ggml-cpu: aarm64: q4_K repack gemm and gemv implementations (dotprod only) (#17494)

* Enabled q4_K_4x8 path

* Fixed generic Q4_K 8x4 implementation

* wip: dotprod gemm

* Working arm q4_K dotprod gemm

Signed-off-by: Alberto Cabrera <alberto.cabrera@liquid.ai>

* Undo acc rename

Signed-off-by: Alberto Cabrera <alberto.cabrera@liquid.ai>

* Q4_K arm dotprod gemm

Signed-off-by: Alberto Cabrera <alberto.cabrera@liquid.ai>

* Fix: q4_qs reinterpret from uint to int

Signed-off-by: Alberto Cabrera <alberto.cabrera@liquid.ai>

* Removed comments

* Fixed macro guards

* Fixed unused vars in generic implementation

* Fixed unused vars in 8x4 repack

* Fixed unused vars in generic implementation, unneeded comment

* Missing arch fallback for x86

* minor : style

---------

Signed-off-by: Alberto Cabrera <alberto.cabrera@liquid.ai>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.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
 
@@ -52,6 +50,7 @@ WORKDIR /app
 
 RUN apt-get update \
     && apt-get install -y \
+    build-essential \
     git \
     python3 \
     python3-pip \

.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/ccache-action@v1.2.16
@@ -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/ccache-action@v1.2.16
@@ -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/ccache-action@v1.2.16
@@ -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:

.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 }}

CODEOWNERS

@@ -2,10 +2,8 @@
 # multiplie collaborators per item can be specified
 
 /.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @slaren @CISC
+/.github/actions/                       @CISC
 /.github/workflows/                     @CISC
-/.github/workflows/release.yml          @slaren
-/.github/workflows/winget.yml           @slaren
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
 /common/CMakeLists.txt                  @ggerganov
@@ -40,21 +38,14 @@
 /examples/passkey/                      @ggerganov
 /examples/retrieval/                    @ggerganov
 /examples/save-load-state/              @ggerganov
-/examples/simple-chat/                  @slaren
-/examples/simple/                       @slaren
 /examples/speculative-simple/           @ggerganov
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov
-/ggml/include/                          @ggerganov @slaren
-/ggml/src/ggml-alloc.c                  @slaren
-/ggml/src/ggml-backend*                 @slaren
-/ggml/src/ggml-blas/                    @slaren
-/ggml/src/ggml-common.h                 @ggerganov @slaren
-/ggml/src/ggml-cpu/                     @ggerganov @slaren
+/ggml/include/                          @ggerganov
+/ggml/src/ggml-common.h                 @ggerganov
+/ggml/src/ggml-cpu/                     @ggerganov
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
-/ggml/src/ggml-cuda/common.cuh          @slaren
 /ggml/src/ggml-cuda/fattn*              @JohannesGaessler
-/ggml/src/ggml-cuda/ggml-cuda.cu        @slaren
 /ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
 /ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
 /ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
@@ -62,19 +53,19 @@
 /ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
 /ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
-/ggml/src/ggml-impl.h                   @ggerganov @slaren
+/ggml/src/ggml-impl.h                   @ggerganov
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
 /ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov
-/ggml/src/ggml-threading.*              @ggerganov @slaren
+/ggml/src/ggml-threading.*              @ggerganov
 /ggml/src/ggml-vulkan/                  @0cc4m
 /ggml/src/ggml-webgpu/                  @reeselevine
 /ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml.c                        @ggerganov @slaren
-/ggml/src/ggml.cpp                      @ggerganov @slaren
+/ggml/src/ggml.c                        @ggerganov
+/ggml/src/ggml.cpp                      @ggerganov
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
 /gguf-py/                               @CISC
 /media/                                 @ggerganov
@@ -86,15 +77,11 @@
 /src/llama-arch.*                       @CISC
 /src/llama-chat.*                       @ngxson
 /src/llama-graph.*                      @CISC
-/src/llama-model-loader.*               @slaren
 /src/llama-model.*                      @CISC
 /src/llama-vocab.*                      @CISC
 /src/models/                            @CISC
 /tests/                                 @ggerganov
-/tests/test-backend-ops.cpp             @slaren
-/tests/test-thread-safety.cpp           @slaren
 /tools/batched-bench/                   @ggerganov
-/tools/llama-bench/                     @slaren
 /tools/main/                            @ggerganov
 /tools/mtmd/                            @ngxson
 /tools/perplexity/                      @ggerganov
@@ -106,8 +93,6 @@
 /tools/tokenize/                        @ggerganov
 /tools/tts/                             @ggerganov
 /vendor/                                @ggerganov
-/.clang-format                          @slaren
-/.clang-tidy                            @slaren
 /AUTHORS                                @ggerganov
 /CMakeLists.txt                         @ggerganov
 /CONTRIBUTING.md                        @ggerganov

README.md

@@ -242,6 +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/arg.cpp

@@ -694,6 +694,12 @@ static bool is_autoy(const std::string & value) {
 }
 
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+    // default values specific to example
+    // note: we place it here instead of inside server.cpp to allow llama-gen-docs to pick it up
+    if (ex == LLAMA_EXAMPLE_SERVER) {
+        params.use_jinja = true;
+    }
+
     // load dynamic backends
     ggml_backend_load_all();
 
@@ -1232,6 +1238,7 @@ 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 +1268,7 @@ 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 +1276,7 @@ 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 +1284,7 @@ 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 +1292,7 @@ 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 +1307,7 @@ 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 +1315,7 @@ 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 +1334,7 @@ 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 +1342,7 @@ 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 +1440,7 @@ 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 +1448,7 @@ 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 +1456,7 @@ 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(
@@ -2476,11 +2494,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--jinja"},
-        "use jinja template for chat (default: disabled)",
+        string_format("use jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
         [](common_params & params) {
             params.use_jinja = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_JINJA"));
+    add_opt(common_arg(
+        {"--no-jinja"},
+        string_format("disable jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.use_jinja = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_NO_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"

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
 //
@@ -942,6 +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);
@@ -953,6 +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/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

@@ -565,7 +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
 
@@ -4183,6 +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):
@@ -10046,6 +10061,25 @@ class LazyTorchTensor(gguf.LazyBase):
         torch.uint8: np.uint8,
     }
 
+    # only used when byteswapping data. Only correct size is needed
+    _dtype_byteswap_map: dict[torch.dtype, type] = {
+        torch.float64: np.float64,
+        torch.float32: np.float32,
+        torch.bfloat16: np.float16,
+        torch.float16: np.float16,
+        torch.int64: np.int64,
+        torch.uint64: np.uint64,
+        torch.int32: np.int32,
+        torch.uint32: np.uint32,
+        torch.int16: np.int16,
+        torch.uint16: np.uint16,
+        torch.int8: np.int8,
+        torch.uint8: np.uint8,
+        torch.bool: np.uint8,
+        torch.float8_e4m3fn: np.uint8,
+        torch.float8_e5m2: np.uint8,
+    }
+
     # used for safetensors slices
     # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
     # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
@@ -10089,8 +10123,14 @@ class LazyTorchTensor(gguf.LazyBase):
     @classmethod
     def from_local_tensor(cls, t: gguf.utility.LocalTensor) -> Tensor:
         def load_tensor(tensor: gguf.utility.LocalTensor) -> Tensor:
+            def byteswap_tensor(tensor: np.ndarray, dtype: type) -> np.ndarray:
+                if sys.byteorder == 'big':
+                    # switch data back to big endian
+                    tensor = tensor.view(dtype).byteswap(inplace=False)
+                return tensor
             dtype = cls._dtype_str_map[tensor.dtype]
-            return torch.from_numpy(tensor.mmap_bytes()).view(dtype).reshape(tensor.shape)
+            numpy_dtype = cls._dtype_byteswap_map[dtype]
+            return torch.from_numpy(byteswap_tensor(tensor.mmap_bytes(), numpy_dtype)).view(dtype).reshape(tensor.shape)
         dtype = cls._dtype_str_map[t.dtype]
         shape = t.shape
         lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(t,), func=lambda r: load_tensor(r))
@@ -10098,10 +10138,16 @@ class LazyTorchTensor(gguf.LazyBase):
 
     @classmethod
     def from_remote_tensor(cls, remote_tensor: gguf.utility.RemoteTensor):
+        def byteswap_tensor(tensor: np.ndarray, dtype: type) -> np.ndarray:
+            if sys.byteorder == 'big':
+                # switch data back to big endian
+                tensor = tensor.view(dtype).byteswap(inplace=False)
+            return tensor
         dtype = cls._dtype_str_map[remote_tensor.dtype]
+        numpy_dtype = cls._dtype_byteswap_map[dtype]
         shape = remote_tensor.shape
         meta = cls.meta_with_dtype_and_shape(dtype, shape)
-        lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.frombuffer(r.data(), dtype=dtype).reshape(shape))
+        lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.from_numpy(byteswap_tensor(np.frombuffer(r.data(), dtype=numpy_dtype), numpy_dtype)).view(dtype).reshape(shape))
         return cast(torch.Tensor, lazy)
 
     @classmethod

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...")

