat16_t out = svmul_f16_m(pg, hy, vx);
+        svst1_f16(pg, (__fp16 *)(y + np), out);
+    }
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
+    for (int i = 0, vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m2(n - i);
+        vfloat16m2_t vy = __riscv_vle16_v_f16m2((_Float16 *)&y[i], vl);
+        vfloat32m4_t vy32 = __riscv_vfwcvt_f_f_v_f32m4(vy, vl);
+        vy32 = __riscv_vfmul_vf_f32m4(vy32, v, vl);
+        vy = __riscv_vfncvt_f_f_w_f16m2(vy32, vl);
+        __riscv_vse16_v_f16m2((_Float16 *)&y[i], vy, vl);
+    }
+#elif defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        // leftovers
-        // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
-        if (np < n) {
-            svbool_t pg = svwhilelt_b16(np, n);
-            svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
-            svfloat16_t out = svmul_f16_m(pg, hy, vx);
-            svst1_f16(pg, (__fp16 *)(y + np), out);
-        }
-    #elif defined(__riscv_v_intrinsic)
-        // todo: RVV impl
-        // scalar
-        for (int i = 0; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
+    }
 
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
-
-                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
-            }
-        }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-        }
-    #endif
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+    }
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -1416,6 +1417,16 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #endif
 }
 
+inline static void ggml_vec_cumsum_f32(const int n, float * y, const float * x) {
+    for (int i = 0; i < n; ++i) {
+        if (i == 0) {
+            y[i] = x[i];
+        } else {
+            y[i] = y[i - 1] + x[i];
+        }
+    }
+}
+
 inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) {
     ggml_float sum = 0.0;
     for (int i = 0; i < n; ++i) {

ggml/src/ggml-cuda/common.cuh

@@ -224,6 +224,10 @@ 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)
 
+#if defined(GGML_USE_HIP) && defined(RDNA4)
+#define AMD_WMMA_AVAILABLE
+#endif // defined(GGML_USE_HIP) && defined(RDNA4)
+
 // 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
@@ -283,6 +287,10 @@ static bool amd_mfma_available(const int cc) {
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
+static bool amd_wmma_available(const int cc) {
+    return GGML_CUDA_CC_IS_RDNA4(cc);
+}
+
 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;
 }

ggml/src/ggml-cuda/convert.cuh

@@ -39,6 +39,15 @@ template<typename dst_t, typename src_t>
         return __float2bfloat16(float(x));
     } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
         return __bfloat162float(x);
+    } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, half2>) {
+        return __float22half2_rn(x);
+    } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, nv_bfloat162>) {
+        // bypass compile error on cuda 12.0.1
+#ifdef GGML_USE_HIP
+        return __float22bfloat162_rn(x);
+#else
+        return {x.x, x.y};
+#endif // GGML_USE_HIP
     } else if constexpr(std::is_same_v<dst_t, int32_t>) {
         return int32_t(x);
     } else {

ggml/src/ggml-cuda/cpy-utils.cuh

@@ -212,6 +212,6 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 }
 
 template<typename src_t, typename dst_t>
-static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
+static __device__ void cpy_1_scalar(const char * cxi, char * cdsti) {
     *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
 }

ggml/src/ggml-cuda/cpy.cu

@@ -12,10 +12,10 @@ 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, 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) {
+static __global__ void cpy_scalar(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 int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -40,7 +40,7 @@ static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
 }
 
 template <typename T>
-static __global__ void cpy_flt_transpose(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_scalar_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) {
@@ -166,7 +166,7 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
 }
 
 template<typename src_t, typename dst_t>
-static __global__ void cpy_flt_contiguous(const char * cx, char * cdst, const int64_t ne) {
+static __global__ void cpy_scalar_contiguous(const char * cx, char * cdst, const int64_t ne) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -180,17 +180,17 @@ static __global__ void cpy_flt_contiguous(const char * cx, char * cdst, const in
 }
 
 template<typename src_t, typename dst_t>
-static void ggml_cpy_flt_contiguous_cuda(
+static void ggml_cpy_scalar_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>>>
+    cpy_scalar_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(
+static void ggml_cpy_scalar_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) {
@@ -212,11 +212,11 @@ static void ggml_cpy_flt_cuda(
                       (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>>>
+        cpy_scalar_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>>>
+        cpy_scalar<cpy_1_scalar<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);
     }
 }
@@ -384,7 +384,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src1_ddc = (char *) src1->data;
 
     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;
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) &&
+        src0->ne[3] == 1 && nb02 == ne00 * ne01 * (int64_t)ggml_element_size(src0);
 
     if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
@@ -398,94 +399,132 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         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);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<float, half>        (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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);
+        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);
+        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);
+        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);
+        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);
     } 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);
+        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);
+        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);
     } 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);
+        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);
+        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);
     } 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);
+        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);
+        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);
+        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) {
         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);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<half, nv_bfloat16>  (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<half, float>        (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         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);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, half>  (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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_I32 && src1->type == GGML_TYPE_I32) {
+        if (can_be_transposed) {
+            ggml_cpy_scalar_cuda<int32_t, int32_t, 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_scalar_cuda<int32_t, 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_F32 && src1->type == GGML_TYPE_I32) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<float, int32_t>     (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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) {
         if (contiguous_srcs) {
-            ggml_cpy_flt_contiguous_cuda<int32_t, float>     (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_scalar_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);
+            ggml_cpy_scalar_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/src/ggml-cuda/ggml-cuda.cu

@@ -2527,6 +2527,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_TRUNC:
                     ggml_cuda_op_trunc(ctx, dst);
                     break;
+                case GGML_UNARY_OP_EXPM1:
+                    ggml_cuda_op_expm1(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_cuda_op_softplus(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2995,6 +3001,10 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
                                                 const ggml_tensor * view,
                                                 const ggml_tensor * set_rows) {
+
+    if (rope->op != GGML_OP_ROPE || view->op != GGML_OP_VIEW || set_rows->op != GGML_OP_SET_ROWS) {
+        return false;
+    }
     // ne3 not tested
     if (rope->src[0]->ne[3] != 1) {
         return false;
@@ -3738,10 +3748,110 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
+#if defined(__linux__)
+// Helper function to get available memory from /proc/meminfo for UMA systems
+static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_kb, long * free_swap_kb) {
+    FILE * meminfo_file = nullptr;
+    // 2KB buffer for reading /proc/meminfo since it does not report size info, should be enough
+    const size_t BUFFER_SIZE = 2048;
+    auto file_buffer = std::make_unique<char[]>(BUFFER_SIZE);
+    size_t bytes_read = 0;
+    long huge_tlb_total_pages = -1;
+    long huge_tlb_free_pages = -1;
+    long huge_tlb_page_size = -1;
+
+    if (available_memory_kb == nullptr || free_swap_kb == nullptr) {
+        return false;
+    }
+
+    meminfo_file = fopen("/proc/meminfo", "r");
+    if (meminfo_file == nullptr) {
+        GGML_LOG_ERROR("%s: failed to open /proc/meminfo\n", __func__);
+        return false;
+    }
+
+    // Read file into buffer
+    bytes_read = fread(file_buffer.get(), 1, BUFFER_SIZE - 1, meminfo_file);
+    fclose(meminfo_file);
+
+    if (bytes_read == 0) {
+        GGML_LOG_ERROR("%s: failed to read from /proc/meminfo\n", __func__);
+        return false;
+    }
+    file_buffer[bytes_read] = '\0';
+
+    *available_memory_kb = -1;
+    *free_swap_kb = -1;
+
+    // Parse the file buffer line by line
+    char * line = file_buffer.get();
+    char * line_next;
+    while (line < file_buffer.get() + bytes_read) {
+        // Find the end of the current line
+        line_next = strchr(line, '\n');
+        if (line_next != nullptr) {
+            *line_next = '\0';
+            line_next++;
+        } else {
+            line_next = file_buffer.get() + bytes_read;
+        }
+
+        long value;
+        if (sscanf(line, "MemAvailable: %ld kB", &value) == 1) {
+            *available_memory_kb = value;
+        } else if (sscanf(line, "SwapFree: %ld kB", &value) == 1) {
+            *free_swap_kb = value;
+        } else if (sscanf(line, "HugePages_Total: %ld", &value) == 1) {
+            huge_tlb_total_pages = value;
+        } else if (sscanf(line, "HugePages_Free: %ld", &value) == 1) {
+            huge_tlb_free_pages = value;
+        } else if (sscanf(line, "Hugepagesize: %ld kB", &value) == 1) {
+            huge_tlb_page_size = value;
+        }
+
+        line = line_next;
+    }
+
+    if (huge_tlb_total_pages != 0 && huge_tlb_total_pages != -1) {
+        *available_memory_kb = huge_tlb_free_pages * huge_tlb_page_size;
+
+        // Hugetlbfs pages are not swappable.
+        *free_swap_kb = 0;
+    }
+
+    GGML_LOG_DEBUG("%s: final available_memory_kb: %ld\n", __func__, *available_memory_kb);
+    return true;
+}
+#endif // defined(__linux__)
+
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     ggml_cuda_set_device(ctx->device);
     CUDA_CHECK(cudaMemGetInfo(free, total));
+
+// ref: https://github.com/ggml-org/llama.cpp/pull/17368
+#if defined(__linux__)
+    // Check if this is a UMA (Unified Memory Architecture) system
+    cudaDeviceProp prop;
+    CUDA_CHECK(cudaGetDeviceProperties(&prop, ctx->device));
+
+    // Check if UMA is explicitly enabled via environment variable
+    bool uma_env = getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr;
+    bool is_uma = prop.unifiedAddressing > 0 || uma_env;
+
+    if (is_uma) {
+        // For UMA systems (like DGX Spark), use system memory info
+        long available_memory_kb = 0;
+        long free_swap_kb = 0;
+
+        if (ggml_backend_cuda_get_available_uma_memory(&available_memory_kb, &free_swap_kb) && available_memory_kb > 0) {
+            *free = (size_t)available_memory_kb * 1024;
+        } else {
+            GGML_LOG_ERROR("%s: /proc/meminfo reading failed, using cudaMemGetInfo\n", __func__);
+        }
+    }
+#endif // defined(__linux__)
+
 }
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
@@ -3829,6 +3939,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_EXPM1:
+                case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_FLOOR:
                 case GGML_UNARY_OP_CEIL:
@@ -4003,6 +4115,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_I32) {
+                    return true;
+                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }

ggml/src/ggml-cuda/mma.cuh

@@ -73,7 +73,7 @@ namespace ggml_cuda_mma {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
 
-#if defined(GGML_USE_HIP)
+#if defined(AMD_MFMA_AVAILABLE)
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
@@ -149,6 +149,34 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 16) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 16) {
+                return 8 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 16) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#endif
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
@@ -236,6 +264,32 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+        static constexpr int ne = I * J / 32;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return 4 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
 #else
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
@@ -285,6 +339,34 @@ namespace ggml_cuda_mma {
     struct tile<I_, J_, nv_bfloat162> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if defined(AMD_WMMA_AVAILABLE)
+        static constexpr int ne = I * J / 32;
+        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return 4 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
@@ -320,6 +402,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
+#endif  // defined(AMD_WMMA_AVAILABLE)
     };
 
     template <int I, int J>
@@ -353,6 +436,21 @@ namespace ggml_cuda_mma {
             const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
             xi[0] = xs[0];
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+        if constexpr (I == 16 && J == 4) {
+            int64_t * xi = (int64_t *) t.x;
+            const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
+            xi[0] = xs[0];
+        }else if constexpr (I == 16 && J == 8) {
+            int64_t * xi = (int64_t *) t.x;
+            const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I));
+            xi[0] = xs[0];
+
+            const int64_t * xs1 = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I) + 2);
+            xi[1] = xs1[0];
+        }else{
+            NO_DEVICE_CODE;
+        }
 #else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
@@ -639,12 +737,34 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#elif defined(AMD_WMMA_AVAILABLE)
+        using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
+        using floatx8_t = __attribute__((ext_vector_type(8))) float;
+        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
+        const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
+        const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
+    static __device__ __forceinline__ void mma(
+            tile<16, 16, float> & D, const tile<16, 8, nv_bfloat162> & A, const tile<16, 8, nv_bfloat162> & B) {
+#if defined(AMD_WMMA_AVAILABLE)
+        using bf16x8_t = __attribute__((ext_vector_type(8))) __bf16;
+        using floatx8_t = __attribute__((ext_vector_type(8))) float;
+        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
+        const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
+        const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // AMPERE_MMA_AVAILABLE
+    }
+
     static __device__ __forceinline__ void mma(
             tile<16, 16, int> & D, const tile<16, 8, int> & A, const tile<16, 8, int> & B) {
 #if defined(AMD_MFMA_AVAILABLE)
@@ -665,6 +785,36 @@ namespace ggml_cuda_mma {
                                                       acc[0],
                                                       0, 0, 0);
 #endif // defined(CDNA3)
+
+#elif defined(AMD_WMMA_AVAILABLE)
+        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+        int32x2_t * a_vec = (int32x2_t *) A.x;
+        int32x2_t * b_vec = (int32x2_t *) B.x;
+
+        using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+        int32x8_t * acc = (int32x8_t *) D.x;
+
+#if defined(RDNA4)
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            true
+        );
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+            true,
+            a_vec[1],
+            true,
+            b_vec[1],
+            acc[0],
+            true
+        );
+#endif // defined(RDNA4)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -691,6 +841,7 @@ namespace ggml_cuda_mma {
                                                      acc[0],
                                                      0, 0, 0);
 #endif // defined(CDNA3)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -735,4 +886,31 @@ namespace ggml_cuda_mma {
         mma(D16[1], A16[1], B);
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     }
+
+static __device__ __forceinline__ void mma(
+            tile<16, 16, int> & D, const tile<16, 4, int> & A, const tile<16, 4, int> & B) {
+#if defined(AMD_WMMA_AVAILABLE)
+    using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+    int32x2_t * a_vec = (int32x2_t *) A.x;
+    int32x2_t * b_vec = (int32x2_t *) B.x;
+
+    using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+    int32x8_t * acc = (int32x8_t *) D.x;
+
+    acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+        true,
+        a_vec[0],
+        true,
+        b_vec[0],
+        acc[0],
+        false
+    );
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif
+    }
 }
+

ggml/src/ggml-cuda/mmf.cu

@@ -151,7 +151,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
             return false;
         }
     } else {
-        if (src1_ncols > 16) {
+        if (src1_ncols > 16 || GGML_CUDA_CC_IS_RDNA4(cc)) {
             return false;
         }
     }
@@ -160,9 +160,9 @@ 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 volta_mma_available(cc) || turing_mma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc);
+            return ampere_mma_available(cc) || amd_wmma_available(cc);
         default:
             return false;
     }

ggml/src/ggml-cuda/mmf.cuh

@@ -2,6 +2,7 @@
 
 #include "mma.cuh"
 #include "common.cuh"
+#include "convert.cuh"
 
 using namespace ggml_cuda_mma;
 