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

@@ -530,6 +530,7 @@ extern "C" {
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
+        GGML_OP_TOP_K,
         GGML_OP_LEAKY_RELU,
         GGML_OP_TRI,
         GGML_OP_FILL,
@@ -2258,18 +2259,25 @@ extern "C" {
             struct ggml_tensor  * a,
             enum ggml_sort_order  order);
 
+    // similar to ggml_top_k but implemented as `argsort` + `view`
+    GGML_API struct ggml_tensor * ggml_argsort_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
+    // top k elements per row
+    // note: the resulting top k indices are in no particular order
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
     GGML_API struct ggml_tensor * ggml_arange(
             struct ggml_context * ctx,
             float                 start,
             float                 stop,
             float                 step);
 
-    // top k elements per row
-    GGML_API struct ggml_tensor * ggml_top_k(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   k);
-
 #define GGML_KQ_MASK_PAD 64
 
     // q:    [n_embd_k, n_batch,     n_head,    ne3 ]

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-cann/aclnn_ops.cpp

@@ -42,6 +42,7 @@
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_fill_scalar.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_ger.h>
 #include <aclnnop/aclnn_group_norm.h>
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_gt_scalar.h>
@@ -2206,78 +2207,120 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context & ctx,
 }
 
 /**
- * @brief Initializes and caches sine/cosine positional encoding values
- *        (used in RoPE, Rotary Position Embedding) for attention layers.
+ * @brief Initializes and caches all intermediate tensors required for RoPE
+ *        (Rotary Position Embedding), including support for Yarn, mRoPE,
+ *        i-mRoPE, Neox repeat strategy, independent sectors, frequency factors，
+ *        and multi-section rotary groups.
  *
- * This function computes and caches the sin/cos values of
- * θ = position * theta_scale for RoPE encoding. The cache is shared
- * across attention layers, and only the first attention layer will
- * trigger initialization. The cache includes repeated sin/cos values
- * with different repeat methods depending on the @param is_neox flag.
+ * This function computes and caches the per-dimension θ coefficients used for
+ * Q/K rotary embedding. The cache is shared across layers, and recomputed only
+ * when any dependent parameter changes.
  *
- * Steps performed by this function:
- *   1. Identify whether the target tensor belongs to Q/K in attention
- *      and restrict computation to the first layer only.
- *   2. Initialize the theta scale array (arange → power → freq scaling).
- *   3. Allocate sin/cos caches if the max prompt length increases.
- *   4. Compute θ = position * theta_scale.
- *   5. Compute sin(θ), cos(θ) and optionally scale by attn_factor.
- *   6. Expand sin/cos values by repeat or repeat_interleave depending
- *      on whether @param is_neox is enabled.
+ * The function now supports:
+ *   - Yarn RoPE extrapolation (via @param corr_dims and @param ext_factor)
+ *   - Per-dimension independent sector exponent rules (indep_sects + sections[])
+ *   - Multi-section RoPE (mRoPE) index mapping (mrope_used + is_imrope)
+ *   - Frequency factor division (src2)
+ *   - Neox / normal repeat expansion modes
  *
- * @param ctx                The CANN backend context, holding memory pool,
- *                           stream, and persistent buffers for rope init/cache.
- * @param dst                The destination ggml_tensor whose computation
- *                           depends on the RoPE values (usually Qcur/Kcur).
- * @param theta_scale        Scalar exponent base for computing theta scale values.
- * @param freq_scale         Frequency scaling factor, applied to theta scale.
- * @param attn_factor        Attention scaling factor, applied to sin/cos.
- * @param is_neox            Whether to use Neox-style repeat strategy
- *                           (dim expansion vs repeat_interleave).
+ * @param ctx                CANN backend context, containing memory pool,
+ *                           cached buffers, and runtime stream.
+ * @param dst                Destination ggml_tensor whose computation
+ *                           depends on RoPE (typically Qcur or Kcur).
+ * @param corr_dims          [low, high] Yarn correction range.
+ * @param ext_factor         Yarn extrapolation strength. 0 = disabled.
+ * @param theta_scale        Base multiplier for per-dimension θ exponent.
+ * @param freq_scale         Global frequency scaling factor.
+ * @param attn_factor        Optional scaling applied to sin/cos (if needed).
+ * @param is_neox            Whether to use Neox-style dimension interleave.
+ * @param sections           4-way sector sizes for independent-section RoPE
+ *                           and multi-section mRoPE (t/h/w/e).
+ * @param mrope_used         Whether to enable multi-section rotary embedding.
+ * @param is_imrope          Whether to apply interleaved mRoPE rules.
+ * @param indep_sects        Whether each dimension runs independent exponent
+ *                           resets based on @p sections.
  */
-static void aclnn_cache_init(ggml_backend_cann_context & ctx,
-                             ggml_tensor *               dst,
-                             float *                     corr_dims,
-                             float                       ext_factor,
-                             float                       theta_scale,
-                             float                       freq_scale,
-                             float                       attn_factor,
-                             bool                        is_neox) {
+static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
+                                  ggml_tensor *               dst,
+                                  float *                     corr_dims,
+                                  float                       ext_factor,
+                                  float                       theta_scale,
+                                  float                       freq_scale,
+                                  float                       attn_factor,
+                                  bool                        is_neox,
+                                  int                         sections[4],
+                                  bool                        mrope_used,
+                                  bool                        is_imrope,
+                                  bool                        indep_sects) {
     ggml_tensor * src0 = dst->src[0];  // input
     ggml_tensor * src1 = dst->src[1];  // position
     ggml_tensor * src2 = dst->src[2];  // freq_factors
 
-    if (src2 == nullptr && ctx.rope_cache.cached && ctx.rope_cache.ext_factor == ext_factor &&
-        ctx.rope_cache.theta_scale == theta_scale && ctx.rope_cache.freq_scale == freq_scale &&
-        ctx.rope_cache.attn_factor == attn_factor && ctx.rope_cache.is_neox == is_neox) {
+    int64_t theta_scale_length = src0->ne[0] / 2;
+    int64_t position_length    = dst->ne[2];
+
+    // TODO: check theta_scale_length and position_length.
+    if (src2 == nullptr && ctx.rope_cache.cached &&
+        ctx.rope_cache.equal(theta_scale_length, position_length, ext_factor, theta_scale, freq_scale, attn_factor,
+                             is_neox, indep_sects, mrope_used, is_imrope, sections)) {
         // use cache.
         return;
     }
 
-    int64_t theta_scale_length = src0->ne[0] / 2;
-    int64_t theta_scale_ne[]   = { theta_scale_length, 1, 1, 1 };
-    size_t  theta_scale_nb[]   = { sizeof(float), sizeof(float), sizeof(float), theta_scale_length * sizeof(float) };
+    // Step0: calculate tensor shape.
+    int64_t theta_scale_ne[] = { theta_scale_length, 1, 1, 1 };
+    size_t  theta_scale_nb[] = { sizeof(float), theta_scale_length * sizeof(float), theta_scale_length * sizeof(float),
+                                 theta_scale_length * sizeof(float) };
 
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    int64_t position_length = src1->ne[0];
-    int64_t position_ne[]   = { 1, 1, position_length, 1 };
-    size_t  position_nb[]   = { sizeof(int32_t), sizeof(int32_t), sizeof(int32_t), sizeof(int32_t) * position_length };
+    int64_t position_ne[] = { 1, 1, position_length, 1 };
+    size_t  position_nb[] = { sizeof(int32_t), sizeof(int32_t), sizeof(int32_t), sizeof(int32_t) * position_length };
 
-    int64_t theta_ne[] = { theta_scale_length, 1, position_length, 1 };
-    size_t  theta_nb[GGML_MAX_DIMS];
-    theta_nb[0] = sizeof(float);
+    int64_t cache_ne[] = { theta_scale_length, 1, position_length, 1 };
+    size_t  cache_nb[GGML_MAX_DIMS];
+    cache_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
+        cache_nb[i] = cache_nb[i - 1] * cache_ne[i - 1];
     }
 