@@ -27,20 +28,35 @@ static __global__ void mul_mat_f(
         const int stride_col_id, const int stride_row_id,
         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)
+// TODO: handle this in a consistent and simpler way after AMD MFMA support has been added
+#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE)
+    // Special case for tf32, just dummy mma layout as wmma doesn't support it.
+    constexpr int tile_B_I = std::is_same_v<T, float> ? 8 : 16;
+    constexpr int tile_C_J = std::is_same_v<T, float> ? 8 : 16;
+    typedef tile<16,       8, T>     tile_A;
+    typedef tile<tile_B_I, 8, T>     tile_B;
+    typedef tile<16,       tile_C_J, float> tile_C;
+
+    constexpr bool a_supported = tile_A::supported();
+    constexpr bool b_supported = tile_B::supported();
+    constexpr bool c_supported = tile_C::supported();
+    constexpr bool supported = a_supported && b_supported && c_supported;
+#else
     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 bool supported = I_16_supported || I_32_supported;
 
     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;
+#endif // defined(AMD_WMMA_AVAILABLE)
+    if constexpr (!supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -161,11 +177,11 @@ static __global__ void mul_mat_f(
 
                     if constexpr (!has_ids) {
                         const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2(0.0f, 0.0f);
-                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = ggml_cuda_cast<T>(tmp);
                     } else {
                         const bool valid = j < cols_per_block && (col_base + j) < ncols_dst_total && slot_map[j] >= 0;
                         float2 tmp = valid ? *(const float2*) &y[slot_map[j]*stride_channel_y + 2*(j*stride_col_y + col)] : make_float2(0.0f, 0.0f);
-                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = ggml_cuda_cast<T>(tmp);
                     }
                 }
             } else {
@@ -239,7 +255,7 @@ static __global__ void mul_mat_f(
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
 }
 
 //This kernel is for larger batch sizes of mul_mat_id
@@ -253,20 +269,35 @@ static __global__ void mul_mat_f_ids(
         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)
+// TODO: handle this in a consistent and simpler way after AMD MFMA support has been added
+#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE)
+    // Special case for tf32, just dummy mma layout as wmma doesn't support it.
+    constexpr int tile_B_I = std::is_same_v<T, float> ? 8 : 16;
+    constexpr int tile_C_J = std::is_same_v<T, float> ? 8 : 16;
+    typedef tile<16,       8, T>     tile_A;
+    typedef tile<tile_B_I, 8, T>     tile_B;
+    typedef tile<16,       tile_C_J, float> tile_C;
+
+    constexpr bool a_supported = tile_A::supported();
+    constexpr bool b_supported = tile_B::supported();
+    constexpr bool c_supported = tile_C::supported();
+    constexpr bool supported = a_supported && b_supported && c_supported;
+#else
     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();
+    constexpr bool supported = I_16_supported || I_32_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.
+    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;
+#endif // defined(AMD_WMMA_AVAILABLE)
+    if constexpr (!supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -408,7 +439,7 @@ static __global__ void mul_mat_f_ids(
 #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};
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = ggml_cuda_cast<T>(tmp);
                 }
 
                 if (itB + 1 < ntB) {
@@ -492,7 +523,7 @@ static __global__ void mul_mat_f_ids(
         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)
+#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
 }
 
 template<typename T, int cols_per_block, int nwarps>
@@ -554,7 +585,8 @@ void mul_mat_f_cuda(
         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;
+    typedef tile<16, 8, T>     tile_B_16;
+    typedef tile< 8, 8, T>     tile_B_8;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
     GGML_ASSERT(stride_row   % 2 == 0);
@@ -581,7 +613,8 @@ void mul_mat_f_cuda(
 
     constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
     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_cols_per_block_pad = amd_wmma_available(cc) ? tile_B_16::I : tile_B_8::I;
+    const int nbytes_shared_combine = GGML_PAD(cols_per_block, nbytes_cols_per_block_pad) * (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;

ggml/src/ggml-cuda/mmq.cu

@@ -306,5 +306,11 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    if (amd_wmma_available(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+            return true;
+        }
+    }
+
+    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ggml/src/ggml-cuda/unary.cu

@@ -81,6 +81,14 @@ static __device__ __forceinline__ float op_log(float x) {
     return logf(x);
 }
 
+static __device__ __forceinline__ float op_expm1(float x) {
+    return expm1f(x);
+}
+
+static __device__ __forceinline__ float op_softplus(float x) {
+    return (x > 20.0f) ? x : logf(1.0f + expf(x));
+}
+
 static __device__ __forceinline__ float op_elu(float x) {
     return (x > 0.f) ? x : expm1f(x);
 }
@@ -233,6 +241,14 @@ void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_trunc>(ctx, dst);
 }
+
+void ggml_cuda_op_expm1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_expm1>(ctx, dst);
+}
+
+void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_softplus>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>

ggml/src/ggml-cuda/unary.cuh

@@ -61,6 +61,10 @@ void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_expm1(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-hexagon/CMakeLists.txt

@@ -43,6 +43,14 @@ set(HTP_CMAKE_ARGS
     -DHEXAGON_TOOLS_ROOT=$ENV{HEXAGON_TOOLS_ROOT}
     -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG})
 
+ExternalProject_Add(htp-v68
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v68 -DPREBUILT_LIB_DIR="toolv19_v68")
+
+ExternalProject_Add(htp-v69
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v69 -DPREBUILT_LIB_DIR="toolv19_v69")
+
 ExternalProject_Add(htp-v73
     SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
     CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v73 -DPREBUILT_LIB_DIR="toolv19_v73")
@@ -61,6 +69,8 @@ ExternalProject_Add(htp-v81
 
 # Install Hexagon skels required at runtime
 install(FILES
+    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v68.so
+    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v69.so
     ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v73.so
     ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v75.so
     ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v79.so

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

@@ -9,6 +9,7 @@
 #include <chrono>
 #include <mutex>
 #include <string>
+#include <stdexcept>
 
 #ifdef _WIN32
 #    include <sal.h>
@@ -240,6 +241,23 @@ struct ggml_hexagon_session {
     uint32_t         prof_pkts;
 };
 
+static inline void hex_print_op_info(const ggml_tensor * op, ggml_hexagon_session * sess, const uint32_t req_flags) {
+    char dims[64 * GGML_MAX_SRC];
+    char strides[64 * GGML_MAX_SRC];
+    char types[16 * GGML_MAX_SRC];
+    char buffs[64 * GGML_MAX_SRC];
+    char names[64 * GGML_MAX_SRC];
+
+    hex_format_op_dims(dims, op);
+    hex_format_op_strides(strides, op);
+    hex_format_op_types(types, op);
+    hex_format_op_buffs(buffs, op);
+    hex_format_op_names(names, op);
+
+    HEX_VERBOSE("ggml-hex: %s %s: %s : %s : %s : %s : %s: flags 0x%x\n", sess->name.c_str(), ggml_op_name(op->op),
+                names, dims, types, strides, buffs, req_flags);
+}
+
 void ggml_hexagon_session::enqueue(struct htp_general_req &req, struct dspqueue_buffer *bufs, uint32_t n_bufs, bool sync) {
     // Bump pending flag (cleared in the session::flush once we get the responce)
     this->op_pending++;  // atomic inc
@@ -1912,6 +1930,15 @@ static bool hex_supported_dims(const struct ggml_tensor * x, const struct ggml_t
     return true;
 }
 
+template <typename... _TTensor>
+static inline bool hex_supported_buffer(const struct ggml_hexagon_session * sess, _TTensor... tensors) {
+    return ([&]() -> bool {
+        return !tensors || !tensors->buffer ||
+               (ggml_backend_buffer_is_hexagon(tensors->buffer) &&
+                ggml_backend_hexagon_buffer_get_sess(tensors->buffer) == sess);
+    }() && ...);
+}
+
 static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * sess, const struct ggml_tensor * dst) {
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
@@ -1959,16 +1986,7 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
     }
 
     // src0 & src1 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, dst)) {
         return false;
     }
 
@@ -2016,20 +2034,7 @@ static bool ggml_hexagon_supported_mul_mat_id(const struct ggml_hexagon_session
 
     // src0 (weights) must be repacked and mapped to the same session
     // src1 & sr2 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (src2->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src2->buffer) || ggml_backend_hexagon_buffer_get_sess(src2->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, src2, dst)) {
         return false;
     }
 
@@ -2063,16 +2068,7 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, dst)) {
         return false;
     }
 
@@ -2104,20 +2100,7 @@ static bool ggml_hexagon_supported_add_id(const struct ggml_hexagon_session * se
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (src2->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src2->buffer) || ggml_backend_hexagon_buffer_get_sess(src2->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, src2, dst)) {
         return false;
     }
 
@@ -2144,12 +2127,7 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
     }
 
     // src0 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, dst)) {
         return false;
     }
 
@@ -2186,16 +2164,7 @@ static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1 && src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, dst)) {
         return false;
     }
 
@@ -2248,16 +2217,7 @@ static bool ggml_hexagon_supported_softmax(const struct ggml_hexagon_session * s
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1 && src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, dst)) {
         return false;
     }
 
@@ -2269,7 +2229,7 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
 
     int mode = op_params[2];
 
-    if ((mode & GGML_ROPE_TYPE_NEOX) || (mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
+    if ((mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
         return false;
     }
     if (mode & 1) {
@@ -2312,20 +2272,7 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
     }
 
     // src0, src1, src2 & dst must be mapped to the same session
-    if (src0->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src0->buffer) || ggml_backend_hexagon_buffer_get_sess(src0->buffer) != sess)) {
-        return false;
-    }
-    if (src1->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src1->buffer) || ggml_backend_hexagon_buffer_get_sess(src1->buffer) != sess)) {
-        return false;
-    }
-    if (src2 && src2->buffer &&
-        (!ggml_backend_buffer_is_hexagon(src2->buffer) || ggml_backend_hexagon_buffer_get_sess(src2->buffer) != sess)) {
-        return false;
-    }
-    if (dst->buffer &&
-        (!ggml_backend_buffer_is_hexagon(dst->buffer) || ggml_backend_hexagon_buffer_get_sess(dst->buffer) != sess)) {
+    if (!hex_supported_buffer(sess, src0, src1, src2, dst)) {
         return false;
     }
 
@@ -2346,6 +2293,26 @@ static void init_htp_tensor(htp_tensor * h, const ggml_tensor * t) {
     h->nb[3] = t->nb[3];
 }
 
+static size_t dspqueue_buffers_init(dspqueue_buffer * buf, const ggml_tensor * t, bool flush_host, bool flush_htp) {
+    if (!t) {
+        return 0;
+    }
+
+    memset(buf, 0, sizeof(*buf));
+    auto tensor_buf = static_cast<ggml_backend_hexagon_buffer_context *>(t->buffer->context);
+    buf->fd      = tensor_buf->fd;
+    buf->ptr     = t->data;
+    buf->offset  = (uint8_t *) t->data - tensor_buf->base;
+    buf->size    = ggml_nbytes(t);
+    buf->flags   = (flush_host ? DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER : 0);        // Flush CPU
+    buf->flags |= (flush_htp ? DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT : 0);  // Invalidate DSP
+    return 1;
+}
+
+static ggml_hexagon_session * get_session_from_tensor(const ggml_tensor * t) {
+    return static_cast<ggml_backend_hexagon_buffer_context *>(t->buffer->context)->sess;
+}
+
 static void hex_dump_dspbuf(const struct ggml_tensor * t, const dspqueue_buffer * d) {
     auto buf  = static_cast<ggml_backend_hexagon_buffer_context *>(t->buffer->context);
     auto sess = buf->sess;
@@ -2360,10 +2327,6 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
     const struct ggml_tensor * src1 = op->src[1];
     const struct ggml_tensor * dst  = op;
 
-    auto src0_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src0->buffer->context);
-    auto src1_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src1->buffer->context);
-    auto dst_buf  = static_cast<ggml_backend_hexagon_buffer_context *>(dst->buffer->context);
-
     uint64_t t1, t2;
     t1 = ggml_time_us();
 
@@ -2385,55 +2348,27 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
     }
 
     dspqueue_buffer bufs[3];
-    memset(bufs, 0, sizeof(bufs));
 
     // First buffer Weights.
     // The content is static, there is no need to do any cache management
-    bufs[0].fd     = src0_buf->fd;
-    bufs[0].ptr    = src0->data;
-    bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
-    bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = 0;
+    dspqueue_buffers_init(bufs, src0, false, false);
 
     // Second buffer Input Activations. This is a buffer that the CPU
     // writes and the DSP reads, so we'll need to flush CPU caches and
     // invalidate DSP ones. On platforms with I/O coherency support the
     // framework will automatically skip cache operations where possible.
-    bufs[1].fd     = src1_buf->fd;
-    bufs[1].ptr    = src1->data;
-    bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
-    bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
+    dspqueue_buffers_init(&bufs[1], src1, true, true);
 
     // Third buffer Output Activations. We'll handle DSP
     // cache maintenance in the response message but need to flush
     // CPU caches to ensure any previously written dirty lines are
     // written out before writes from the DSP start.
-    bufs[2].fd     = dst_buf->fd;
-    bufs[2].ptr    = dst->data;
-    bufs[2].offset = (uint8_t *) dst->data - dst_buf->base;
-    bufs[2].size   = ggml_nbytes(dst);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    dspqueue_buffers_init(&bufs[2], dst, true, false);
 
-    // Primary DSP session from the src0 (normally weight) tensor
-    auto sess = src0_buf->sess;
+    auto * sess = get_session_from_tensor(src0);
 
     if (opt_verbose) {
-        char dims[64 * GGML_MAX_SRC];
-        char strides[64 * GGML_MAX_SRC];
-        char types[16 * GGML_MAX_SRC];
-        char buffs[64 * GGML_MAX_SRC];
-        char names[64 * GGML_MAX_SRC];
-
-        hex_format_op_dims(dims, op);
-        hex_format_op_strides(strides, op);
-        hex_format_op_types(types, op);
-        hex_format_op_buffs(buffs, op);
-        hex_format_op_names(names, op);
-
-        HEX_VERBOSE("ggml-hex: %s %s: %s : %s : %s : %s : %s: flags 0x%x\n", sess->name.c_str(), ggml_op_name(op->op),
-                    names, dims, types, strides, buffs, req.flags);
+        hex_print_op_info(op, sess, req.flags);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2463,11 +2398,6 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     const struct ggml_tensor * src2 = op->src[2];
     const struct ggml_tensor * dst  = op;
 
-    auto src0_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src0->buffer->context);
-    auto src1_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src1->buffer->context);
-    auto src2_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src2->buffer->context);
-    auto dst_buf  = static_cast<ggml_backend_hexagon_buffer_context *>(dst->buffer->context);
-
     uint64_t t1, t2;
     t1 = ggml_time_us();
 
@@ -2490,66 +2420,32 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     }
 
     dspqueue_buffer bufs[4];
-    memset(bufs, 0, sizeof(bufs));
-
     // First buffer Weights.
     // The content is static, there is no need to do any cache management
-    bufs[0].fd     = src0_buf->fd;
-    bufs[0].ptr    = src0->data;
-    bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
-    bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = 0;
+    dspqueue_buffers_init(bufs, src0, false, false);
 
     // Second buffer Input Activations. This is a buffer that the CPU
     // writes and the DSP reads, so we'll need to flush CPU caches and
     // invalidate DSP ones. On platforms with I/O coherency support the
     // framework will automatically skip cache operations where possible.
-    bufs[1].fd     = src1_buf->fd;
-    bufs[1].ptr    = src1->data;
-    bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
-    bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
+    dspqueue_buffers_init(&bufs[1], src1, true, true);
 
     // Third buffer expert IDs. This is a buffer that the CPU
     // writes and the DSP reads, so we'll need to flush CPU caches and
     // invalidate DSP ones. On platforms with I/O coherency support the
     // framework will automatically skip cache operations where possible.
-    bufs[2].fd     = src2_buf->fd;
-    bufs[2].ptr    = src2->data;
-    bufs[2].offset = (uint8_t *) src2->data - src2_buf->base;
-    bufs[2].size   = ggml_nbytes(src2);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
+    dspqueue_buffers_init(&bufs[2], src2, true, true);
 
     // Forth buffer Output Activations. We'll handle DSP
     // cache maintenance in the response message but need to flush
     // CPU caches to ensure any previously written dirty lines are
     // written out before writes from the DSP start.
-    bufs[3].fd     = dst_buf->fd;
-    bufs[3].ptr    = dst->data;
-    bufs[3].offset = (uint8_t *) dst->data - dst_buf->base;
-    bufs[3].size   = ggml_nbytes(dst);
-    bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    dspqueue_buffers_init(&bufs[3], dst, true, false);
 
-    // Primary DSP session from the src0 (normally weight) tensor
-    auto sess = src0_buf->sess;
+    auto * sess = get_session_from_tensor(src0);
 
     if (opt_verbose) {
-        char dims[64 * GGML_MAX_SRC];
-        char strides[64 * GGML_MAX_SRC];
-        char types[16 * GGML_MAX_SRC];
-        char buffs[64 * GGML_MAX_SRC];
-        char names[64 * GGML_MAX_SRC];
-
-        hex_format_op_dims(dims, op);
-        hex_format_op_types(types, op);
-        hex_format_op_buffs(buffs, op);
-        hex_format_op_names(names, op);
-
-        HEX_VERBOSE("ggml-hex: %s %s: %s : %s : %s : %s : %s: flags 0x%x\n", sess->name.c_str(), ggml_op_name(op->op),
-                    names, dims, types, strides, buffs, req.flags);
-
+        hex_print_op_info(op, sess, req.flags);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2581,10 +2477,6 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
     const struct ggml_tensor * src1 = node->src[1];
     const struct ggml_tensor * dst  = node;
 
-    auto src0_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src0->buffer->context);
-    auto src1_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src1->buffer->context);
-    auto dst_buf  = static_cast<ggml_backend_hexagon_buffer_context *>(dst->buffer->context);
-
     uint64_t t1 = 0;
     uint64_t t2 = 0;
 
@@ -2621,60 +2513,30 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
     init_htp_tensor(&req.dst, dst);
 
     dspqueue_buffer bufs[3];
-    memset(bufs, 0, sizeof(bufs));
-
     // First buffer = First Operand of Binary op
     // This is a buffer that the CPU writes and the DSP reads, so we'll
     // need to flush CPU caches and invalidate DSP ones. On platforms
     // with I/O coherency support the framework will automatically skip
     // cache operations where possible.
-    bufs[0].fd     = src0_buf->fd;
-    bufs[0].ptr    = src0->data;
-    bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
-    bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
+    dspqueue_buffers_init(bufs, src0, true, true);
 
     // Second buffer = Second Operand of Binary op
     // This is a buffer that the CPU writes and the DSP reads, so we'll
     // need to flush CPU caches and invalidate DSP ones. On platforms
     // with I/O coherency support the framework will automatically skip
     // cache operations where possible.
-    bufs[1].fd     = src1_buf->fd;
-    bufs[1].ptr    = src1->data;
-    bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
-    bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
+    dspqueue_buffers_init(&bufs[1], src1, true, true);
 
     // Third buffer = Output Activations. We'll handle DSP
     // cache maintenance in the response message but need to flush
     // CPU caches to ensure any previously written dirty lines are
     // written out before writes from the DSP start.
-    bufs[2].fd     = dst_buf->fd;
-    bufs[2].ptr    = dst->data;
-    bufs[2].offset = (uint8_t *) dst->data - dst_buf->base;
-    bufs[2].size   = ggml_nbytes(dst);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    dspqueue_buffers_init(&bufs[2], dst, true, false);
 
-    // Primary DSP session from the src0 tensor
-    ggml_hexagon_session * sess = src0_buf->sess;
+    auto * sess = get_session_from_tensor(src0);
 
     if (opt_verbose) {
-        char dims[64 * GGML_MAX_SRC];
-        char strides[16 * GGML_MAX_SRC];
-        char types[16 * GGML_MAX_SRC];
-        char buffs[64 * GGML_MAX_SRC];
-        char names[64 * GGML_MAX_SRC];
-
-        hex_format_op_dims(dims, op);
-        hex_format_op_strides(strides, op);
-        hex_format_op_types(types, op);
-        hex_format_op_buffs(buffs, op);
-        hex_format_op_names(names, op);
-
-        HEX_VERBOSE("ggml-hex: %s %s : %s : %s : %s : %s : %s : flags 0x%x\n", sess->name.c_str(),
-                    ggml_op_name(node->op), names, dims, types, strides, buffs, req.flags);
+        hex_print_op_info(op, sess, req.flags);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2705,11 +2567,6 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     const struct ggml_tensor * src2 = node->src[2];
     const struct ggml_tensor * dst  = node;
 
-    auto src0_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src0->buffer->context);
-    auto src1_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src1->buffer->context);
-    auto src2_buf = static_cast<ggml_backend_hexagon_buffer_context *>(src2->buffer->context);
-    auto dst_buf  = static_cast<ggml_backend_hexagon_buffer_context *>(dst->buffer->context);
-
     uint64_t t1 = 0;
     uint64_t t2 = 0;
 
@@ -2741,58 +2598,19 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     init_htp_tensor(&req.dst, dst);
 
     dspqueue_buffer bufs[4];
-    memset(bufs, 0, sizeof(bufs));
-
     // First buffer = input activations
-    bufs[0].fd     = src0_buf->fd;
-    bufs[0].ptr    = src0->data;
-    bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
-    bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
-
+    dspqueue_buffers_init(bufs, src0, true, true);
     // Second buffer = experts bias
-    bufs[1].fd     = src1_buf->fd;
-    bufs[1].ptr    = src1->data;
-    bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
-    bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
-
+    dspqueue_buffers_init(&bufs[1], src1, true, true);
     // Third buffer = activated experts
-    bufs[2].fd     = src2_buf->fd;
-    bufs[2].ptr    = src2->data;
-    bufs[2].offset = (uint8_t *) src2->data - src2_buf->base;
-    bufs[2].size   = ggml_nbytes(src2);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                     DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
-
+    dspqueue_buffers_init(&bufs[2], src2, true, true);
     // Forth buffer = output activations
-    bufs[3].fd     = dst_buf->fd;
-    bufs[3].ptr    = dst->data;
-    bufs[3].offset = (uint8_t *) dst->data - dst_buf->base;
-    bufs[3].size   = ggml_nbytes(dst);
-    bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    dspqueue_buffers_init(&bufs[3], dst, true, true);
 
-    // Primary DSP session from the src0 tensor
-    ggml_hexagon_session * sess = src0_buf->sess;
+    auto * sess = get_session_from_tensor(src0);
 
     if (opt_verbose) {
-        char dims[64 * GGML_MAX_SRC];
-        char strides[16 * GGML_MAX_SRC];
-        char types[16 * GGML_MAX_SRC];
-        char buffs[64 * GGML_MAX_SRC];
-        char names[64 * GGML_MAX_SRC];
-
-        hex_format_op_dims(dims, op);
-        hex_format_op_strides(strides, op);
-        hex_format_op_types(types, op);
-        hex_format_op_buffs(buffs, op);
-        hex_format_op_names(names, op);
-
-        HEX_VERBOSE("ggml-hex: %s %s : %s : %s : %s : %s : %s : flags 0x%x\n", sess->name.c_str(),
-                    ggml_op_name(node->op), names, dims, types, strides, buffs, req.flags);
-
+        hex_print_op_info(op, sess, req.flags);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2886,71 +2704,33 @@ static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
     }
 
     dspqueue_buffer bufs[3];
-    int             n_bufs = 0;
-
-    memset(bufs, 0, sizeof(bufs));
 
     // First buffer = Only Operand of Unary op
     // This is a buffer that the CPU writes and the DSP reads, so we'll
     // need to flush CPU caches and invalidate DSP ones. On platforms
     // with I/O coherency support the framework will automatically skip
     // cache operations where possible.
-    auto src0_buf       = static_cast<ggml_backend_hexagon_buffer_context *>(src0->buffer->context);
-    bufs[n_bufs].fd     = src0_buf->fd;
-    bufs[n_bufs].ptr    = src0->data;
-    bufs[n_bufs].offset = (uint8_t *) src0->data - src0_buf->base;
-    bufs[n_bufs].size   = ggml_nbytes(src0);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
-    ++n_bufs;
+    size_t n_bufs = dspqueue_buffers_init(bufs, src0, true, true);
 
-    if (src1) {
-        // Second buffer = Second Operand of Binary op
-        // This is a buffer that the CPU writes and the DSP reads, so we'll
-        // need to flush CPU caches and invalidate DSP ones. On platforms
-        // with I/O coherency support the framework will automatically skip
-        // cache operations where possible.
-        auto src1_buf       = static_cast<ggml_backend_hexagon_buffer_context *>(src1->buffer->context);
-        bufs[n_bufs].fd     = src1_buf->fd;
-        bufs[n_bufs].ptr    = src1->data;
-        bufs[n_bufs].offset = (uint8_t *) src1->data - src1_buf->base;
-        bufs[n_bufs].size   = ggml_nbytes(src1);
-        bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                              DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
-        ++n_bufs;
-    }
+    // Second buffer(nullable) = Second Operand of Binary op
+    // This is a buffer that the CPU writes and the DSP reads, so we'll
+    // need to flush CPU caches and invalidate DSP ones. On platforms
+    // with I/O coherency support the framework will automatically skip
+    // cache operations where possible.
+    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], src1, true, true);
 
     // Second or third buffer = Output Activations. We'll handle DSP
     // Second buffer = Output Activations. We'll handle DSP
     // cache maintenance in the response message but need to flush
     // CPU caches to ensure any previously written dirty lines are
     // written out before writes from the DSP start.
-    auto dst_buf        = static_cast<ggml_backend_hexagon_buffer_context *>(dst->buffer->context);
-    bufs[n_bufs].fd     = dst_buf->fd;
-    bufs[n_bufs].ptr    = dst->data;
-    bufs[n_bufs].offset = (uint8_t *) dst->data - dst_buf->base;
-    bufs[n_bufs].size   = ggml_nbytes(dst);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
-    ++n_bufs;
+    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], dst, true, false);
 
     // Primary DSP session from the src0 tensor
-    ggml_hexagon_session * sess = src0_buf->sess;
+    auto * sess = get_session_from_tensor(src0);
 
     if (opt_verbose) {
-        char dims[64 * GGML_MAX_SRC];
-        char strides[64 * GGML_MAX_SRC];
-        char types[16 * GGML_MAX_SRC];
-        char buffs[64 * GGML_MAX_SRC];
-        char names[64 * GGML_MAX_SRC];
-
-        hex_format_op_dims(dims, op);
-        hex_format_op_strides(strides, op);
-        hex_format_op_types(types, op);
-        hex_format_op_buffs(buffs, op);
-        hex_format_op_names(names, op);
-
-        HEX_VERBOSE("ggml-hex: %s %s : %s : %s : %s : %s : %s : flags 0x%x\n", sess->name.c_str(), ggml_op_name(op->op),
-                    names, dims, types, strides, buffs, req.flags);
+        hex_print_op_info(op, sess, req.flags);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             if (src1) {
@@ -3023,85 +2803,40 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
     }
 
     dspqueue_buffer bufs[4];
-    int             n_bufs = 0;
-
-    memset(bufs, 0, sizeof(bufs));
 
     // First buffer
     // This is a buffer that the CPU writes and the DSP reads, so we'll
     // need to flush CPU caches and invalidate DSP ones. On platforms
     // with I/O coherency support the framework will automatically skip
     // cache operations where possible.
-    auto src0_buf       = static_cast<ggml_backend_hexagon_buffer_context *>(src0->buffer->context);
-    bufs[n_bufs].fd     = src0_buf->fd;
-    bufs[n_bufs].ptr    = src0->data;
-    bufs[n_bufs].offset = (uint8_t *) src0->data - src0_buf->base;
-    bufs[n_bufs].size   = ggml_nbytes(src0);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
-    ++n_bufs;
+    size_t n_bufs = dspqueue_buffers_init(bufs, src0, true, true);
 
     // Second buffer
     // This is a buffer that the CPU writes and the DSP reads, so we'll
     // need to flush CPU caches and invalidate DSP ones. On platforms
     // with I/O coherency support the framework will automatically skip
     // cache operations where possible.
-    auto src1_buf       = static_cast<ggml_backend_hexagon_buffer_context *>(src1->buffer->context);
-    bufs[n_bufs].fd     = src1_buf->fd;
-    bufs[n_bufs].ptr    = src1->data;
-    bufs[n_bufs].offset = (uint8_t *) src1->data - src1_buf->base;
-    bufs[n_bufs].size   = ggml_nbytes(src1);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
-    ++n_bufs;
+    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], src1, true, true);
 
-    if (src2) {
-        // Third buffer
-        // This is a buffer that the CPU writes and the DSP reads, so we'll
-        // need to flush CPU caches and invalidate DSP ones. On platforms
-        // with I/O coherency support the framework will automatically skip
-        // cache operations where possible.
-        auto src2_buf       = static_cast<ggml_backend_hexagon_buffer_context *>(src2->buffer->context);
-        bufs[n_bufs].fd     = src2_buf->fd;
-        bufs[n_bufs].ptr    = src2->data;
-        bufs[n_bufs].offset = (uint8_t *) src2->data - src2_buf->base;
-        bufs[n_bufs].size   = ggml_nbytes(src2);
-        bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
-                              DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
-        ++n_bufs;
-    }
+    // Third buffer(nullable)
+    // This is a buffer that the CPU writes and the DSP reads, so we'll
+    // need to flush CPU caches and invalidate DSP ones. On platforms
+    // with I/O coherency support the framework will automatically skip
+    // cache operations where possible.
+    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], src2, true, true);
 