-    // theta_scale arange, [0,1,...,ne00/2 - 1]
+    // Step1: Compute the coefficient of theta. During the cache_init process, aside from
+    // (1) multiplying by the position,
+    // (2) dividing by freq_factors,
+    // (3) computing the sine and cosine,
+    // the other parameters used in the computation generally do not change in most scenarios.
+    // Therefore, we can first compute this part of the result and then cache it.
+
+    // Step1.1: prepare theta_scale exponent. if this exponent updated, should update theta_scale_tensor.
     acl_tensor_ptr acl_theta_scale_tensor;
-    // cache theta scale
-    if (ctx.rope_cache.theta_scale_length != theta_scale_length ||
-        // theta_scale and freq_scale should not change during the current token inference process,
-        // so we can directly use == here instead of comparing the absolute difference.
-        ctx.rope_cache.theta_scale != theta_scale || ctx.rope_cache.freq_scale != freq_scale) {
-        ctx.rope_cache.theta_scale_length = theta_scale_length;
+    bool           theta_scale_updated = false;
+    if (ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.theta_scale != theta_scale ||
+        ctx.rope_cache.indep_sects != indep_sects) {
+        theta_scale_updated = true;
+        if (ctx.rope_cache.theta_scale_exp_host != nullptr) {
+            free(ctx.rope_cache.theta_scale_exp_host);
+        }
+        ctx.rope_cache.theta_scale_exp_host = (float *) malloc(theta_scale_length * sizeof(float));
+        GGML_ASSERT(ctx.rope_cache.theta_scale_exp_host != nullptr);
+        if (!indep_sects) {
+            ctx.rope_cache.theta_scale_exp_host[0] = 1;
+            for (int i = 1; i < theta_scale_length; i++) {
+                ctx.rope_cache.theta_scale_exp_host[i] = ctx.rope_cache.theta_scale_exp_host[i - 1] * theta_scale;
+            }
+        } else {
+            int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+            int sec_w     = sections[1] + sections[0];
+            int sec_e     = sections[2] + sec_w;
+
+            ctx.rope_cache.theta_scale_exp_host[0] = 1;
+            for (int i = 1; i < theta_scale_length; i++) {
+                int sector = i % sect_dims;
+                if (sector == 0 || sector == sections[0] || sector == sec_w || sector == sec_e) {
+                    ctx.rope_cache.theta_scale_exp_host[i] = 1;
+                    continue;
+                }
+                ctx.rope_cache.theta_scale_exp_host[i] = ctx.rope_cache.theta_scale_exp_host[i - 1] * theta_scale;
+            }
+        }
 
         if (ctx.rope_cache.theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(ctx.rope_cache.theta_scale_cache));
@@ -2285,74 +2328,138 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
         ACL_CHECK(aclrtMalloc(&ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float),
                               ACL_MEM_MALLOC_HUGE_FIRST));
 
+        ACL_CHECK(aclrtMemcpyAsync(ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float),
+                                   ctx.rope_cache.theta_scale_exp_host, theta_scale_length * sizeof(float),
+                                   ACL_MEMCPY_HOST_TO_DEVICE, ctx.stream()));
+
         acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
                                                          theta_scale_ne, theta_scale_nb, 1);
+    }
 
-        float start      = 0;
-        float step       = 1;
-        float stop       = theta_scale_length;
-        float n_elements = theta_scale_length;
-        aclnn_arange(ctx, acl_theta_scale_tensor.get(), start, stop, step, n_elements);
+    // Step1.2: prepare rope_yarn_ramp, if this part updated, should update theta_scale_tensor.
+    bool                 yarn_ramp_tensor_updated = false;
+    ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
+    acl_tensor_ptr       acl_yarn_ramp_tensor;
+    if (ext_factor != 0 &&
+        // TODO: check more parameter.
+        (ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.freq_scale != freq_scale)) {
+        yarn_ramp_tensor_updated = true;
 
-        ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
-        acl_tensor_ptr       acl_yarn_ramp_tensor;
-        if (ext_factor != 0) {
-            // -rope_yarn_ramp
-            // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
-            // return MIN(1, MAX(0, y)) - 1;
-            yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
-            void * yarn_ramp_buffer = yarn_ramp_allocator.get();
-            acl_yarn_ramp_tensor =
-                ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
-            float          zero_value = 0, one_value = 1;
-            float          denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
-            acl_scalar_ptr low              = ggml_cann_create_scalar(&corr_dims[0], aclDataType::ACL_FLOAT);
-            acl_scalar_ptr zero             = ggml_cann_create_scalar(&zero_value, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr one              = ggml_cann_create_scalar(&one_value, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr denom_safe       = ggml_cann_create_scalar(&denom_safe_value, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr ext_factor_sc    = ggml_cann_create_scalar(&ext_factor, aclDataType::ACL_FLOAT);
+        // -rope_yarn_ramp
+        // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
+        // return MIN(1, MAX(0, y)) - 1;
+        yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
+        void * yarn_ramp_buffer = yarn_ramp_allocator.get();
+        acl_yarn_ramp_tensor =
+            ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
+        float          zero_value = 0, one_value = 1;
+        float          denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
+        acl_scalar_ptr low              = ggml_cann_create_scalar(&corr_dims[0], aclDataType::ACL_FLOAT);
+        acl_scalar_ptr zero             = ggml_cann_create_scalar(&zero_value, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr one              = ggml_cann_create_scalar(&one_value, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr denom_safe       = ggml_cann_create_scalar(&denom_safe_value, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr ext_factor_sc    = ggml_cann_create_scalar(&ext_factor, aclDataType::ACL_FLOAT);
 
-            GGML_CANN_CALL_ACLNN_OP(ctx, Subs, acl_theta_scale_tensor.get(), low.get(), one.get(),
-                                    acl_yarn_ramp_tensor.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceDivs, acl_yarn_ramp_tensor.get(), denom_safe.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceThreshold, acl_yarn_ramp_tensor.get(), zero.get(), zero.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceClampMax, acl_yarn_ramp_tensor.get(), one.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor.get(), one.get(), one.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), ext_factor_sc.get());
+        aclnn_arange(ctx, acl_yarn_ramp_tensor.get(), 0, theta_scale_length, 1, theta_scale_length);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor.get(), low.get(), one.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceDivs, acl_yarn_ramp_tensor.get(), denom_safe.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceThreshold, acl_yarn_ramp_tensor.get(), zero.get(), zero.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceClampMax, acl_yarn_ramp_tensor.get(), one.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor.get(), one.get(), one.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), ext_factor_sc.get());
 
-            // theta_interp = freq_scale * theta_extrap;
-            // theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-            // theta = freq_scale * theta_extrap * (1 - ramp_mix) + theta_extrap * ramp_mix;
-            // theta = freq_scale * theta_extrap - freq_scale * theta_extrap * ramp_mix + theta_extrap * ramp_mix;
-            // theta = theta_extrap * (freq_scale - freq_scale * ramp_mix + ramp_mix);
-            //
-            // we cache (freq_scale - freq_scale * ramp_mix + ramp_mix), Considering that the rope_yarn_ramp here is the inverse
-            // cache freq_scale + (freq_scale - 1) * ramp_mix
-            float          freq_scale_1    = freq_scale - 1;
-            acl_scalar_ptr freq_scale_sc   = ggml_cann_create_scalar(&freq_scale, aclDataType::ACL_FLOAT);
-            acl_scalar_ptr freq_scale_1_sc = ggml_cann_create_scalar(&freq_scale_1, aclDataType::ACL_FLOAT);
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), freq_scale_1_sc.get());
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor.get(), freq_scale_sc.get(), one.get());
-        }
+        // theta_interp = freq_scale * theta_extrap;
+        // theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+        // theta = freq_scale * theta_extrap * (1 - ramp_mix) + theta_extrap * ramp_mix;
+        // theta = freq_scale * theta_extrap - freq_scale * theta_extrap * ramp_mix + theta_extrap * ramp_mix;
+        // theta = theta_extrap * (freq_scale - freq_scale * ramp_mix + ramp_mix);
+        //
+        // we cache (freq_scale - freq_scale * ramp_mix + ramp_mix), Considering that the rope_yarn_ramp here is the inverse
+        // cache freq_scale + (freq_scale - 1) * ramp_mix
+        float          freq_scale_1    = freq_scale - 1;
+        acl_scalar_ptr freq_scale_sc   = ggml_cann_create_scalar(&freq_scale, aclDataType::ACL_FLOAT);
+        acl_scalar_ptr freq_scale_1_sc = ggml_cann_create_scalar(&freq_scale_1, aclDataType::ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), freq_scale_1_sc.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor.get(), freq_scale_sc.get(), one.get());
+    }
 
-        // power
-        acl_scalar_ptr acl_theta_scale = ggml_cann_create_scalar(&theta_scale, aclDataType::ACL_FLOAT);
-        GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale.get(), acl_theta_scale_tensor.get(),
-                                acl_theta_scale_tensor.get());
-
-        if (ext_factor != 0) {
+    // Step 1.3: update theta_scale_tensor according to ext_factor or freq_scale.
+    if (ext_factor != 0) {
+        if (theta_scale_updated || yarn_ramp_tensor_updated) {
+            theta_scale_updated = true;
             aclnn_mul(ctx, acl_theta_scale_tensor.get(), acl_yarn_ramp_tensor.get());
-        } else if (freq_scale != 1) {
-            aclnn_muls(ctx, acl_theta_scale_tensor.get(), freq_scale, nullptr, true);
         }
     } else {
-        // use cache
+        if (freq_scale != 1 && (ctx.rope_cache.freq_scale != freq_scale || theta_scale_updated)) {
+            theta_scale_updated = true;
+            aclnn_muls(ctx, acl_theta_scale_tensor.get(), freq_scale, nullptr, true);
+        }
+    }
+
+    // Nothing changed, use cache.
+    if (!theta_scale_updated) {
         acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
                                                          theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
     }
 