     // Final buffer = Output Activations. We'll handle DSP
     // Second buffer = Output Activations. We'll handle DSP
     // cache maintenance in the response message but need to flush
     // CPU caches to ensure any previously written dirty lines are
     // written out before writes from the DSP start.
-    auto dst_buf        = static_cast<ggml_backend_hexagon_buffer_context *>(dst->buffer->context);
-    bufs[n_bufs].fd     = dst_buf->fd;
-    bufs[n_bufs].ptr    = dst->data;
-    bufs[n_bufs].offset = (uint8_t *) dst->data - dst_buf->base;
-    bufs[n_bufs].size   = ggml_nbytes(dst);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
-    ++n_bufs;
+    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], dst, true, false);
 
     // Primary DSP session from the src0 tensor
-    ggml_hexagon_session * sess = src0_buf->sess;
+    auto * sess = get_session_from_tensor(src0);
 
     if (opt_verbose) {
-        char dims[64 * GGML_MAX_SRC];
-        char strides[64 * GGML_MAX_SRC];
-        char types[16 * GGML_MAX_SRC];
-        char buffs[64 * GGML_MAX_SRC];
-        char names[64 * GGML_MAX_SRC];
-
-        hex_format_op_dims(dims, op);
-        hex_format_op_strides(strides, op);
-        hex_format_op_types(types, op);
-        hex_format_op_buffs(buffs, op);
-        hex_format_op_names(names, op);
-
-        HEX_VERBOSE("ggml-hex: %s %s : %s : %s : %s : %s : %s : flags 0x%x\n", sess->name.c_str(), ggml_op_name(op->op),
-                    names, dims, types, strides, buffs, req.flags);
+        hex_print_op_info(op, sess, req.flags);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             if (src1) {

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

@@ -390,6 +390,12 @@ int get_hex_arch_ver(int domain, int * arch) {
     }
 
     switch (arch_ver.capability & 0xff) {
+        case 0x68:
+            *arch = 68;
+            return 0;
+        case 0x69:
+            *arch = 69;
+            return 0;
         case 0x73:
             *arch = 73;
             return 0;

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

@@ -106,33 +106,32 @@ static void glu_swiglu_fp32_per_thread(const struct htp_tensor * src0,
     t1 = HAP_perf_get_qtimer_count();
 
     int is_aligned = 1;
-    int opt_path   = 0;
     if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) {
         is_aligned = 0;
         FARF(HIGH, "swiglu-f32: unaligned addresses in elementwise op, possibly slower execution\n");
     }
-    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
-        opt_path = 1;
-    }
 
     const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
     const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
     uint8_t * restrict data_dst        = (uint8_t *) dst->data;
 
-    bool src1_valid = src1->ne[0];
+    const bool src1_valid = src1->ne[0];
+    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
     if (!src1_valid) {
-        data_src1     = data_src0;
-        src1_row_size = src0_row_size;
+        const int32_t swapped = op_params[1];
+        data_src1             = data_src0;
+        src1_row_size         = src0_row_size;
+
+        const size_t nc_in_bytes = nc * SIZEOF_FP32;
+        data_src0 += swapped ? nc_in_bytes : 0;
+        data_src1 += swapped ? 0 : nc_in_bytes;
     }
 
     uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size);
     uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_row_size);
     uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_row_size);
 
-    const int32_t swapped = op_params[1];
-
-    const int nc = (src1_valid) ? ne0 : ne0 / 2;
-
+    const bool opt_path = ((1 == is_aligned) && !(nb01 & (VLEN - 1)));
     for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
         const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size));
         const float * restrict src1 = (float *) (data_src1 + (ir * src1_row_size));
@@ -142,12 +141,7 @@ static void glu_swiglu_fp32_per_thread(const struct htp_tensor * src0,
             htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size);
         }
 
-        if (!src1_valid) {
-            src0 += swapped ? nc : 0;
-            src1 += swapped ? 0 : nc;
-        }
-
-        if (1 == opt_path) {
+        if (opt_path) {
             hvx_fast_sigmoid_f32((const uint8_t *) src0, (uint8_t *) src0_spad_data, nc);
             hvx_mul_mul_f32_opt((const uint8_t *) src0, (const uint8_t *) src0_spad_data, (const uint8_t *) src1,
                                 (uint8_t *) dst, nc);
@@ -218,7 +212,7 @@ static void glu_swiglu_oai_fp32_per_thread(const struct htp_tensor * src0,
     const float   alpha   = ((const float *) (op_params))[2];
     const float   limit   = ((const float *) (op_params))[3];
 
-    const int nc = (src1_valid) ? ne0 : ne0 / 2;
+    const int nc = (src1_valid) ? ne00 : ne00 / 2;
 
     for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
         const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size));

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

@@ -66,6 +66,13 @@ static inline bool dma_queue_push(dma_queue *  q,
     desc->desctype       = HEXAGON_UDMA_DESC_DESCTYPE_TYPE1;
     desc->dstbypass      = 1;
     desc->srcbypass      = 1;
+#if __HVX_ARCH__ >= 73
+    desc->dstbypass      = 1;
+    desc->srcbypass      = 1;
+#else
+    desc->dstbypass      = 0;
+    desc->srcbypass      = 1;
+#endif
     desc->order          = 0;
     desc->dstate         = HEXAGON_UDMA_DESC_DSTATE_INCOMPLETE;
     desc->src            = (void *) src;

ggml/src/ggml-hexagon/htp/hvx-exp.c

@@ -16,6 +16,14 @@
 #include "hvx-utils.h"
 #include "ops-utils.h"
 
+static inline HVX_Vector hvx_vec_exp_fp32_guard(HVX_Vector in_vec, HVX_Vector max_exp, HVX_Vector inf) {
+    const HVX_VectorPred pred0 = Q6_Q_vcmp_gt_VsfVsf(in_vec, max_exp);
+
+    HVX_Vector out = hvx_vec_exp_fp32(in_vec);
+
+    return Q6_V_vmux_QVV(pred0, inf, out);
+}
+
 void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems, bool negate) {
     int left_over       = num_elems & (VLEN_FP32 - 1);
     int num_elems_whole = num_elems - left_over;
@@ -34,6 +42,12 @@ void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int
 
     HVX_Vector vec_out = Q6_V_vzero();
 
+    static const float kInf    = INFINITY;
+    static const float kMaxExp = 88.02f;  // log(INF)
+
+    const HVX_Vector max_exp = hvx_vec_splat_fp32(kMaxExp);
+    const HVX_Vector inf     = hvx_vec_splat_fp32(kInf);
+
     if (0 == unaligned_loop) {
         HVX_Vector * p_vec_in1 = (HVX_Vector *) src;
         HVX_Vector * p_vec_out = (HVX_Vector *) dst;
@@ -42,9 +56,9 @@ void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int
         for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
             if (true == negate) {
                 HVX_Vector neg_vec_in = hvx_vec_neg_fp32(*p_vec_in1++);
-                *p_vec_out++          = hvx_vec_exp_fp32(neg_vec_in);
+                *p_vec_out++          = hvx_vec_exp_fp32_guard(neg_vec_in, max_exp, inf);
             } else {
-                *p_vec_out++ = hvx_vec_exp_fp32(*p_vec_in1++);
+                *p_vec_out++ = hvx_vec_exp_fp32_guard(*p_vec_in1++, max_exp, inf);
             }
         }
     } else {
@@ -54,9 +68,9 @@ void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int
 
             if (true == negate) {
                 HVX_Vector neg_vec_in                    = hvx_vec_neg_fp32(in);
-                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32(neg_vec_in);
+                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32_guard(neg_vec_in, max_exp, inf);
             } else {
-                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32(in);
+                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32_guard(in, max_exp, inf);
             }
         }
     }
@@ -70,9 +84,9 @@ void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int
         if (true == negate) {
             HVX_Vector neg_vec_in = hvx_vec_neg_fp32(in);
 
-            vec_out = hvx_vec_exp_fp32(neg_vec_in);
+            vec_out = hvx_vec_exp_fp32_guard(neg_vec_in, max_exp, inf);
         } else {
-            vec_out = hvx_vec_exp_fp32(in);
+            vec_out = hvx_vec_exp_fp32_guard(in, max_exp, inf);
         }
 
         hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, vec_out);

ggml/src/ggml-hexagon/htp/hvx-inverse.c

@@ -16,6 +16,15 @@
 #include "hvx-utils.h"
 #include "ops-utils.h"
 
+static inline HVX_Vector hvx_vec_inverse_fp32_guard(HVX_Vector v_sf, HVX_Vector nan_inf_mask) {
+    HVX_Vector out = hvx_vec_inverse_fp32(v_sf);
+
+    HVX_Vector           masked_out = Q6_V_vand_VV(out, nan_inf_mask);
+    const HVX_VectorPred pred       = Q6_Q_vcmp_eq_VwVw(nan_inf_mask, masked_out);
+
+    return Q6_V_vmux_QVV(pred, Q6_V_vzero(), out);
+}
+
 void hvx_inverse_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems) {
     int left_over       = num_elems & (VLEN_FP32 - 1);
     int num_elems_whole = num_elems - left_over;
@@ -32,19 +41,22 @@ void hvx_inverse_f32(const uint8_t * restrict src, uint8_t * restrict dst, const
         FARF(HIGH, "hvx_inverse_f32: unaligned loop in hvx op, possibly slower execution\n");
     }
 
+    static const uint32_t kNanInfMask  = 0x7f800000;
+    const HVX_Vector      nan_inf_mask = Q6_V_vsplat_R(kNanInfMask);
+
     if (0 == unaligned_loop) {
         HVX_Vector * p_vec_in  = (HVX_Vector *) src;
         HVX_Vector * p_vec_out = (HVX_Vector *) dst;
 
         #pragma unroll(4)
         for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            *p_vec_out++ = hvx_vec_inverse_fp32(*p_vec_in++);
+            *p_vec_out++ = hvx_vec_inverse_fp32_guard(*p_vec_in++, nan_inf_mask);
         }
     } else {
         #pragma unroll(4)
         for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
             HVX_Vector in                            = *(HVX_UVector *) (src + i * SIZEOF_FP32);
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_inverse_fp32(in);
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_inverse_fp32_guard(in, nan_inf_mask);
         }
     }
 
@@ -53,7 +65,7 @@ void hvx_inverse_f32(const uint8_t * restrict src, uint8_t * restrict dst, const
         float *       dstf = (float *) dst + num_elems_whole;
 
         HVX_Vector in  = *(HVX_UVector *) srcf;
-        HVX_Vector out = hvx_vec_inverse_fp32(in);
+        HVX_Vector out = hvx_vec_inverse_fp32_guard(in, nan_inf_mask);
 
         hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, out);
     }

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

@@ -401,7 +401,9 @@ void hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t *
         FARF(HIGH, "hvx_add_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
     }
 
-    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
+    static const float kInf    = INFINITY;
+    const HVX_Vector   inf     = hvx_vec_splat_fp32(kInf);
+    HVX_Vector         val_vec = hvx_vec_splat_fp32(val);
 
     if (0 == unaligned_loop) {
         HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
@@ -409,17 +411,24 @@ void hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t *
 
         #pragma unroll(4)
         for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*vec_in1++, val_vec);
-            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
+            HVX_Vector           in       = *vec_in1++;
+            const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, in);
+            HVX_Vector           v        = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
+            v                             = Q6_Vsf_equals_Vqf32(v);
+            v                             = Q6_V_vmux_QVV(pred_inf, inf, v);
+            *vec_out++                    = v;
         }
     } else {
         #pragma unroll(4)
         for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
             HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
 
-            HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
+            const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, in);
+            HVX_Vector           out      = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
+            out                           = Q6_Vsf_equals_Vqf32(out);
+            out                           = Q6_V_vmux_QVV(pred_inf, inf, out);
 
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = out;
         }
     }
 
@@ -429,8 +438,12 @@ void hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t *
 
         HVX_Vector in = *(HVX_UVector *) srcf;
 
-        HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
+        const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, in);
+        HVX_Vector           out      = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
+        out                           = Q6_Vsf_equals_Vqf32(out);
+        out                           = Q6_V_vmux_QVV(pred_inf, inf, out);
+
+        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, out);
     }
 }
 

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

@@ -12,6 +12,35 @@
 #define VLEN_FP32   (VLEN / SIZEOF_FP32)
 #define VLEN_FP16   (VLEN / SIZEOF_FP16)
 
+typedef union {
+    HVX_Vector v;
+    uint8_t    b[VLEN];
+    uint16_t   h[VLEN_FP16];
+    uint32_t   w[VLEN_FP32];
+    __fp16     fp16[VLEN_FP16];
+    float      fp32[VLEN_FP32];
+} __attribute__((aligned(VLEN), packed)) HVX_VectorAlias;
+
+/* Q6_Vsf_equals_Vw is only available on v73+.*/
+#if __HVX_ARCH__ < 73
+static inline HVX_Vector int32_to_qfloat(HVX_Vector const in)
+{
+    HVX_Vector const vzero = Q6_V_vzero();
+    HVX_VectorPred is_zero = Q6_Q_vcmp_eq_VwVw(in, vzero);
+    HVX_Vector lshift = Q6_Vw_vnormamt_Vw(in);
+    HVX_Vector normalized = Q6_Vw_vasl_VwVw(in, lshift);
+    HVX_Vector vexp = Q6_Vw_vsub_VwVw(Q6_V_vsplat_R(0x7f + 30), lshift);
+    HVX_Vector mant = Q6_V_vand_VV(Q6_V_vsplat_R(0xFFFFFF00), normalized);
+    HVX_Vector ret = Q6_V_vmux_QVV(is_zero, vzero, Q6_Vw_vadd_VwVw(mant, vexp));
+    return ret;
+}
+
+static inline HVX_Vector Q6_Vsf_equals_Vw(HVX_Vector const in)
+{
+    return Q6_Vsf_equals_Vqf32(int32_to_qfloat(in));
+}
+#endif
+
 static inline HVX_Vector hvx_vec_splat_fp32(float i) {
     union {
         float   f;
@@ -243,19 +272,16 @@ static __attribute__((always_inline)) int32_t is_in_one_chunk(void * addr, uint3
 }
 
 static void hvx_vec_dump_fp16_n(char * pref, HVX_Vector v, uint32_t n) {
-    union {
-        HVX_Vector v;
-        __fp16 d[64];
-    } u = { .v = v };
+    HVX_VectorAlias u = { .v = v };
 
     const uint32_t n0 = n / 16;
     const uint32_t n1 = n % 16;
     int            i  = 0;
     for (; i < n0; i++) {
-        htp_dump_fp16_line(pref, u.d + (16 * i), 16);
+        htp_dump_fp16_line(pref, u.fp16 + (16 * i), 16);
     }
     if (n1) {
-        htp_dump_fp16_line(pref, u.d + (16 * i), n1);
+        htp_dump_fp16_line(pref, u.fp16 + (16 * i), n1);
     }
 }
 
@@ -411,8 +437,8 @@ static inline HVX_Vector hvx_vec_fp32_reduce_sum_n(HVX_Vector in, unsigned int n
 
     HVX_Vector sum = in, sum_t;
     while (width < total) {
-        sum_t = Q6_V_vror_VR(sum, width);       // rotate right
-        sum   = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(sum, sum_t)); // elementwise sum
+        sum_t = Q6_V_vror_VR(sum, width);                               // rotate right
+        sum   = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(sum, sum_t));  // elementwise sum
         width = width << 1;
     }
     return sum;
@@ -491,7 +517,7 @@ static inline HVX_Vector hvx_vec_abs_fp16(HVX_Vector v) {
 static inline HVX_Vector hvx_vec_neg_fp16(HVX_Vector v) {
     // neg by setting the fp16 sign bit
     HVX_Vector mask = Q6_Vh_vsplat_R(0x8000);
-    return Q6_V_vor_VV(v, mask);
+    return Q6_V_vxor_VV(v, mask);
 }
 
 static inline HVX_Vector hvx_vec_abs_fp32(HVX_Vector v) {
@@ -506,7 +532,7 @@ static inline HVX_Vector hvx_vec_neg_fp32(HVX_Vector v) {
 #else
     // neg by setting the fp32 sign bit
     HVX_Vector mask = Q6_V_vsplat_R(0x80000000);
-    return Q6_V_vor_VV(v, mask);
+    return Q6_V_vxor_VV(v, mask);
 #endif  // __HTP_ARCH__ > 75
 }
 
@@ -934,6 +960,18 @@ static inline HVX_Vector hvx_vec_rsqrt_fp32(HVX_Vector in_vec) {
     return Q6_Vsf_equals_Vqf32(temp);
 }
 
+static inline HVX_Vector hvx_vec_fast_sigmoid_fp32_guard(HVX_Vector v,
+                                                         HVX_Vector one,
+                                                         HVX_Vector max_exp,
+                                                         HVX_Vector min_exp) {
+    const HVX_VectorPred pred_max = Q6_Q_vcmp_gt_VsfVsf(max_exp, v);
+    const HVX_VectorPred pred_min = Q6_Q_vcmp_gt_VsfVsf(v, min_exp);
+
+    HVX_Vector out = hvx_vec_fast_sigmoid_fp32(v);
+    out            = Q6_V_vmux_QVV(pred_max, out, one);
+    return Q6_V_vmux_QVV(pred_min, out, Q6_V_vzero());
+}
+
 static inline void hvx_fast_sigmoid_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems) {
     int step_of_1 = num_elems >> 5;
     int remaining = num_elems - step_of_1 * VLEN_FP32;
@@ -943,9 +981,16 @@ static inline void hvx_fast_sigmoid_f32(const uint8_t * restrict src, uint8_t *
     const HVX_Vector * restrict v_src = (HVX_Vector *) src;
     HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
 
+    static const float kMinExp = -87.f;  // 0
+    static const float kMaxExp = 87.f;   // 1
+
+    const HVX_Vector one     = hvx_vec_splat_fp32(1.f);
+    const HVX_Vector max_exp = hvx_vec_splat_fp32(kMaxExp);
+    const HVX_Vector min_exp = hvx_vec_splat_fp32(kMinExp);
+
     #pragma unroll(4)
     for (int i = 0; i < step_of_1; i++) {
-        v_dst[i] = hvx_vec_fast_sigmoid_fp32(v_src[i]);
+        v_dst[i] = hvx_vec_fast_sigmoid_fp32_guard(v_src[i], one, max_exp, min_exp);
     }
 }
 

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

@@ -143,16 +143,25 @@ AEEResult htp_iface_disable_etm(remote_handle64 handle) {
 }
 
 static int vtcm_acquire(struct htp_context * ctx) {
+    int err;
     if (!ctx->vtcm_valid) {
         // Temporarily bump thread priority to make sure it's higher than other sessions.
         // This way the resource manager will notify the other thread to release VTCM.
         // Note that we need to reaquire VTCM at normal priority for this to work next time.
         qurt_thread_set_priority(qurt_thread_get_id(), ctx->thread_prio - 10);
-        HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000);
+        err = HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000);
+        if (err != 0) {
+            FARF(ERROR, "Failed to acquire VTCM: 0x%08x", (unsigned)err);
+            abort();
+        }
         HAP_compute_res_release_cached(ctx->vtcm_rctx);
         qurt_thread_set_priority(qurt_thread_get_id(), ctx->thread_prio);
 
-        HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000);
+        err = HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000);
+        if (err != 0) {
+            FARF(ERROR, "Failed to acquire VTCM: 0x%08x", (unsigned)err);
+            abort();
+        }
         ctx->vtcm_valid = true;
     }
 
@@ -201,7 +210,7 @@ static int vtcm_alloc(struct htp_context * ctx) {
     HAP_compute_res_attr_init(&attr);
     HAP_compute_res_attr_set_serialize(&attr, 0);
     HAP_compute_res_attr_set_cache_mode(&attr, 1);
-    HAP_compute_res_attr_set_vtcm_param_v2(&attr, vtcm_size, vtcm_size, vtcm_size);
+    HAP_compute_res_attr_set_vtcm_param_v2(&attr, vtcm_size, 0, vtcm_size);
     HAP_compute_res_attr_set_release_callback(&attr, vtcm_release_callback, (void *) ctx);
     HAP_compute_res_attr_set_hmx_param(&attr, 1);
 

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

@@ -24,6 +24,10 @@
 #include "hvx-utils.h"
 #include "ops-utils.h"
 
+// Redefined the types GGML_ROPE_TYPE_NORMAL & GGML_ROPE_TYPE_NEOX as we cant include ggml.h
+#define HTP_ROPE_TYPE_NORMAL 0
+#define HTP_ROPE_TYPE_NEOX   2
+
 #define htp_rope_preamble              \
     const uint32_t ne00 = src0->ne[0]; \
     const uint32_t ne01 = src0->ne[1]; \
@@ -146,6 +150,57 @@ static void init_rope_ctx(struct rope_th_ctx * rope_ctx, struct htp_ops_context
          rope_ctx->ext_factor, rope_ctx->theta_scale, rope_ctx->attn_factor);
 }
 
+static void hvx_calc_rope_neox_f32(const float * restrict src0,
+                              float * restrict dst,
+                              const int num_elems,
+                              const float * restrict theta_cache) {
+    // for (int i = 0; i < num_elems; i += 2) {
+    //const float cos_theta = theta_cache[i + 0];
+    //const float sin_theta = theta_cache[i + 1];
+
+    //const float x0 = src[0];
+    //const float x1 = src[num_elems/2];
+
+    //dst[0] = x0*cos_theta - x1*sin_theta;
+    //dst[num_elems/2] = x0*sin_theta + x1*cos_theta;
+
+    //src += 1;
+    //dst += 1;
+    // }
+
+    const uint8_t * restrict src0_curr  = (const uint8_t *) src0;
+    const uint8_t * restrict theta_curr = (const uint8_t *) theta_cache;
+    uint8_t * restrict dst_curr         = (uint8_t *) dst;
+
+    int step_of_1 = num_elems >> 6;  // 6 because we process two vectors at once
+    int half_size = (sizeof(float) * (num_elems / 2));
+
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v0 = *(HVX_Vector *) src0_curr;
+        HVX_Vector v1 = *(HVX_Vector *) (src0_curr + half_size);
+
+        HVX_Vector v2 = *(HVX_Vector *) theta_curr;
+        HVX_Vector v3 = *(HVX_Vector *) (theta_curr + VLEN);
+
+        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);  // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
+
+        HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_hi_W(vcos_sin));
+        HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_hi_W(vcos_sin));
+
+        HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
+        HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
+
+        *(HVX_Vector *) dst_curr          = Q6_Vsf_equals_Vqf32(v4);
+        *(HVX_Vector *) (dst_curr + half_size) = Q6_Vsf_equals_Vqf32(v5);
+
+        src0_curr += VLEN;
+        theta_curr += 2 * VLEN;
+        dst_curr += VLEN;
+    }
+}
+
 static void hvx_calc_rope_f32(const float * restrict src0,
                               float * restrict dst,
                               const int num_elems,
@@ -212,6 +267,9 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
     const struct htp_tensor * src2 = &octx->src2;
     struct htp_tensor *       dst  = &octx->dst;
 
+    const int32_t mode  = rope_ctx->mode;
+    const bool is_neox  = mode & HTP_ROPE_TYPE_NEOX;
+
     htp_rope_preamble;
 
     const int32_t * pos = (const int32_t *) src1->data;
@@ -247,20 +305,35 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
                 float *       dst_data_loc = dst_data;
 
                 if (1 == opt_path) {
-                    hvx_calc_rope_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
+                    if (is_neox) {
+                        hvx_calc_rope_neox_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
+                    } else {
+                        hvx_calc_rope_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
+                    }
                 } else {
                     for (uint32_t i0 = 0; i0 < rope_ctx->n_dims; i0 += 2) {
                         const float cos_theta = wp0[i0 + 0];
                         const float sin_theta = wp0[i0 + 1];
 
-                        const float x0 = src_loc[0];
-                        const float x1 = src_loc[1];
+                        if (is_neox) {
+                            const float x0 = src_loc[0];
+                            const float x1 = src_loc[rope_ctx->n_dims/2];
 
-                        dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
-                        dst_data_loc[1] = x0 * sin_theta + x1 * cos_theta;
+                            dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
+                            dst_data_loc[rope_ctx->n_dims/2] = x0 * sin_theta + x1 * cos_theta;
 
-                        src_loc += 2;
-                        dst_data_loc += 2;
+                            src_loc += 1;
+                            dst_data_loc += 1;
+                        } else {
+                            const float x0 = src_loc[0];
+                            const float x1 = src_loc[1];
+
+                            dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
+                            dst_data_loc[1] = x0 * sin_theta + x1 * cos_theta;
+
+                            src_loc += 2;
+                            dst_data_loc += 2;
+                        }
                     }
                 }
 

ggml/src/ggml-impl.h

@@ -102,7 +102,7 @@ static bool ggml_op_is_empty(enum ggml_op op) {
     }
 }
 
-static inline float ggml_softplus(float input) {
+static inline float ggml_compute_softplus_f32(float input) {
     return (input > 20.0f) ? input : logf(1 + expf(input));
 }
 //

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

@@ -318,6 +318,44 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows(ggml_metal_librar
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_blk(ggml_metal_library_t lib, const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_CUMSUM);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_cumsum_blk_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_add(ggml_metal_library_t lib, const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_CUMSUM);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_cumsum_add_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max(ggml_metal_library_t lib, const ggml_tensor * op) {
     GGML_ASSERT(!op->src[1] || op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32);
 
@@ -943,6 +981,34 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort(ggml_metal_library
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_ARGSORT);
+
+    char base[256];
+    char name[256];
+
+    ggml_sort_order order = (ggml_sort_order) op->op_params[0];
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_merge_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
@@ -1438,6 +1504,30 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d(ggml_met
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_CONV_2D);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type         == GGML_TYPE_F32);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_conv_2d_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_UPSCALE);
 

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

@@ -113,6 +113,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_unary             (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_blk        (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -125,6 +127,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id         (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -133,6 +136,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope              (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);

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

@@ -870,6 +870,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_SUM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_CUMSUM:
         case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
@@ -885,6 +886,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return true;
         case GGML_OP_IM2COL:
             return ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
+        case GGML_OP_CONV_2D:
+            return ggml_is_contiguous(op->src[0]) &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   op->type == GGML_TYPE_F32 &&
+                   (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
         case GGML_OP_POOL_1D:
             return false;
         case GGML_OP_UPSCALE:
@@ -899,8 +905,6 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_LEAKY_RELU:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ARGSORT:
-            // TODO: Support arbitrary column width
-            return op->src[0]->ne[0] <= 1024;
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
@@ -985,7 +989,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                                 return false;
                         }
                     case GGML_TYPE_I32:
-                        return op->type == GGML_TYPE_F32;
+                        return op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_I32;
                     default:
                         return false;
                 };

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

@@ -528,6 +528,36 @@ typedef struct {
     uint64_t nb2;
 } ggml_metal_kargs_conv_transpose_2d;
 
+typedef struct {
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  IW;
+    int32_t  IH;
+    int32_t  KW;
+    int32_t  KH;
+    int32_t  IC;
+    int32_t  OC;
+    int32_t  OW;
+    int32_t  OH;
+    int32_t  N;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  p0;
+    int32_t  p1;
+    int32_t  d0;
+    int32_t  d1;
+} ggml_metal_kargs_conv_2d;
+
 typedef struct {
     uint64_t  ofs0;
     uint64_t  ofs1;
@@ -582,6 +612,45 @@ typedef struct {
     uint64_t nb3;
 } ggml_metal_kargs_sum_rows;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  net0;
+    int64_t  net1;
+    int64_t  net2;
+    int64_t  net3;
+    uint64_t nbt0;
+    uint64_t nbt1;
+    uint64_t nbt2;
+    uint64_t nbt3;
+    bool     outb;
+} ggml_metal_kargs_cumsum_blk;
+
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  net0;
+    int64_t  net1;
+    int64_t  net2;
+    int64_t  net3;
+    uint64_t nbt0;
+    uint64_t nbt1;
+    uint64_t nbt2;
+    uint64_t nbt3;
+} ggml_metal_kargs_cumsum_add;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -763,10 +832,28 @@ typedef struct {
 } ggml_metal_kargs_leaky_relu;
 
 typedef struct {
-    int64_t  ncols;
-    int64_t  ncols_pad;
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
 } ggml_metal_kargs_argsort;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  len;
+} ggml_metal_kargs_argsort_merge;
+
 typedef struct {
     int64_t  ne0;
     float    start;

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

@@ -10,6 +10,8 @@
 
 #include <cassert>
 #include <algorithm>
+#include <limits>
+#include <cmath>
 
 static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
     if (!t) {
@@ -310,6 +312,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_sum_rows(ctx, idx);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                n_fuse = ggml_metal_op_cumsum(ctx, idx);
+            } break;
         case GGML_OP_SOFT_MAX:
             {
                 n_fuse = ggml_metal_op_soft_max(ctx, idx);
@@ -364,6 +370,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_im2col(ctx, idx);
             } break;
+        case GGML_OP_CONV_2D:
+            {
+                n_fuse = ggml_metal_op_conv_2d(ctx, idx);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
                 n_fuse = ggml_metal_op_conv_transpose_1d(ctx, idx);
@@ -534,7 +544,7 @@ int ggml_metal_op_repeat(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_repeat(lib, op->type);
 
@@ -580,7 +590,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
@@ -689,7 +699,7 @@ int ggml_metal_op_scale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float scale;
     float bias;
@@ -728,7 +738,7 @@ int ggml_metal_op_clamp(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float min;
     float max;
@@ -767,7 +777,7 @@ int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     int64_t n = ggml_nelements(op);
 