+    // Step 1.4: prepare select index if mrope
+    acl_tensor_ptr position_select_index_tensor;
+    if (mrope_used) {
+        if (ctx.rope_cache.sections[0] != sections[0] || ctx.rope_cache.sections[1] != sections[1] ||
+            ctx.rope_cache.sections[2] != sections[2] || ctx.rope_cache.sections[3] != sections[3] ||
+            ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.is_imrope != is_imrope) {
+            if (ctx.rope_cache.position_select_index_host != nullptr) {
+                free(ctx.rope_cache.position_select_index_host);
+            }
+            ctx.rope_cache.position_select_index_host = (int *) malloc(theta_scale_length * sizeof(int));
+            GGML_ASSERT(ctx.rope_cache.position_select_index_host != nullptr);
+            int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+            int sec_w     = sections[1] + sections[0];
+            int sec_e     = sections[2] + sec_w;
+            // t,h,w,e
+            for (int i = 0; i < theta_scale_length; i++) {
+                int sector = i % sect_dims;
+
+                if (is_imrope) {  // qwen3vl apply interleaved mrope
+                    if (sector % 3 == 1 && sector < 3 * sections[1]) {
+                        ctx.rope_cache.position_select_index_host[i] = 1;
+                    } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
+                        ctx.rope_cache.position_select_index_host[i] = 2;
+                    } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                        ctx.rope_cache.position_select_index_host[i] = 0;
+                    } else {
+                        ctx.rope_cache.position_select_index_host[i] = 3;
+                    }
+                } else {
+                    if (sector >= sections[0] && sector < sec_w) {
+                        ctx.rope_cache.position_select_index_host[i] = 1;
+                    } else if (sector >= sec_w && sector < sec_e) {
+                        ctx.rope_cache.position_select_index_host[i] = 2;
+                    } else if (sector >= sec_e) {
+                        ctx.rope_cache.position_select_index_host[i] = 3;
+                    } else {
+                        ctx.rope_cache.position_select_index_host[i] = 0;
+                    }
+                }
+            }
+
+            if (ctx.rope_cache.position_select_index != nullptr) {
+                ACL_CHECK(aclrtFree(ctx.rope_cache.position_select_index));
+            }
+            ACL_CHECK(aclrtMalloc(&ctx.rope_cache.position_select_index, theta_scale_length * sizeof(int),
+                                  ACL_MEM_MALLOC_HUGE_FIRST));
+
+            ACL_CHECK(aclrtMemcpyAsync(ctx.rope_cache.position_select_index, theta_scale_length * sizeof(int),
+                                       ctx.rope_cache.position_select_index_host, theta_scale_length * sizeof(int),
+                                       ACL_MEMCPY_HOST_TO_DEVICE, ctx.stream()));
+        }
+
+        position_select_index_tensor = ggml_cann_create_tensor(ctx.rope_cache.position_select_index, ACL_INT32,
+                                                               sizeof(int), theta_scale_ne, theta_scale_nb, 1);
+    }
+
+    // Step2: divide by freq_factors
     ggml_cann_pool_alloc freq_fac_res_allocator(ctx.pool());
-    // freq_factors
     if (src2) {
         freq_fac_res_allocator.alloc(theta_scale_length * sizeof(float));
         void *         freq_fac_res_ptr = freq_fac_res_allocator.get();
@@ -2365,6 +2472,85 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
         std::swap(acl_theta_scale_tensor, acl_freq_fac_res_tensor);
     }
 
+    // Step3: prepare position_tensor
+    acl_tensor_ptr       acl_position_tensor;
+    ggml_cann_pool_alloc mrope_position_acllocator(ctx.pool());
+    if (mrope_used) {
+        // Step3.1: select current position;
+        // position :
+        // pos1: [[0, 1 ,2 ,3 ],
+        // pos2:  [4, 5 ,6 ,7 ],
+        // pos3:  [8, 9 ,10,11],
+        // pos4:  [12,13,14,15] ]
+        //
+        // select index = [0, 1, 2, 2, 1, 0]
+        //
+        // selected_tensor:
+        // [[0, 1 ,2 ,3 ],
+        //  [4, 5 ,6 ,7 ],
+        //  [8, 9 ,10,11],
+        //  [8, 9 ,10,11],
+        //  [4, 5 ,6 ,7 ],
+        //  [0, 1 ,2 ,3 ]]
+        //
+        // transpose, from [seq_len:dims] to [dims:seq_len]
+        // [0, 4, 8 ,8 ,4, 0],
+        // [1, 5, 9, 9, 5, 1],
+        // [2, 6, 10,10,6 ,2],
+        // [3, 7, 11,11,7 3 ]]
+        //
+        // multipy by theta_scale_tensor
+        // [theta_scale^0, theta_scale^1, ..., theta_scale ^ n]
+
+        int64_t        mrope_position_ne[] = { position_length, 4 };
+        size_t         mrope_position_nb[] = { sizeof(int), position_length * sizeof(int) };
+        acl_tensor_ptr mrope_position =
+            ggml_cann_create_tensor(src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type),
+                                    mrope_position_ne, mrope_position_nb, 2);
+
+        // selected position tensor's shape is a transpose of cache tensor.
+        int64_t selected_position_ne[] = { position_length, theta_scale_length };
+        size_t  selected_position_nb[] = { sizeof(float), position_length * sizeof(float) };
+        mrope_position_acllocator.alloc(theta_scale_length * position_length * sizeof(float));
+        void * mrope_position_buffer = mrope_position_acllocator.get();
+        acl_position_tensor =
+            ggml_cann_create_tensor(mrope_position_buffer, ggml_cann_type_mapping(src1->type),
+                                    ggml_type_size(src1->type), selected_position_ne, selected_position_nb, 2);
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, mrope_position.get(), 0, position_select_index_tensor.get(),
+                                acl_position_tensor.get());
+
+        // transpose
+        int64_t transposed_ne[] = { position_length, 1, theta_scale_length, 1 };
+        size_t  transposed_nb[GGML_MAX_DIMS];
+        transposed_nb[0] = sizeof(float);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            transposed_nb[i] = transposed_nb[i - 1] * transposed_ne[i - 1];
+        }
+
+        std::swap(transposed_ne[0], transposed_ne[2]);
+        std::swap(transposed_nb[0], transposed_nb[2]);
+
+        acl_position_tensor =
+            ggml_cann_create_tensor(mrope_position_buffer, ggml_cann_type_mapping(src1->type),
+                                    ggml_type_size(src1->type), transposed_ne, transposed_nb, GGML_MAX_DIMS);
+
+    } else {
+        // auto bcast.
+        acl_position_tensor =
+            ggml_cann_create_tensor(src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type),
+                                    position_ne, position_nb, GGML_MAX_DIMS);
+    }
+
+    // Step4: multiply by the position
+    int64_t              theta_length = theta_scale_length * position_length;
+    ggml_cann_pool_alloc theta_allocator(ctx.pool(), theta_length * sizeof(float));
+    void *               theta_buffer = theta_allocator.get();
+
+    acl_tensor_ptr acl_theta_tensor =
+        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float), cache_ne, cache_nb, GGML_MAX_DIMS);
+    aclnn_mul(ctx, acl_position_tensor.get(), acl_theta_scale_tensor.get(), acl_theta_tensor.get());
+
+    // Step5: calculate sin cos.
     // init sin_repeat && cos_repeat, only to accelerate first layer on each device
     if (position_length > ctx.rope_cache.position_length) {
         ctx.rope_cache.position_length = position_length;
@@ -2381,44 +2567,30 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
             aclrtMalloc(&ctx.rope_cache.cos_cache, repeat_theta_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
     }
 
-    // position
-    acl_tensor_ptr acl_position_tensor =
-        ggml_cann_create_tensor(src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type), position_ne,
-                                position_nb, GGML_MAX_DIMS);
-
-    // power * position
-    int64_t              theta_length = theta_scale_length * position_length;
-    ggml_cann_pool_alloc theta_allocator(ctx.pool(), theta_length * sizeof(float));
-    void *               theta_buffer = theta_allocator.get();
-
-    acl_tensor_ptr acl_theta_tensor =
-        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb, GGML_MAX_DIMS);
-    aclnn_mul(ctx, acl_position_tensor.get(), acl_theta_scale_tensor.get(), acl_theta_tensor.get());
-
     // sin/cos
     ggml_cann_pool_alloc sin_allocator(ctx.pool(), theta_length * sizeof(float));
     void *               sin_buffer = sin_allocator.get();
     acl_tensor_ptr       acl_sin_tensor =
-        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float), cache_ne, cache_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_sin(ctx, acl_theta_tensor.get(), acl_sin_tensor.get());
 
     ggml_cann_pool_alloc cos_allocator(ctx.pool(), theta_length * sizeof(float));
     void *               cos_buffer = cos_allocator.get();
     acl_tensor_ptr       acl_cos_tensor =
-        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float), cache_ne, cache_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_theta_tensor.get(), acl_cos_tensor.get());
 
     if (ext_factor != 0) {
         attn_factor *= 1.0f + 0.1f * logf(1.0f / freq_scale);
     }
 
-    // attn_factor
+    // Step 5: multiply by attn_factor
     if (attn_factor != 1) {
         aclnn_muls(ctx, acl_sin_tensor.get(), attn_factor, nullptr, true);
         aclnn_muls(ctx, acl_cos_tensor.get(), attn_factor, nullptr, true);
     }
 
-    int64_t sin_reshape_ne[4] = { src0->ne[0], 1, src0->ne[2], 1 };
+    int64_t sin_reshape_ne[4] = { src0->ne[0], 1, dst->ne[2], 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
     sin_reshape_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
@@ -2429,8 +2601,9 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
     acl_tensor_ptr acl_cos_repeat_tensor = ggml_cann_create_tensor(ctx.rope_cache.cos_cache, ACL_FLOAT, sizeof(float),
                                                                    sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
 
-    // repeat
+    // Step 6: repeat
     if (is_neox) {
+        // [sinθ1, sinθ1, sinθ2, sinθ2, ..., sinθn, sinθn]
         int64_t repeatsArray[] = { 1, 1, 1, 2 };
         aclnn_repeat(ctx, acl_sin_tensor.get(), acl_sin_repeat_tensor.get(), repeatsArray);
         aclnn_repeat(ctx, acl_cos_tensor.get(), acl_cos_repeat_tensor.get(), repeatsArray);
@@ -2438,17 +2611,15 @@ static void aclnn_cache_init(ggml_backend_cann_context & ctx,
         int64_t num_repeats = 2;
         int64_t dim         = 3;
         int64_t output_size = theta_scale_length * num_repeats;
+        // [sinθ1, sinθ2, ..., sinθn, sinθ1, sinθ2, ..., sinθn]
         aclnn_repeat_interleave(ctx, acl_sin_tensor.get(), acl_sin_repeat_tensor.get(), dim, num_repeats, output_size);
         aclnn_repeat_interleave(ctx, acl_cos_tensor.get(), acl_cos_repeat_tensor.get(), dim, num_repeats, output_size);
     }
 
-    // Other layers use cache except first layer.
-    ctx.rope_cache.cached      = true;
-    ctx.rope_cache.ext_factor  = ext_factor;
-    ctx.rope_cache.theta_scale = theta_scale;
-    ctx.rope_cache.freq_scale  = freq_scale;
-    ctx.rope_cache.attn_factor = attn_factor;
-    ctx.rope_cache.is_neox     = is_neox;
+    // Update cached value.
+    ctx.rope_cache.cached = true;
+    ctx.rope_cache.set(theta_scale_length, position_length, ext_factor, theta_scale, freq_scale, attn_factor, is_neox,
+                       indep_sects, mrope_used, is_imrope, sections);
 }
 
 #ifdef __cplusplus
@@ -2474,6 +2645,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     // param
     float     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
     // const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
     const int mode       = ((int32_t *) dst->op_params)[2];
@@ -2482,12 +2654,13 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
-    memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
-    memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
-    memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
-    memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     // TODO: n_dims <= ne0
     GGML_ASSERT(n_dims == ne0);
@@ -2498,10 +2671,25 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
+    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (mrope_used) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
+
+    if (is_imrope || mrope_used) {
+        is_neox = true;
+    }
 
     // init ctx.rope_cos/rope_sin cache
-    aclnn_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox);
+    aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections, mrope_used, is_imrope, is_vision);
 
     int64_t sin_reshape_ne[4] = { ne00, 1, ne02, 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
@@ -2657,8 +2845,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     return;
 #endif
 
-    // ggml_mode = 0 --> aclnn_model = 1
-    int64_t acl_mode = mode == 0 ? 1 : mode;
+    int64_t acl_mode = is_neox ? 0 : 1;
 
     switch (src0->type) {
         case GGML_TYPE_F32:
@@ -3236,3 +3423,64 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst
         GGML_ABORT("Function is not implemented.");
     }
 }
+
+static void ggml_cann_out_prod_fp(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];  // weight
+    ggml_tensor * src1 = dst->src[1];  // input
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_dst.get());
+
+    const int64_t dps2 = ne2 / ne02;
+    const int64_t dps3 = ne3 / ne03;
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            const int64_t i02 = i2 / dps2;
+            const int64_t i03 = i3 / dps3;
+
+            const int64_t  i12 = i2;
+            const int64_t  i13 = i3;
+            acl_tensor_ptr accumulator =
+                ggml_cann_create_tensor((char *) dst->data + i2 * nb2 + i3 * nb3, ggml_cann_type_mapping(dst->type),
+                                        ggml_type_size(dst->type), dst->ne, dst->nb, 2);
+
+            // The outer product needs to be accumulated in this dimension.
+            for (int64_t i1 = 0; i1 < ne11; i1++) {
+                acl_tensor_ptr acl_input = ggml_cann_create_tensor(
+                    (char *) src1->data + i1 * nb11 + i12 * nb12 + i13 * nb13, ggml_cann_type_mapping(src0->type),
+                    ggml_type_size(src0->type), src1->ne, src1->nb, 1);
+
+                acl_tensor_ptr acl_weight = ggml_cann_create_tensor(
+                    (char *) src0->data + i1 * nb01 + i02 * nb02 + i03 * nb03, ggml_cann_type_mapping(src0->type),
+                    ggml_type_size(src0->type), src0->ne, src0->nb, 1);
+
+                ggml_cann_pool_alloc output_allocator(ctx.pool());
+                void *               output_buffer = output_allocator.alloc(ggml_nbytes(dst));
+                acl_tensor_ptr       acl_out = ggml_cann_create_tensor(output_buffer, ggml_cann_type_mapping(dst->type),
+                                                                       ggml_type_size(dst->type), dst->ne, dst->nb, 2);
+
+                GGML_CANN_CALL_ACLNN_OP(ctx, Ger, acl_input.get(), acl_weight.get(), acl_out.get());
+                float       alpha_value = 1.0f;
+                aclScalar * alpha       = aclCreateScalar(&alpha_value, ACL_FLOAT);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdd, accumulator.get(), acl_out.get(), alpha);
+            }
+        }
+    }
+}
+
+void ggml_cann_out_prod(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+
+    const enum ggml_type type = src0->type;
+
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+            ggml_cann_out_prod_fp(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupport type for GGML_OP_OUT_PROD");
+            break;
+    }
+}

ggml/src/ggml-cann/aclnn_ops.h

@@ -1125,3 +1125,23 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
     } while (0)
 
 #endif  // CANN_ACLNN_OPS
+
+/**
+ * @brief Performs outer product operation on two ggml tensors using the CANN backend.
+ *
+ * @details This function computes the outer product of two input tensors (src0 and src1)
+ * and stores the result in the destination tensor. The outer product operation is defined as:
+ * dst[i,j,k,l] = sum_m (src0[i,m,k,l] * src1[j,m,k,l])
+ *
+ * The function supports multiple data types including F32, F16. For floating-point
+ * types, it uses batch matrix multiplication for efficient computation.
+ *
+ * The implementation handles 4D tensor broadcasting and batch processing automatically.
+ *
+ * @param ctx The CANN backend context for operation execution and memory management.
+ * @param dst The destination ggml_tensor where the outer product result will be stored.
+ *            The input tensors are assumed to be `dst->src[0]` and `dst->src[1]`.
+ *
+ * @see GGML_CANN_CALL_ACLNN_OP for CANN operator invocation
+ */
+void ggml_cann_out_prod(ggml_backend_cann_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cann/common.h

@@ -300,30 +300,92 @@ struct ggml_cann_graph_lru_cache {
 
 struct ggml_cann_rope_cache {
     ~ggml_cann_rope_cache() {
-        if (theta_scale_cache != nullptr) {
+        if (theta_scale_cache) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
-        if (sin_cache != nullptr) {
+        if (sin_cache) {
             ACL_CHECK(aclrtFree(sin_cache));
         }
-        if (cos_cache != nullptr) {
+        if (cos_cache) {
             ACL_CHECK(aclrtFree(cos_cache));
         }
+        if (position_select_index) {
+            ACL_CHECK(aclrtFree(position_select_index));
+        }
+        if (theta_scale_exp_host) {
+            free(theta_scale_exp_host);
+        }
+        if(position_select_index_host) {
+            free(position_select_index_host);
+        }
     }
 