@@ -797,7 +807,7 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     if (op->src[1]) {
         GGML_ASSERT(ggml_are_same_shape(op->src[0], op->src[1]));
@@ -829,18 +839,6 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
 
     const int32_t nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00/2);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-    //if (src1) {
-    //    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-    //} else {
-    //    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-    //}
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:3];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
@@ -902,7 +900,7 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_sum_rows args = {
         /*.ne00 =*/ ne00,
@@ -936,14 +934,6 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
 
     const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:0];
-    //[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-    //[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
@@ -956,6 +946,149 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_cumsum(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_pipeline_t pipeline_blk = ggml_metal_library_get_pipeline_cumsum_blk(lib, op);
+
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_blk)) {
+        nth *= 2;
+    }
+
+    GGML_ASSERT(ne00 <= nth*nth);
+
+    const int64_t net0 = (ne00 + nth - 1) / nth;
+    const int64_t net1 = ne01;
+    const int64_t net2 = ne02;
+    const int64_t net3 = ne03;
+
+    const uint64_t nbt0 = sizeof(float);
+    const uint64_t nbt1 = net0*nbt0;
+    const uint64_t nbt2 = net1*nbt1;
+    const uint64_t nbt3 = net2*nbt2;
+
+    const size_t smem = GGML_PAD(32*sizeof(float), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += ggml_nbytes(op);
+
+    {
+        ggml_metal_kargs_cumsum_blk args = {
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.ne03 =*/ ne03,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.net0 =*/ net0,
+            /*.net1 =*/ net1,
+            /*.net2 =*/ net2,
+            /*.net3 =*/ net3,
+            /*.nbt0 =*/ nbt0,
+            /*.nbt1 =*/ nbt1,
+            /*.nbt2 =*/ nbt2,
+            /*.nbt3 =*/ nbt3,
+            /*.outb =*/ ne00 > nth,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_blk);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
+
+        ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, net0*ne01, ne02, ne03, nth, 1, 1);
+    }
+
+    if (ne00 > nth) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            ggml_metal_kargs_cumsum_blk args = {
+                /*.ne00 =*/ net0,
+                /*.ne01 =*/ net1,
+                /*.ne02 =*/ net2,
+                /*.ne03 =*/ net3,
+                /*.nb00 =*/ nbt0,
+                /*.nb01 =*/ nbt1,
+                /*.nb02 =*/ nbt2,
+                /*.nb03 =*/ nbt3,
+                /*.net0 =*/ net0,
+                /*.net1 =*/ net1,
+                /*.net2 =*/ net2,
+                /*.net3 =*/ net3,
+                /*.nbt0 =*/ nbt0,
+                /*.nbt1 =*/ nbt1,
+                /*.nbt2 =*/ nbt2,
+                /*.nbt3 =*/ nbt3,
+                /*.outb =*/ false,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline_blk);
+            ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 3);
+
+            ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, net1, net2, net3, nth, 1, 1);
+        }
+
+        ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            ggml_metal_pipeline_t pipeline_add = ggml_metal_library_get_pipeline_cumsum_add(lib, op);
+
+            ggml_metal_kargs_cumsum_add args = {
+                /*.ne00 =*/ ne00,
+                /*.ne01 =*/ ne01,
+                /*.ne02 =*/ ne02,
+                /*.ne03 =*/ ne03,
+                /*.nb00 =*/ nb00,
+                /*.nb01 =*/ nb01,
+                /*.nb02 =*/ nb02,
+                /*.nb03 =*/ nb03,
+                /*.net0 =*/ net0,
+                /*.net1 =*/ net1,
+                /*.net2 =*/ net2,
+                /*.net3 =*/ net3,
+                /*.nbt0 =*/ nbt0,
+                /*.nbt1 =*/ nbt1,
+                /*.nbt2 =*/ nbt2,
+                /*.nbt3 =*/ nbt3,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline_add);
+            ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_dst, 2);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, net0*ne01, ne02, ne03, nth, 1, 1);
+        }
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -967,7 +1100,7 @@ int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_get_rows(lib, op->src[0]->type);
 
@@ -1012,7 +1145,7 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);
 
@@ -1036,11 +1169,6 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
 
     nth = std::min(nth, nk0);
 
-    if (nth*nrptg > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
-        nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
-        nrptg = 1;
-    }
-
     ggml_metal_kargs_set_rows args = {
         /*.nk0  =*/ nk0,
         /*.ne01 =*/ ne01,
@@ -1081,7 +1209,7 @@ int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float scale;
     float max_bias;
@@ -1169,7 +1297,7 @@ int ggml_metal_op_ssm_conv(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_ssm_conv args = {
         /*.ne00 =*/ ne00,
@@ -1224,7 +1352,7 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne6, op->src[6], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb6, op->src[6], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const ggml_tensor * src3 = op->src[3];
     const ggml_tensor * src4 = op->src[4];
@@ -1310,7 +1438,7 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int64_t B = op->op == GGML_OP_RWKV_WKV6 ? op->src[5]->ne[1] : op->src[6]->ne[1];
     const int64_t T = op->src[0]->ne[2];
@@ -1351,7 +1479,7 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
 
@@ -1424,7 +1552,7 @@ int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t * opts = op->op_params;
     ggml_op_pool op_pool = (ggml_op_pool) opts[0];
@@ -1488,7 +1616,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     GGML_ASSERT(ne00 == ne10);
 
@@ -1729,7 +1857,7 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     // src2 = ids
     GGML_ASSERT(op->src[2]->type == GGML_TYPE_I32);
@@ -1975,7 +2103,9 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
     const bool has_mask = op->src[3] != nullptr;
 
     if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
-        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+        // note: always reserve the padding space to avoid graph reallocations
+        //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+        const bool has_kvpad = true;
 
         if (has_kvpad) {
             res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
@@ -1984,7 +2114,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
                 (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
         }
     } else {
-        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+        //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+        const bool has_kvpad = true;
 
         if (has_kvpad) {
             res += OP_FLASH_ATTN_EXT_NCPSG*(
@@ -2020,9 +2151,10 @@ size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
     const bool is_vec = ggml_metal_op_flash_attn_ext_use_vec(op);
 
     // this optimization is not useful for the vector kernels
-    if (is_vec) {
-        return res;
-    }
+    // note: always reserve the blk buffer to avoid graph reallocations
+    //if (is_vec) {
+    //    return res;
+    //}
 
     const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
     const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
@@ -2049,13 +2181,16 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
 
     size_t res = 0;
 
-    if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    // note: always reserve the temp buffer to avoid graph reallocations
+    //if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    if (true) {
         const int64_t nwg = 32;
+        const int64_t ne01_max = std::min(ne01, 32);
 
         // temp buffer for writing the results from each workgroup
         // - ne20: the size of the Value head
         // -  + 2: the S and M values for each intermediate result
-        res += ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+        res += ggml_type_size(GGML_TYPE_F32)*(ne01_max*ne02*ne03*nwg*(ne20 + 2));
     }
 
     return res;
@@ -2184,8 +2319,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
 
             need_sync = true;
-        } else {
-            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
         }
 
         if (has_mask) {
@@ -2215,8 +2348,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_dispatch_threadgroups(enc, nblk0, nblk1, ne32*ne33, 32, 1, 1);
 
             need_sync = true;
-        } else {
-            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) == 0);
         }
 
         if (need_sync) {
@@ -2356,8 +2487,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
 
             need_sync = true;
-        } else {
-            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
         }
 
         if (need_sync) {
@@ -2688,7 +2817,7 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
@@ -2736,7 +2865,7 @@ int ggml_metal_op_group_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t ngrp = ((const int32_t *) op->op_params)[0];
 
@@ -2791,7 +2920,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
@@ -2927,7 +3056,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     // make sure we have one or more position id(ne10) per token(ne02)
     GGML_ASSERT(ne10 % ne02 == 0);
@@ -3021,7 +3150,7 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
     const int32_t s1 = ((const int32_t *)(op->op_params))[1];
@@ -3082,6 +3211,84 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_conv_2d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t *) op->op_params)[0];
+    const int32_t s1 = ((const int32_t *) op->op_params)[1];
+    const int32_t p0 = ((const int32_t *) op->op_params)[2];
+    const int32_t p1 = ((const int32_t *) op->op_params)[3];
+    const int32_t d0 = ((const int32_t *) op->op_params)[4];
+    const int32_t d1 = ((const int32_t *) op->op_params)[5];
+
+    ggml_metal_kargs_conv_2d args = {
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+        /*.IW   =*/ ne10,
+        /*.IH   =*/ ne11,
+        /*.KW   =*/ ne00,
+        /*.KH   =*/ ne01,
+        /*.IC   =*/ ne02,
+        /*.OC   =*/ ne03,
+        /*.OW   =*/ ne0,
+        /*.OH   =*/ ne1,
+        /*.N    =*/ ne3,
+        /*.s0   =*/ s0,
+        /*.s1   =*/ s1,
+        /*.p0   =*/ p0,
+        /*.p1   =*/ p1,
+        /*.d0   =*/ d0,
+        /*.d1   =*/ d1,
+    };
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_conv_2d(lib, op);
+
+    int nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+    nth = std::min(nth, 256);
+    nth = std::max(nth, 1);
+
+    const uint64_t n_out = ggml_nelements(op);
+
+    uint64_t tg = (n_out + nth - 1)/nth;
+    tg = std::max<uint64_t>(tg, 1);
+    tg = std::min<uint64_t>(tg, (uint64_t) std::numeric_limits<int>::max());
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, tg, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -3093,7 +3300,7 @@ int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
 
@@ -3138,7 +3345,7 @@ int ggml_metal_op_conv_transpose_2d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
 
@@ -3192,7 +3399,7 @@ int ggml_metal_op_upscale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const float sf0 = (float)ne0/op->src[0]->ne[0];
     const float sf1 = (float)ne1/op->src[0]->ne[1];
@@ -3245,7 +3452,7 @@ int ggml_metal_op_pad(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_pad args = {
         /*.ne00 =*/ ne00,
@@ -3289,7 +3496,7 @@ int ggml_metal_op_pad_reflect_1d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_pad_reflect_1d args = {
         /*.ne00 =*/ ne00,
@@ -3333,7 +3540,7 @@ int ggml_metal_op_arange(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_t enc = ctx->enc;
 
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float start;
     float step;
@@ -3351,12 +3558,6 @@ int ggml_metal_op_arange(ggml_metal_op_t ctx, int idx) {
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_arange(lib, op);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:0];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:1];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op), 1);
@@ -3375,7 +3576,7 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int dim        = op->op_params[0];
     const int max_period = op->op_params[1];
@@ -3409,7 +3610,7 @@ int ggml_metal_op_argmax(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_argmax args = {
         /*.ne00 = */ ne00,
@@ -3445,38 +3646,93 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
-
-    // bitonic sort requires the number of elements to be power of 2
-    int64_t ne00_padded = 1;
-    while (ne00_padded < ne00) {
-        ne00_padded *= 2;
-    }
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_argsort(lib, op);
 
-    const int64_t nrows = ggml_nrows(op->src[0]);
+    // bitonic sort requires the number of elements to be power of 2
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    const int npr = (ne00 + nth - 1)/nth;
 
     // Metal kernels require the buffer size to be multiple of 16 bytes
     // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
-    const size_t smem = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+    const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += ggml_nbytes(op);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
 
     ggml_metal_kargs_argsort args = {
-        /*.ncols =*/ ne00,
-        /*.ncols_pad =*/ ne00_padded
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
     };
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, 1, nrows, 1, ne00_padded, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_argsort_merge(lib, op);
+
+    int len = nth;
+
+    while (len < ne00) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            .ne00 = ne00,
+            .ne01 = ne01,
+            .ne02 = ne02,
+            .ne03 = ne03,
+            .nb00 = nb00,
+            .nb01 = nb01,
+            .nb02 = nb02,
+            .nb03 = nb03,
+            .len  = len,
+        };
+
+        // merges per row
+        const int nm = (ne00 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge));
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
+        ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
+
+        std::swap(bid_dst, bid_tmp);
+
+        len <<= 1;
+    }
 
     return 1;
 }
@@ -3490,7 +3746,7 @@ int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float slope;
     memcpy(&slope, op->op_params, sizeof(float));
@@ -3526,7 +3782,7 @@ int ggml_metal_op_opt_step_adamw(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_opt_step_adamw(lib, op);
 
@@ -3562,7 +3818,7 @@ int ggml_metal_op_opt_step_sgd(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_opt_step_sgd(lib, op);
 

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

@@ -52,6 +52,7 @@ int ggml_metal_op_unary             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_glu               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_sum               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_sum_rows          (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
@@ -70,6 +71,7 @@ int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rope              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_im2col            (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_conv_2d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_1d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_2d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_upscale           (ggml_metal_op_t ctx, int idx);

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

@@ -197,6 +197,11 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
                 res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
                 res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
             } break;
+        case GGML_OP_CUMSUM:
+        case GGML_OP_ARGSORT:
+            {
+                res *= 2;
+            } break;
         default:
             break;
     }

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

@@ -1832,6 +1832,117 @@ typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
 template [[host_name("kernel_sum_rows_f32")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
 template [[host_name("kernel_mean_f32")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
 
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;
+
 template<typename T>
 kernel void kernel_soft_max(
         constant ggml_metal_kargs_soft_max & args,
@@ -4146,6 +4257,120 @@ template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 //template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
 //template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
 
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
 typedef void (conv_transpose_1d_t)(
         constant ggml_metal_kargs_conv_transpose_1d & args,
         device const float * src0,
@@ -4427,69 +4652,227 @@ kernel void kernel_timestep_embedding_f32(
 // bitonic sort implementation following the CUDA kernels as reference
 typedef void (argsort_t)(
         constant   ggml_metal_kargs_argsort & args,
-        device  const float * x,
+        device   const char * src0,
         device      int32_t * dst,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
 
 template<ggml_sort_order order>
 kernel void kernel_argsort_f32_i32(
         constant   ggml_metal_kargs_argsort & args,
-        device const float  * x,
+        device   const char * src0,
         device      int32_t * dst,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
-    int col = tpitg[0];
-    int row = tgpig[1];
+    const int col = tpitg[0];
 
-    if (col >= args.ncols_pad) return;
+    const int i00 = (tgpig[0]/args.ne01)*ntg.x;
+    const int i01 =  tgpig[0]%args.ne01;
+    const int i02 =  tgpig[1];
+    const int i03 =  tgpig[2];
 
-    device const float   * x_row   = x + row * args.ncols;
-    threadgroup int32_t  * dst_row = shared_values;
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
     // initialize indices
-    dst_row[col] = col;
+    shmem_i32[col] = i00 + col;
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (int k = 2; k <= args.ncols_pad; k *= 2) {
+    for (int k = 2; k <= ntg.x; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (dst_row[col] >= args.ncols ||
-                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 } else {
-                    if (dst_row[ixj] >= args.ncols ||
-                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 }
             }
+
             threadgroup_barrier(mem_flags::mem_threadgroup);
         }
     }
 
     // copy the result to dst without the padding
-    if (col < args.ncols) {
-        dst[row * args.ncols + col] = dst_row[col];
+    if (i00 + col < args.ne00) {
+        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
+
+        dst[col] = shmem_i32[col];
     }
 }
 
 template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
 template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
 