-    void *  theta_scale_cache  = nullptr;
-    int64_t theta_scale_length = 0;
+    bool equal(int64_t theta_scale_length,
+               int64_t position_length,
+               float   ext_factor,
+               float   theta_scale,
+               float   freq_scale,
+               float   attn_factor,
+               bool    is_neox,
+               bool    indep_sects,
+               bool    mrope_used,
+               bool    is_imrope,
+               int     sections[4]) {
+        return this->theta_scale_length == theta_scale_length && this->position_length == position_length &&
+               this->ext_factor == ext_factor && this->theta_scale == theta_scale && this->freq_scale == freq_scale &&
+               this->attn_factor == attn_factor && this->is_neox == is_neox && this->indep_sects == indep_sects &&
+               this->mrope_used == mrope_used && this->is_imrope == is_imrope && this->sections[0] == sections[0] &&
+               this->sections[1] == sections[1] && this->sections[2] == sections[2] && this->sections[3] == sections[3];
+    }
+
+    void set(int64_t theta_scale_length,
+             int64_t position_length,
+             float    ext_factor,
+             float   theta_scale,
+             float   freq_scale,
+             float   attn_factor,
+             bool    is_neox,
+             bool    indep_sects,
+             bool    mrope_used,
+             bool    is_imrope,
+             int     sections[4]) {
+        this->theta_scale_length = theta_scale_length;
+        this->position_length    = position_length;
+        this->ext_factor         = ext_factor;
+        this->theta_scale        = theta_scale;
+        this->freq_scale         = freq_scale;
+        this->attn_factor        = attn_factor;
+        this->is_neox            = is_neox;
+        this->indep_sects        = indep_sects;
+        this->mrope_used         = mrope_used;
+        this->is_imrope          = is_imrope;
+        this->sections[0]        = sections[0];
+        this->sections[1]        = sections[1];
+        this->sections[2]        = sections[2];
+        this->sections[3]        = sections[3];
+    }
+
+    // memory cache, prepare before inferencing.
+    void *  theta_scale_cache          = nullptr;
+    float * theta_scale_exp_host       = nullptr;
+    int *   position_select_index_host = nullptr;
+    void *  position_select_index      = nullptr;
     // sin/cos cache, used only to accelerate first layer on each device
-    void *  sin_cache          = nullptr;
-    void *  cos_cache          = nullptr;
-    int64_t position_length    = 0;
+    void *  sin_cache                  = nullptr;
+    void *  cos_cache                  = nullptr;
     // Properties to check before reusing the sincos cache
-    bool    cached             = false;
-    float   ext_factor         = 0.0f;
-    float   theta_scale        = 0.0f;
-    float   freq_scale         = 0.0f;
-    float   attn_factor        = 0.0f;
-    bool    is_neox            = false;
+    int64_t theta_scale_length         = 0;
+    int64_t position_length            = 0;
+    bool    cached                     = false;
+    float   ext_factor                 = 0.0f;
+    float   theta_scale                = 0.0f;
+    float   freq_scale                 = 0.0f;
+    float   attn_factor                = 0.0f;
+    bool    is_neox                    = false;
+    bool    indep_sects                = false;
+    bool    mrope_used                 = false;
+    int     sections[4]                = { 0, 0, 0, 0 };
+    bool    is_imrope                  = false;
 };
 
 struct ggml_cann_tensor_cache {

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

@@ -1886,6 +1886,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_FLASH_ATTN_EXT:
             ggml_cann_flash_attn_ext(ctx, dst);
             break;
+        case GGML_OP_OUT_PROD:
+            ggml_cann_out_prod(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2246,8 +2249,7 @@ 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
@@ -2271,12 +2273,14 @@ 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
@@ -2302,9 +2306,9 @@ 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) {
@@ -2335,7 +2339,6 @@ 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);
 
@@ -2477,13 +2480,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                     return false;
                 }
 
-                const int mode = ((const int32_t *) op->op_params)[2];
-                if (mode & GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                if (mode & GGML_ROPE_TYPE_VISION) {
-                    return false;
-                }
                 if (op->src[0]->ne[0] > 896) {
                     return false;
                 }
@@ -2563,6 +2559,16 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_OUT_PROD:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                        return true;
+                    default:
+                        return false;
+                }
+            }
         case GGML_OP_CONV_TRANSPOSE_1D:
             // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255.
             return (op->src[0]->ne[0] - 1) <= 255;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -224,7 +224,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
             include(CheckCXXSourceCompiles)
             set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS}")
+            string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
+            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
             foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
                 set(ARM_FEATURE "HAVE_${feature}")
                 check_cxx_source_compiles(
@@ -452,22 +453,35 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ggml-cpu/spacemit/ime_kernels.h
             )
         endif()
-        set(MARCH_STR "rv64gc")
-        if (GGML_RV_ZFH)
-            string(APPEND MARCH_STR "_zfh")
-        endif()
-        if (GGML_XTHEADVECTOR)
-            string(APPEND MARCH_STR "_xtheadvector")
-        elseif (GGML_RVV)
-            string(APPEND MARCH_STR "_v")
-            if (GGML_RV_ZVFH)
-                string(APPEND MARCH_STR "_zvfh")
+        if(NOT GGML_CPU_ALL_VARIANTS)
+            set(MARCH_STR "rv64gc")
+            if (GGML_RV_ZFH)
+                string(APPEND MARCH_STR "_zfh")
             endif()
+            if (GGML_XTHEADVECTOR)
+                string(APPEND MARCH_STR "_xtheadvector")
+            elseif (GGML_RVV)
+                string(APPEND MARCH_STR "_v")
+                if (GGML_RV_ZVFH)
+                    string(APPEND MARCH_STR "_zvfh")
+                endif()
+            endif()
+            if (GGML_RV_ZICBOP)
+                string(APPEND MARCH_STR "_zicbop")
+            endif()
+            list(APPEND ARCH_FLAGS "-march=${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

@@ -33,10 +33,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -44,27 +46,30 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #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)
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__POWERPC__) || defined(__powerpc__)
 // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
@@ -76,10 +81,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -87,6 +94,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -101,10 +109,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -112,6 +122,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -134,15 +145,18 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -163,10 +177,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -174,6 +190,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -196,10 +213,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -207,6 +226,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0

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 +497,295 @@ 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_8x4_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_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4; // 0123 and 4567
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[col_groups];
+
+    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 < col_groups; 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_0123 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_4567 = 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_0123   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_4567   = 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) };
+            int32x4_t acc_lo[col_groups];
+            int32x4_t acc_hi[col_groups];
+
+            // 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_groups; 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];
+                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));
+                }
+
+                int8x16_t q8_qs[64 / 16];
+                for (int i = 0; i < 64 / 16; i++) {
+                    q8_qs[i] = vld1q_s8(q8_ptr[b].qs + sb * 64 + i * 16);
+                }
+
+                for (int c = 0; c < col_groups; c++) {
+                    uint8x16_t q4_cols[8];
+                    for (int i = 0; i < 8; i++) {
+                        q4_cols[i] = vld1q_u8(q4_ptr[b].qs + sb * QK_K + i * 32 + 16 * c);
+                    }
+
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[0], m4b)), q8_qs[0], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[1], m4b)), q8_qs[0], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[2], m4b)), q8_qs[0], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[3], m4b)), q8_qs[0], 3);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[4], m4b)), q8_qs[1], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[5], m4b)), q8_qs[1], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[6], m4b)), q8_qs[1], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[7], m4b)), q8_qs[1], 3);
+
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[0], 4)), q8_qs[2], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[1], 4)), q8_qs[2], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[2], 4)), q8_qs[2], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[3], 4)), q8_qs[2], 3);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[4], 4)), q8_qs[3], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[5], 4)), q8_qs[3], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[6], 4)), q8_qs[3], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[7], 4)), q8_qs[3], 3);
+                }
+
+                // Scales
+                // row c0123 blk0 and blk1
+                const int16x4_t   sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                const int16x4_t   sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                const float32x4_t sumf_0123  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[0]),
+                                                                       vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[0])));
+                acc_f32[0]                   = vfmaq_f32(acc_f32[0], sb_scale_0123, sumf_0123);
+                // row c4567 blk0 and blk1
+                const int16x4_t   sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                const int16x4_t   sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                const float32x4_t sumf_4567  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[1]),
+                                                                       vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[1])));
+                acc_f32[1]                   = vfmaq_f32(acc_f32[1], sb_scale_4567, sumf_4567);
+
+                // Bias Correction
+                const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                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]));
+                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_0123);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_4567);
+        }  // 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  // #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x4_q8_K_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) && defined(__ARM_FEATURE_DOTPROD)
+    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
+                    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) && defined(__ARM_FEATURE_DOTPROD)
+    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 +2201,412 @@ 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_8x4_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          = 4;
+
+    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_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    acc_size  = 2 * 4;  // 2 row pairs × 4 col pairs
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[acc_size];
+
+    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 < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // d4 0 1 2 3, 4 5 6 7
+                float32x4_t q4_d_0123    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));
+                float32x4_t q4_d_4567    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));
+                // d8 0 1 2 3
+                float32x4_t q8_d_0123    = vld1q_f32(q8_ptr[b].d);
+                // mins
+                float32x4_t q4_dmin_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));
+                float32x4_t q4_dmin_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));
+
+                // Precomputation of scales and mins
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+                float32x4_t sbd_min_0123[q8_k_blocklen];
+                float32x4_t sbd_min_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 0);
+                sbd_min_0123[0]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 0);
+                sbd_min_4567[0]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 0);
+
+                sbd_scale_0123[1] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 1);
+                sbd_min_0123[1]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 1);
+                sbd_min_4567[1]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 1);
+
+                sbd_scale_0123[2] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 2);
+                sbd_min_0123[2]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 2);
+                sbd_min_4567[2]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 2);
+
+                sbd_scale_0123[3] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 3);
+                sbd_min_0123[3]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 3);
+                sbd_min_4567[3]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 3);
+
+                // Precomputation of bsums, each vpaddq calcs all the bsums for each row
+                const int16x8_t bsums[q8_k_blocklen] = {
+                    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[QK_K / 64][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                // interleaved bias_acc: [0]->r0 0123, [1]->r1 0123, .., [4]->r0 4567, [5]->r1 4567 ..
+                int32x4_t bias_acc[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Int accumulators for qs vecdot (4 row x 2 col quartets)
+                    int32x4_t acc_lo[acc_size];
+                    int32x4_t acc_hi[acc_size];
+                    for (int i = 0; i < acc_size; 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_scales[2];
+                    int16x8_t q4sb_mins[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));
+                    }
+
+                    constexpr int reads_per_sb = 8;  // 8 * 16 bytes each => 32 qs * 4 rows
+                    for (int k = 0; k < reads_per_sb; k++) {
+                        const int8x16_t q8_blk0 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k);
+                        const int8x16_t q8_blk1 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k + 128);
+
+                        // 0..3 & 32..35
+                        const uint8x16_t q4_0123 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k);
+                        const uint8x16_t q4_4567 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k + 16);
+
+                        const int8x16_t q4_0123_lo = vreinterpretq_s8_u8(vandq_u8(q4_0123, m4b));
+                        const int8x16_t q4_0123_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_0123, 4));
+
+                        acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q4_0123_lo, q8_blk0, 0);  //  0..3  r0 c0123
+                        acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q4_0123_lo, q8_blk0, 1);  //  0..3  r1 c0123
+                        acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q4_0123_lo, q8_blk0, 2);  //  0..3  r2 c0123
+                        acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q4_0123_lo, q8_blk0, 3);  //  0..3  r3 c0123
+
+                        acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q4_0123_hi, q8_blk1, 0);  // 32..35 r0 c0123
+                        acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q4_0123_hi, q8_blk1, 1);  // 32..35 r1 c0123
+                        acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q4_0123_hi, q8_blk1, 2);  // 32..35 r2 c0123
+                        acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q4_0123_hi, q8_blk1, 3);  // 32..35 r3 c0123
+
+                        const int8x16_t q4_4567_lo = vreinterpretq_s8_u8(vandq_u8(q4_4567, m4b));
+                        const int8x16_t q4_4567_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_4567, 4));
+
+                        acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q4_4567_lo, q8_blk0, 0);  //  0..3  r0 c4567
+                        acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q4_4567_lo, q8_blk0, 1);  //  0..3  r1 c4567
+                        acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q4_4567_lo, q8_blk0, 2);  //  0..3  r2 c4567
+                        acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q4_4567_lo, q8_blk0, 3);  //  0..3  r3 c4567
+
+                        acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q4_4567_hi, q8_blk1, 0);  // 32..35 r0 c4567
+                        acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q4_4567_hi, q8_blk1, 1);  // 32..35 r1 c4567
+                        acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q4_4567_hi, q8_blk1, 2);  // 32..35 r2 c4567
+                        acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q4_4567_hi, q8_blk1, 3);  // 32..35 r3 c4567
+                    }
+
+                    // Scale and bias application
+                    // acc is stored interleaved to match output layout
+                    const int16x4_t sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                    const int16x4_t sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                    const int16x4_t sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                    const int16x4_t sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                    for (int row = 0; row < q8_k_blocklen; row++) {
+                        // Bias correction
+                        // row c0123 blk0 and blk1
+                        const float32x4_t sumf_0123 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[row]),
+                                                    vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[row])));
+                        acc_f32[2 * row] = vfmaq_f32(acc_f32[2 * row], sbd_scale_0123[row], sumf_0123);
+
+                        // row c4567 blk0 and blk1
+                        const float32x4_t sumf_4567 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[row + 4]),
+                                                    vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[row + 4])));
+                        acc_f32[2 * row + 1] = vfmaq_f32(acc_f32[2 * row + 1], sbd_scale_4567[row], sumf_4567);
+
+                        // Bias
+                        const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][row * 2]);
+                        const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][row * 2 + 1]);
+
+                        // row c0123 blk0 and blk1
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                        // row c4567 blk0 and blk1
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    acc_f32[2 * row] = vmlsq_f32(acc_f32[2 * row], vcvtq_f32_s32(bias_acc[2 * row]), sbd_min_0123[row]);
+                    acc_f32[2 * row + 1] =
+                        vmlsq_f32(acc_f32[2 * row + 1], vcvtq_f32_s32(bias_acc[2 * row + 1]), sbd_min_4567[row]);
+                }
+            }  // for b
+
+            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_DOTPROD)
+    ggml_gemm_q4_K_8x4_q8_K_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/ggml-cpu.c

@@ -1927,6 +1927,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_argsort(params, tensor);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                ggml_compute_forward_top_k(params, tensor);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
@@ -2311,6 +2315,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
+        case GGML_OP_TOP_K:
         case GGML_OP_FLASH_ATTN_EXT:
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
@@ -2834,6 +2839,10 @@ struct ggml_cplan ggml_graph_plan(
                         cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
                         cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
                     } break;
+                case GGML_OP_TOP_K:
+                    {
+                        cur += sizeof(int32_t)*node->src[0]->ne[0]*n_tasks;
+                    } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
                         const int64_t ne10 = node->src[1]->ne[0]; // DK

ggml/src/ggml-cpu/ops.cpp

@@ -7794,7 +7794,7 @@ void ggml_compute_forward_timestep_embedding(
 // ggml_compute_forward_argsort
 
 template<enum ggml_sort_order order>
-struct argsort_cmp {
+struct cmp_argsort {
     const float * data;
     bool operator()(int32_t a, int32_t b) const {
         if constexpr (order == GGML_SORT_ORDER_ASC) {
@@ -7833,11 +7833,11 @@ static void ggml_compute_forward_argsort_f32(
 
         switch (order) {
             case GGML_SORT_ORDER_ASC:
-                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_ASC>{src_data});
+                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_ASC>{src_data});
                 break;
 
             case GGML_SORT_ORDER_DESC:
-                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_DESC>{src_data});
+                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_DESC>{src_data});
                 break;
 
             default:
@@ -7864,6 +7864,72 @@ void ggml_compute_forward_argsort(
     }
 }
 
+// ggml_compute_forward_top_k
+
+struct cmp_top_k {
+    const float * data;
+    bool operator()(int32_t a, int32_t b) const {
+        return data[a] > data[b];
+    }
+};
+
+static void ggml_compute_forward_top_k_f32(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    const int top_k = ne0;
+
+    int32_t * tmp = (int32_t *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
+
+        for (int64_t j = 0; j < ne00; j++) {
+            tmp[j] = j;
+        }
+
+        std::partial_sort(tmp, tmp + top_k, tmp + ne00, cmp_top_k{src_data});
+
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+
+        std::copy(tmp, tmp + top_k, dst_data);
+
+        // emphasize that the order is not important
+        if (top_k > 1) {
+            std::swap(dst_data[0], dst_data[1]);
+        }
+    }
+}
+
+void ggml_compute_forward_top_k(
+    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_top_k_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_flash_attn_ext
 
 static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(

ggml/src/ggml-cpu/ops.h

@@ -81,6 +81,7 @@ void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 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_top_k(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);

ggml/src/ggml-cpu/repack.cpp

@@ -124,6 +124,58 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
     }
 }
 