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne00 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne00 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne00
+        + i02*args.ne00*args.ne01
+        + i03*args.ne00*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.ne00
+        + i02*args.ne00*args.ne01
+        + i03*args.ne00*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = min(k0 + chunk, total);
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;
+
 kernel void kernel_leaky_relu_f32(
         constant     ggml_metal_kargs_leaky_relu & args,
         device const float * src0,
@@ -6177,6 +6560,7 @@ template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_
 template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
 template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
 template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
 #endif

ggml/src/ggml-opencl/CMakeLists.txt

@@ -119,6 +119,7 @@ set(GGML_OPENCL_KERNELS
     pad
     repeat
     mul_mat_f16_f32
+    mul_mm_f16_f32_kq_kqv
     conv2d
     conv2d_f16_f32
     flash_attn_f32_f16

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

@@ -407,6 +407,8 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_f32_f32;
     cl_program program_mul;
     cl_program program_mul_mat_f16_f32_tiled;
+    cl_program program_mul_mm_f16_f32_kqv;
+    cl_program program_mul_mm_f16_f32_kq;
     cl_program program_div;
     cl_program program_sub;
     cl_program program_norm;
@@ -481,6 +483,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_f16_f32;
     cl_kernel kernel_mul_mat_f16_f32_l4;
     cl_kernel kernel_mul_mat_f16_f32_tiled;
+    cl_kernel kernel_mul_mm_f16_f32_kqv;
+    cl_kernel kernel_mul_mm_f16_f32_kq;
     cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
     cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
     cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4, kernel_restore_block_mxfp4_trans;
@@ -1235,6 +1239,25 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mm_f16_f32_kq_kqv
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f16_f32_kq_kqv.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f16_f32_kq_kqv.cl");
+#endif
+        backend_ctx->program_mul_mm_f16_f32_kqv =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts+" -DKQV ");
+        backend_ctx->program_mul_mm_f16_f32_kq =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_kqv = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_kqv, "mul_mm_f16_f32_kqv", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_kq = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_kq, "mul_mm_f16_f32_kq", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -5682,7 +5705,7 @@ static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor *
     CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong),      &nb2));
     CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),      &nb3));
     CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float),         &eps));
-    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*nth/sgs, NULL));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,     NULL));
 
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
@@ -6665,6 +6688,146 @@ static void ggml_cl_conv_2d(ggml_backend_t backend, const ggml_tensor * src0, co
     backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
 }
 
+static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+    const int  ne02 = src0->ne[2];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const int  ne10 = src1->ne[0];
+    const int  ne11 = src1->ne[1];
+    const int  ne12 = src1->ne[2];
+
+    const cl_ulong nb10 = src1->nb[0];
+
+    const int  ne0 = dst->ne[0];
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 == ne10);
+
+    cl_kernel kernel;
+    cl_context context = backend_ctx->context;
+
+    cl_int              status;
+    cl_image_format     img_fmt_1d;
+    cl_image_desc       img_desc_1d;
+    cl_buffer_region    region;
+    cl_mem              A_image1d;
+    cl_mem              A_sub_buffer;
+    cl_mem              B_sub_buffer;
+    cl_mem              D_image1d;
+    cl_mem              D_sub_buffer;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    if (nb01 > nb02) {
+        // KQ
+        kernel = backend_ctx->kernel_mul_mm_f16_f32_kq;
+    } else {
+        // KQV
+        kernel = backend_ctx->kernel_mul_mm_f16_f32_kqv;
+    }
+    // create sub-buffer for A
+    // <--------------------------------------------> //
+    extra0 = src0->view_src ? (ggml_tensor_extra_cl *)src0->view_src->extra : (ggml_tensor_extra_cl *)src0->extra;
+
+    region.origin = (extra0->offset);
+    if (nb01 > nb02) {
+        // KQ
+        region.size = nb01 * ne01;
+    } else {
+        // KQV
+        region.size = nb02 * ne02;
+    }
+
+    A_sub_buffer = clCreateSubBuffer((extra0->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+
+    // <--------------------------------------------> //
+
+    // create sub-buffer for B
+    // <--------------------------------------------> //
+    region.origin = (extra1->offset);
+    region.size = nb10 * ne10 * ne11 * ne12;
+    B_sub_buffer = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+    // <--------------------------------------------> //
+
+    img_fmt_1d = {CL_RGBA, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    if (nb01 > nb02) {
+        img_desc_1d.image_width = (nb01 * ne01 / 4)/4;
+    }
+    else {
+        img_desc_1d.image_width = (nb02 * ne02 / 4)/4;
+    }
+    img_desc_1d.buffer = A_sub_buffer;
+    A_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // create sub-buffer for output C
+    // <--------------------------------------------> //
+    region.origin = (extrad->offset);
+    region.size = ne0 * ne1 * dst->ne[2] * dst->nb[0]; // size of C in bytes
+    D_sub_buffer = clCreateSubBuffer((extrad->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+    // <--------------------------------------------> //
+
+    // create image for C output
+    // <--------------------------------------------> //
+    img_fmt_1d = {CL_R, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = ne0 * ne1 * dst->ne[2] * dst->nb[0] / 4;
+    img_desc_1d.buffer = D_sub_buffer;
+    D_image1d = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+    // <--------------------------------------------> //
+
+    int offset_src0 = 0;
+    int offset_src1 = 0;
+
+    // set kernel args
+    // <--------------------------------------------> //
+    cl_uint k_arg = 0;
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem), &A_image1d));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &offset_src0));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem), &B_sub_buffer));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &offset_src1));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem), &D_image1d));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &extrad->offset));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &M));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &K));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &N));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &ne12));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &nb01));
+
+    size_t global_work_size[3] = {64, static_cast<size_t>(((M+63)/64)), static_cast<size_t>(((N+31)/32)*ne12)};
+    size_t local_work_size[3] = {64, 1, 2};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+    // deallocate sub buffers and images
+    // <--------------------------------------------> //
+    CL_CHECK(clReleaseMemObject(A_image1d));
+    CL_CHECK(clReleaseMemObject(D_image1d));
+    CL_CHECK(clReleaseMemObject(A_sub_buffer));
+    CL_CHECK(clReleaseMemObject(B_sub_buffer));
+    CL_CHECK(clReleaseMemObject(D_sub_buffer));
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -6731,6 +6894,27 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     cl_context context = backend_ctx->context;
 
+    if(src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32){
+        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0) {
+            // For KQ
+            if (ggml_is_permuted(src0) && ggml_is_permuted(src1) &&
+                nb00 <= nb02 &&
+                nb02 <= nb01 &&
+                nb01 <= nb03 &&
+                nb10 <= nb12 &&
+                nb12 <= nb11 &&
+                nb11 <= nb13) {
+                ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
+                return;
+            }
+            // For KQV
+            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+                ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
+                return;
+            }
+        }
+    }
+
     if (ne01 && ne1 && use_adreno_kernels(backend_ctx, src0)) {
 
     // init CL objects

ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl

@@ -0,0 +1,273 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#define LM_FIRST_256B   0
+#define LM_SECOND_256B  64
+#define LM_THIRD_256B   128
+#define LM_FOURTH_256B  192
+
+
+inline float16 mm_load_a(
+    image1d_buffer_t matrix_A,
+    uint subMatrixAStartInElements,
+    int nb01,
+    int line_stride_matrix_A_in_bytes
+) {
+    __private float8 regA;
+    size_t sub_block_id_m = get_local_id(0);
+
+#ifdef KQV
+    uint a_texCoord = subMatrixAStartInElements/2 + (sub_block_id_m * nb01/4);
+#else // KQ
+    uint a_texCoord = subMatrixAStartInElements/2 + (sub_block_id_m * line_stride_matrix_A_in_bytes/4);
+#endif
+
+    regA.s0123  = read_imagef(matrix_A, a_texCoord/4);
+    regA.s4567  = read_imagef(matrix_A, (a_texCoord+4)/4);
+
+    return convert_float16(as_half16(regA));
+}
+
+inline float4 alu_32(
+    float16 regA,
+    __local float4* matrix_B_vec
+) {
+
+    __private float4 rC = 0;
+    int i = get_sub_group_id() * 64;
+
+    rC += regA.s0  * matrix_B_vec[i];
+    rC += regA.s1  * matrix_B_vec[i + 16];
+    rC += regA.s4  * matrix_B_vec[i + 1];
+    rC += regA.s5  * matrix_B_vec[i + 17];
+    rC += regA.s8  * matrix_B_vec[i + 2];
+    rC += regA.s9  * matrix_B_vec[i + 18];
+    rC += regA.sc  * matrix_B_vec[i + 3];
+    rC += regA.sd  * matrix_B_vec[i + 19];
+
+    i += 32;
+
+    rC += regA.s2  * matrix_B_vec[i];
+     rC += regA.s3  * matrix_B_vec[i + 16];
+    rC += regA.s6  * matrix_B_vec[i + 1];
+    rC += regA.s7  * matrix_B_vec[i + 17];
+    rC += regA.sa  * matrix_B_vec[i + 2];
+    rC += regA.sb  * matrix_B_vec[i + 18];
+    rC += regA.se  * matrix_B_vec[i + 3];
+    rC += regA.sf  * matrix_B_vec[i + 19];
+
+    return rC;
+}
+
+inline float16 alu_16(
+    float16 regA,
+    __local float* matrix_B_local
+) {
+    float16 out;
+    __local float4* matrix_B_vec = (__local float4*)matrix_B_local;
+
+    out.s0123 = alu_32(regA, matrix_B_vec);
+    out.s4567 = alu_32(regA, matrix_B_vec + 4);
+    out.s89ab = alu_32(regA, matrix_B_vec + 8);
+    out.scdef = alu_32(regA, matrix_B_vec + 12);
+
+    return out;
+}
+
+inline void mm_mad(
+    __local float* matrix_B_local,
+    float16 regA,
+    float8 regB,
+    uint b_localOffsetInWords,
+    float16* regC0_ptr,
+    float16* regC1_ptr
+) {
+    int offset = b_localOffsetInWords + get_sub_group_id() * 256;
+
+    matrix_B_local[offset + LM_FIRST_256B] = regB.s0;
+    matrix_B_local[offset + LM_SECOND_256B] = regB.s1;
+    matrix_B_local[offset + LM_THIRD_256B] = regB.s2;
+    matrix_B_local[offset + LM_FOURTH_256B] = regB.s3;
+
+    float16 add0 = alu_16(regA, matrix_B_local);
+    *regC0_ptr += add0;
+
+    matrix_B_local[offset + LM_FIRST_256B] = regB.s4;
+    matrix_B_local[offset + LM_SECOND_256B] = regB.s5;
+    matrix_B_local[offset + LM_THIRD_256B] = regB.s6;
+    matrix_B_local[offset + LM_FOURTH_256B] = regB.s7;
+
+    float16 add1 = alu_16(regA, matrix_B_local);
+    *regC1_ptr += add1;
+}
+
+inline void mm_store_c_N(
+    __write_only image1d_buffer_t matrix_C,
+    float16 regC0,
+    float16 regC1,
+    uint subMatrixCStartInElements,
+    int line_stride_matrix_C_in_bytes,
+    int mask
+) {
+    size_t sub_block_id_m = get_local_id(0);
+
+    uint strideInWords     = line_stride_matrix_C_in_bytes/4;
+    uint c_coordInWords_0  = (subMatrixCStartInElements + sub_block_id_m);
+
+    uint c_coordInWords_1  = c_coordInWords_0 + 1  * strideInWords;
+    uint c_coordInWords_2  = c_coordInWords_0 + 2  * strideInWords;
+    uint c_coordInWords_3  = c_coordInWords_0 + 3  * strideInWords;
+    uint c_coordInWords_4  = c_coordInWords_0 + 4  * strideInWords;
+    uint c_coordInWords_5  = c_coordInWords_0 + 5  * strideInWords;
+    uint c_coordInWords_6  = c_coordInWords_0 + 6  * strideInWords;
+    uint c_coordInWords_7  = c_coordInWords_0 + 7  * strideInWords;
+    uint c_coordInWords_8  = c_coordInWords_0 + 8  * strideInWords;
+    uint c_coordInWords_9  = c_coordInWords_0 + 9  * strideInWords;
+    uint c_coordInWords_10 = c_coordInWords_0 + 10 * strideInWords;
+    uint c_coordInWords_11 = c_coordInWords_0 + 11 * strideInWords;
+    uint c_coordInWords_12 = c_coordInWords_0 + 12 * strideInWords;
+    uint c_coordInWords_13 = c_coordInWords_0 + 13 * strideInWords;
+    uint c_coordInWords_14 = c_coordInWords_0 + 14 * strideInWords;
+    uint c_coordInWords_15 = c_coordInWords_0 + 15 * strideInWords;
+    uint c_coordInWords_16 = c_coordInWords_0 + 16 * strideInWords;
+    uint c_coordInWords_17 = c_coordInWords_0 + 17 * strideInWords;
+    uint c_coordInWords_18 = c_coordInWords_0 + 18 * strideInWords;
+    uint c_coordInWords_19 = c_coordInWords_0 + 19 * strideInWords;
+    uint c_coordInWords_20 = c_coordInWords_0 + 20 * strideInWords;
+    uint c_coordInWords_21 = c_coordInWords_0 + 21 * strideInWords;
+    uint c_coordInWords_22 = c_coordInWords_0 + 22 * strideInWords;
+    uint c_coordInWords_23 = c_coordInWords_0 + 23 * strideInWords;
+    uint c_coordInWords_24 = c_coordInWords_0 + 24 * strideInWords;
+    uint c_coordInWords_25 = c_coordInWords_0 + 25 * strideInWords;
+    uint c_coordInWords_26 = c_coordInWords_0 + 26 * strideInWords;
+    uint c_coordInWords_27 = c_coordInWords_0 + 27 * strideInWords;
+    uint c_coordInWords_28 = c_coordInWords_0 + 28 * strideInWords;
+    uint c_coordInWords_29 = c_coordInWords_0 + 29 * strideInWords;
+    uint c_coordInWords_30 = c_coordInWords_0 + 30 * strideInWords;
+    uint c_coordInWords_31 = c_coordInWords_0 + 31 * strideInWords;
+
+    if (mask > 0)  { write_imagef(matrix_C, c_coordInWords_0, regC0.s0);  }
+    if (mask > 1)  { write_imagef(matrix_C, c_coordInWords_1, regC0.s1);  }
+    if (mask > 2)  { write_imagef(matrix_C, c_coordInWords_2, regC0.s2);  }
+    if (mask > 3)  { write_imagef(matrix_C, c_coordInWords_3, regC0.s3);  }
+    if (mask > 4)  { write_imagef(matrix_C, c_coordInWords_4, regC0.s4);  }
+    if (mask > 5)  { write_imagef(matrix_C, c_coordInWords_5, regC0.s5);  }
+    if (mask > 6)  { write_imagef(matrix_C, c_coordInWords_6, regC0.s6);  }
+    if (mask > 7)  { write_imagef(matrix_C, c_coordInWords_7, regC0.s7);  }
+    if (mask > 8)  { write_imagef(matrix_C, c_coordInWords_8, regC0.s8);  }
+    if (mask > 9)  { write_imagef(matrix_C, c_coordInWords_9, regC0.s9);  }
+    if (mask > 10) { write_imagef(matrix_C, c_coordInWords_10, regC0.sa); }
+    if (mask > 11) { write_imagef(matrix_C, c_coordInWords_11, regC0.sb); }
+    if (mask > 12) { write_imagef(matrix_C, c_coordInWords_12, regC0.sc); }
+    if (mask > 13) { write_imagef(matrix_C, c_coordInWords_13, regC0.sd); }
+    if (mask > 14) { write_imagef(matrix_C, c_coordInWords_14, regC0.se); }
+    if (mask > 15) { write_imagef(matrix_C, c_coordInWords_15, regC0.sf); }
+    if (mask > 16) { write_imagef(matrix_C, c_coordInWords_16, regC1.s0); }
+    if (mask > 17) { write_imagef(matrix_C, c_coordInWords_17, regC1.s1); }
+    if (mask > 18) { write_imagef(matrix_C, c_coordInWords_18, regC1.s2); }
+    if (mask > 19) { write_imagef(matrix_C, c_coordInWords_19, regC1.s3); }
+    if (mask > 20) { write_imagef(matrix_C, c_coordInWords_20, regC1.s4); }
+    if (mask > 21) { write_imagef(matrix_C, c_coordInWords_21, regC1.s5); }
+    if (mask > 22) { write_imagef(matrix_C, c_coordInWords_22, regC1.s6); }
+    if (mask > 23) { write_imagef(matrix_C, c_coordInWords_23, regC1.s7); }
+    if (mask > 24) { write_imagef(matrix_C, c_coordInWords_24, regC1.s8); }
+    if (mask > 25) { write_imagef(matrix_C, c_coordInWords_25, regC1.s9); }
+    if (mask > 26) { write_imagef(matrix_C, c_coordInWords_26, regC1.sa); }
+    if (mask > 27) { write_imagef(matrix_C, c_coordInWords_27, regC1.sb); }
+    if (mask > 28) { write_imagef(matrix_C, c_coordInWords_28, regC1.sc); }
+    if (mask > 29) { write_imagef(matrix_C, c_coordInWords_29, regC1.sd); }
+    if (mask > 30) { write_imagef(matrix_C, c_coordInWords_30, regC1.se); }
+    if (mask > 31) { write_imagef(matrix_C, c_coordInWords_31, regC1.sf); }
+}
+
+#define TILESIZE_K 16
+#define TILESIZE_M 64
+#define TILESIZE_N 32
+#ifdef KQV
+__kernel void mul_mm_f16_f32_kqv(
+#else
+__kernel void mul_mm_f16_f32_kq(
+#endif
+        __read_only  image1d_buffer_t matrix_A,
+        int offset0,
+        __global float* matrix_B,
+        int offset1,
+        __write_only image1d_buffer_t matrix_C,
+        int offsetd,
+        int M, int K, int N,
+        int D_A,
+        int D_B,
+        int nb01
+) {
+
+    uint block_id_m = get_global_id(1);
+    uint block_id_n = get_global_id(2) % ((N+TILESIZE_N-1)/TILESIZE_N);
+    uint block_id_d = get_global_id(2) / ((N+TILESIZE_N-1)/TILESIZE_N);
+
+    __private float16  regA;
+    __private float8   regB;
+    __private float16 regC0;
+    __private float16 regC1;
+
+    const uint col   = block_id_m * TILESIZE_M;
+    const uint row   = block_id_n * TILESIZE_N;
+    const uint depth_A = block_id_d / (D_B/D_A);
+    const uint depth_B = block_id_d;
+
+#ifdef KQV
+    int line_stride_matrix_A_in_bytes = nb01 * M;
+    int line_stride_matrix_B_in_bytes = K * N * 4;
+#else
+    int line_stride_matrix_A_in_bytes = K * D_A * 2;
+    int line_stride_matrix_B_in_bytes = K * D_B * 4;
+#endif
+
+    int line_stride_matrix_C_in_bytes = M * 4;
+
+    const uint strideAinElements = line_stride_matrix_A_in_bytes / 2;
+    const uint strideBinElements = line_stride_matrix_B_in_bytes / 4;
+
+    size_t sub_block_id_m = get_local_id(0);
+
+    uint b_localOffsetInWords = (sub_block_id_m/16)*16
+                           + ((((sub_block_id_m)>>0)&1)<<2)
+                           + ((((sub_block_id_m)>>1)&1)<<3)
+                           + ((((sub_block_id_m)>>2)&1)<<0)
+                           + ((((sub_block_id_m)>>3)&1)<<1);
+
+    uint2 b_globalOffsetInWords_xy = {((sub_block_id_m%4)*4), (sub_block_id_m>>2)};
+    uint b_globalOffsetInWords00, b_globalOffsetInWords16;
+#ifdef KQV
+    b_globalOffsetInWords00 = b_globalOffsetInWords_xy.x + b_globalOffsetInWords_xy.y*K;
+    b_globalOffsetInWords16 = b_globalOffsetInWords00 + (16 * K);
+    uint subMatrixAStartInElements = depth_A * strideAinElements + col * nb01 / 2;
+    uint subMatrixBStartInElements = depth_B * strideBinElements + row * K;
+#else
+    b_globalOffsetInWords00 = b_globalOffsetInWords_xy.x + b_globalOffsetInWords_xy.y*line_stride_matrix_B_in_bytes/4;
+    b_globalOffsetInWords16 = b_globalOffsetInWords00 + (16 * line_stride_matrix_B_in_bytes/4);
+    uint subMatrixAStartInElements = col * strideAinElements + depth_A * K;
+    uint subMatrixBStartInElements = row * strideBinElements + depth_B * K;
+#endif
+
+    __local float matrix_B_local[1024];
+
+    for (uint step=0; step < K; step+=TILESIZE_K) {
+        size_t sub_block_id_m = get_local_id(0);
+        regA = mm_load_a(matrix_A, subMatrixAStartInElements, nb01, line_stride_matrix_A_in_bytes);
+
+        uint b_coordInWords00 = subMatrixBStartInElements + b_globalOffsetInWords00;
+        uint b_coordInWords16 = subMatrixBStartInElements + b_globalOffsetInWords16;
+
+        regB.s0123 = vload4(b_coordInWords00/4, matrix_B);
+        regB.s4567 = vload4(b_coordInWords16/4, matrix_B);
+
+        mm_mad(matrix_B_local, regA, regB, b_localOffsetInWords, &regC0, &regC1);
+
+        subMatrixAStartInElements += TILESIZE_K;
+        subMatrixBStartInElements += TILESIZE_K;
+    }
+
+    uint subMatrixCStartInElements = depth_B * N * M + row * M + col;
+    mm_store_c_N(matrix_C, regC0, regC1, subMatrixCStartInElements, line_stride_matrix_C_in_bytes, (N-block_id_n*32));
+}
+