+
+void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK_K == 256);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK_K];
+    float iscale[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+            float max = 0;
+
+            for (int j = 0; j < QK_K; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
+                // Update the maximum value of the corresponding super block
+                if(amax < fabsf(srcv[row_iter][j])) {
+                    amax = fabsf(srcv[row_iter][j]);
+                    max = srcv[row_iter][j];
+                }
+            }
+
+            iscale[row_iter] = amax ? -127.f/max : 0;
+
+            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
+        }
+
+        for (int j = 0; j < QK_K / 4; j++) {
+            y[i].bsums[j] = 0;
+        }
+
+        // Quants values are interleaved in sequence of four bytes from corresponding super blocks
+        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
+        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
+        for (int j = 0; j < QK_K * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id     = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+            int index = (((j & 15) >> 2) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
+
+            float x0 = srcv[src_id][src_offset] * iscale[src_id];
+            y[i].qs[j] = nearest_int(x0);
+            y[i].bsums[index] += y[i].qs[j];
+        }
+    }
+}
+
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK_K == 256);
     assert(k % QK_K == 0);
@@ -192,6 +244,12 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTR
     ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row);
 }
 
+template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_K_4x4(x, vy, n_per_row);
+}
+
 template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
     assert(nrow == 4);
     UNUSED(nrow);
@@ -333,6 +391,77 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemv_q4_K_8x4_q8_K_generic(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 = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    float sum_minf[8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
 void ggml_gemv_q4_K_8x8_q8_K_generic(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 = QK_K;
     const int nb = n / qk;
@@ -727,6 +856,89 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemm_q4_K_8x4_q8_K_generic(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 = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2] = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
 void ggml_gemm_q4_K_8x8_q8_K_generic(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 = QK_K;
     const int nb = n / qk;
@@ -1228,9 +1440,10 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
 
     GGML_UNUSED(data_size);
 }
+
 static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 8 || interleave_block == 4);
     constexpr int nrows_interleaved = 8;
 
     block_q4_Kx8 * dst = (block_q4_Kx8*)t->data;
@@ -1468,6 +1681,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q4_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_K_to_q4_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
@@ -1501,6 +1718,10 @@ template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1529,6 +1750,10 @@ template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -1731,12 +1956,13 @@ 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 < nth || disable_chunking) {
+        int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
+        // Only increase nchunk0 to nth if it won't make chunks too small
+        if (nth == 1 || ((nchunk0 < nth || disable_chunking) && (nr0 + nth - 1) / nth >= min_chunk_size)) {
             nchunk0 = nth;
+            dr0 = (nr0 + nchunk0 - 1) / nchunk0;
         }
 
-        const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
-
         // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
         // This prevents creating too many tiny chunks that could overlap after alignment
         const int64_t max_nchunk = (nr0 + min_chunk_size - 1) / min_chunk_size;
@@ -1930,6 +2156,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+
+    // instance for Q4_K
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
     // instance for Q2
@@ -1961,6 +2190,16 @@ 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;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x4_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/repack.h

@@ -80,10 +80,12 @@ extern "C" {
 
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_gemv_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);
 void ggml_gemv_q4_0_4x8_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);
 void ggml_gemv_q4_0_8x8_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);
+void ggml_gemv_q4_K_8x4_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);
 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);
 void ggml_gemv_q2_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);
 void ggml_gemv_iq4_nl_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);
@@ -91,6 +93,7 @@ void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 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);
 void ggml_gemm_q4_0_4x8_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);
 void ggml_gemm_q4_0_8x8_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);
+void ggml_gemm_q4_K_8x4_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);
 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);
 void ggml_gemm_q2_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);
 void ggml_gemm_iq4_nl_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);
@@ -99,10 +102,12 @@ void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -110,6 +115,7 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);

ggml/src/ggml-cpu/vec.h

@@ -397,119 +397,118 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 }
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, 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 = 8 * 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 = 8 * ggml_f16_epr;
 
-        GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+    GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
 
-        const int np= (n & ~(ggml_f16_step - 1));
+    int np = (n & ~(ggml_f16_step - 1));
 
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
+    svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+    svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+    for (int i = 0; i < np; i += ggml_f16_step) {
+        ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
+        GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
 
-            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
+        ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
+        GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
 
-            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
+        ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
+        ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+        ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
+        GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
 
-            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
+        ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
+        ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+        ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
+        GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
 
-            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
+        ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
+        ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+        ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
+        GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
 
-            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
+        ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
+        ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+        ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
+        GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
 
-            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
+        ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
+        ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+        ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
+        GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
 
-            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
+        ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
+        ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+        ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+        GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+    }
+    const int np2 = (n & ~(ggml_f16_epr - 1));
+    for (int k = np; k < np2; k += ggml_f16_epr) {
+        svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
+        svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+        ry = GGML_F16x_VEC_FMA(ry, rx, vx);
+
+        GGML_F16x_VEC_STORE(y + k, ry, 0);
+    }
+
+    if (np2 < n) {
+        svbool_t pg = svwhilelt_b16(np2, n);
+        svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
+        svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+        hy = svmad_f16_x(pg, hx, vx, hy);
+        svst1_f16(pg, (__fp16 *)(y + np2), hy);
+    }
+    np = n;
+#elif defined(__riscv_zvfh) // implies __riscv_v_intrinsic
+    const int np = n;
+    _Float16 hv = (_Float16)v;
+    for (int i = 0, avl; i < n; i += avl) {
+        avl = __riscv_vsetvl_e16m8(n - i);
+        vfloat16m8_t ax = __riscv_vle16_v_f16m8((const _Float16 *)&x[i], avl);
+        vfloat16m8_t ay = __riscv_vle16_v_f16m8((_Float16 *)&y[i], avl);
+        vfloat16m8_t ny = __riscv_vfmadd_vf_f16m8(ax, hv, ay, avl);
+        __riscv_vse16_v_f16m8((_Float16 *)&y[i], ny, avl);
+    }
+#elif defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    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++) {
+            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        const int np2 = (n & ~(ggml_f16_epr - 1));
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-            ry = GGML_F16x_VEC_FMA(ry, rx, vx);
-
-            GGML_F16x_VEC_STORE(y + k, ry, 0);
-        }
-
-        if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
-            hy = svmad_f16_x(pg, hx, vx, hy);
-            svst1_f16(pg, (__fp16 *)(y + np2), hy);
-        }
-
-    #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]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
-
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ax[GGML_F16_ARR];
-        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++) {
-                ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_FMA(ay[j], ax[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]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #endif
+    }
 #else
-    // scalar
-    for (int i = 0; i < n; ++i) {
+    const int np = 0;
+#endif
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
-#endif
 }
 
 // xs and vs are byte strides of x and v

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);
     }
 }
@@ -399,94 +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

@@ -4115,6 +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 +436,30 @@ 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 (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
+
+        } else if constexpr (std::is_same_v<T, int>) {
+            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 {
+            NO_DEVICE_CODE;
+        }
 #else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
@@ -639,12 +746,34 @@ 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 +794,36 @@ 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 +850,7 @@ namespace ggml_cuda_mma {
                                                      acc[0],
                                                      0, 0, 0);
 #endif // defined(CDNA3)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -735,4 +895,31 @@ 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-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-metal/ggml-metal-device.cpp

@@ -1009,6 +1009,64 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge(ggml_metal_l
     return res;
 }
 
+// note: reuse the argsort kernel for top_k
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    // note: the top_k kernel is always descending order
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    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_%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_top_k_merge(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    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,

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

@@ -128,6 +128,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id         (ggml_me
 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_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_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);

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

@@ -905,6 +905,7 @@ 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:
+        case GGML_OP_TOP_K:
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:

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

@@ -832,14 +832,19 @@ typedef struct {
 } ggml_metal_kargs_leaky_relu;
 
 typedef struct {
-    int64_t  ne00;
-    int64_t  ne01;
-    int64_t  ne02;
-    int64_t  ne03;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
     uint64_t nb00;
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
 } ggml_metal_kargs_argsort;
 
 typedef struct {
@@ -851,6 +856,11 @@ typedef struct {
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
     int32_t  len;
 } ggml_metal_kargs_argsort_merge;
 

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

@@ -11,6 +11,7 @@
 #include <cassert>
 #include <algorithm>
 #include <limits>
+#include <cmath>
 
 static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
     if (!t) {
@@ -405,6 +406,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_argsort(ctx, idx);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                n_fuse = ggml_metal_op_top_k(ctx, idx);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
@@ -3677,14 +3682,19 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     }
 
     ggml_metal_kargs_argsort args = {
-        /*.ne00 =*/ ne00,
-        /*.ne01 =*/ ne01,
-        /*.ne02 =*/ ne02,
-        /*.ne03 =*/ ne03,
-        /*.nb00 =*/ nb00,
-        /*.nb01 =*/ nb01,
-        /*.nb02 =*/ nb02,
-        /*.nb03 =*/ nb03,
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ nth,
     };
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
@@ -3704,15 +3714,20 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
         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,
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ ne00,
+            /*.len   =*/ len,
         };
 
         // merges per row
@@ -3736,6 +3751,118 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_top_k(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 = ggml_metal_library_get_pipeline_top_k(lib, op);
+
+    // 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;
+    }
+
+    // blocks per row
+    const int npr = (ne00 + nth - 1)/nth;
+
+    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 += sizeof(int32_t)*ggml