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

@@ -134,6 +134,15 @@ kernel void kernel_rms_norm_mul(
     src1 = src1 + offset1;
     dst  = dst  + offsetd;
 
+    // The size of sum is sizeof(float)*subgroup_size.
+    // Each subgroup writes its partial sum to this array.
+    // So the number of subgroups per workgroup for this kernel cannot exceed the subgroup size.
+    // This is generally true -
+    // for subgroup size 64, workgroup size should be less than 4096 (the max is usually 1024).
+    if (get_sub_group_id() == 0) {
+        sum[get_sub_group_local_id()] = 0.0f;
+    }
+
     int i03 = get_group_id(2);
     int i02 = get_group_id(1);
     int i01 = get_group_id(0);
@@ -148,24 +157,30 @@ kernel void kernel_rms_norm_mul(
         sumf += dot(x[i00], x[i00]);
     }
     sumf = sub_group_reduce_add(sumf);
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
     if (get_sub_group_local_id() == 0) {
         sum[get_sub_group_id()] = sumf;
     }
 
     barrier(CLK_LOCAL_MEM_FENCE);
 
-    for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
-       if (get_local_id(0) < i) {
-           sum[get_local_id(0)] += sum[get_local_id(0) + i];
-       }
-    }
-    if (get_local_id(0) == 0) {
-        sum[0] /= ne00;
-    }
+    //for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
+    //   if (get_local_id(0) < i) {
+    //       sum[get_local_id(0)] += sum[get_local_id(0) + i];
+    //   }
+    //}
+    //if (get_local_id(0) == 0) {
+    //    sum[0] /= ne00;
+    //}
 
-    barrier(CLK_LOCAL_MEM_FENCE);
+    //barrier(CLK_LOCAL_MEM_FENCE);
 
-    float mean  = sum[0];
+    sumf = sum[get_sub_group_local_id()];
+    sumf = sub_group_reduce_add(sumf);
+
+    float mean  = sumf / ne00;
     float scale = 1.0f/sqrt(mean + eps);
 
     global float4 * y = (global float4 *) (dst + i03*nb3 + i02*nb2 + i01*nb1);

ggml/src/ggml-sycl/element_wise.cpp

@@ -170,73 +170,31 @@ static __dpct_inline__ T op_trunc(T x) {
     return sycl::trunc(x);
 }
 
-template<typename T>
-static void unary_op_sgn_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_sgn(x[i]);
-    }
-}
+template<typename T, typename F>
+static void unary_op_generic_kernel(
+        const T * x,
+        T * dst,
+        const int k,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+        const size_t nb0,  const size_t nb1,  const size_t nb2,  const size_t nb3,
+        const size_t nbd0, const size_t nbd1, const size_t nbd2, const size_t nbd3,
+        const sycl::nd_item<1> & item_ct1,
+        F func) {
 
-template<typename T>
-static void unary_op_abs_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+        (void) ne3;
     SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_abs(x[i]);
-    }
-}
+        const int64_t i0 =  i % ne0;
+        const int64_t i1 = (i / ne0)        % ne1;
+        const int64_t i2 = (i / (ne0*ne1))  % ne2;
+        const int64_t i3 =  i / (ne0*ne1*ne2);
 
-template<typename T>
-static void unary_op_elu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_elu(x[i]);
-    }
-}
+        const char * src_base = (const char *) x;
+        char       * dst_base = (char *) dst;
 
-template<typename T>
-static void unary_op_gelu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_gelu(x[i]);
-    }
-}
+        const T * srcp = (const T *)(src_base + i0*nb0  + i1*nb1  + i2*nb2  + i3*nb3 );
+        T *       dstp = (T *)(dst_base + i0*nbd0 + i1*nbd1 + i2*nbd2 + i3*nbd3);
 
-template<typename T>
-static void unary_op_silu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_silu(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_gelu_quick_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_gelu_quick(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_gelu_erf_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_gelu_erf(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_tanh_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_tanh(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_relu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_relu(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_sigmoid_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_sigmoid(x[i]);
+        *dstp = func(*srcp);
     }
 }
 
@@ -261,27 +219,6 @@ static void unary_op_cos_kernel(const T * x, T * dst, const int k, const sycl::n
     }
 }
 
-template<typename T>
-static void unary_op_hardsigmoid_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_hardsigmoid(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_hardswish_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_hardswish(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_exp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_exp(x[i]);
-    }
-}
-
 template<typename T>
 static void unary_op_log_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
     SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
@@ -289,19 +226,6 @@ static void unary_op_log_kernel(const T * x, T * dst, const int k, const sycl::n
     }
 }
 
-template<typename T>
-static void unary_op_neg_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_neg(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_step_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
-    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
-        dst[i] = op_step(x[i]);
-    }
-}
 
 template<typename T>
 static void unary_op_leaky_relu_kernel(const T * x, T * dst, const int k, float negative_slope, const sycl::nd_item<1> &item_ct1) {
@@ -620,6 +544,48 @@ static inline void dispatch_ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx
     }
 }
 
+template<typename F>
+static inline void ggml_sycl_op_unary(
+        ggml_backend_sycl_context & ctx, ggml_tensor * dst, F func) {
+
+    ggml_tensor * src0 = dst->src[0];
+
+    const int64_t ne0  = dst->ne[0];
+    const int64_t ne1  = dst->ne[1];
+    const int64_t ne2  = dst->ne[2];
+    const int64_t ne3  = dst->ne[3];
+
+    const size_t  nb0  = src0->nb[0];
+    const size_t  nb1  = src0->nb[1];
+    const size_t  nb2  = src0->nb[2];
+    const size_t  nb3  = src0->nb[3];
+
+    const size_t  nbd0 = dst->nb[0];
+    const size_t  nbd1 = dst->nb[1];
+    const size_t  nbd2 = dst->nb[2];
+    const size_t  nbd3 = dst->nb[3];
+
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [=](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+
+            const int num_blocks = ceil_div(k_elements, 256);
+
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_generic_kernel(
+                        src, dst_ptr, k_elements,
+                        ne0, ne1, ne2, ne3,
+                        nb0, nb1, nb2, nb3,
+                        nbd0, nbd1, nbd2, nbd3,
+                        item_ct1,
+                        func
+                    );
+                });
+        });
+}
+
 
 static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
@@ -645,159 +611,75 @@ static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_ten
 
 
 static inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, 256);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
-                                  sycl::range<1>(256)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_sgn_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_sgn(x);
+    });
 }
 
+
 static inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, 256);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
-                                  sycl::range<1>(256)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_abs_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_abs(x);
+    });
 }
 
 static inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, 256);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
-                                  sycl::range<1>(256)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_elu_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_elu(x);
+    });
 }
-
 static inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_SILU_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SILU_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_silu_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_silu(x);
+    });
 }
 
 static inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_gelu_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_gelu(x);
+    });
 }
 
-static inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_gelu_quick_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+static inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_gelu_quick(x);
+    });
 }
 
-static inline void ggml_sycl_op_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_gelu_erf_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+static inline void ggml_sycl_op_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_gelu_erf(x);
+    });
 }
 
 static inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_TANH_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_TANH_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_TANH_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_tanh_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_tanh(x);
+    });
 }
 
 static inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_relu_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_relu(x);
+    });
 }
 
 static inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_HARDSIGMOID_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_hardsigmoid_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_hardsigmoid(x);
+    });
 }
 
 static inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_HARDSWISH_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_hardswish_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_hardswish(x);
+    });
 }
 
 static inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_exp_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_exp(x);
+    });
 }
 
 static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -814,42 +696,22 @@ static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor
 }
 
 static inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_neg_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_neg(x);
+    });
 }
 
+
 static inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE); // Using NEG block size
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_step_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_step(x);
+    });
 }
 
 static inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
-            const int num_blocks = ceil_div(k_elements, SYCL_SIGMOID_BLOCK_SIZE);
-            stream->parallel_for(
-                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE),
-                                  sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
-                    unary_op_sigmoid_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_sigmoid(x);
+    });
 }
 
 static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4360,21 +4360,22 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             }
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_ABS:
                 case GGML_UNARY_OP_NEG:
                 case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
-                case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_SIGMOID:
-                case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_GELU_ERF:
-                case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
-                case GGML_UNARY_OP_SGN:
-                case GGML_UNARY_OP_ABS:
                 case GGML_UNARY_OP_ELU:
+                    return true;
                 case GGML_UNARY_OP_FLOOR:
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND: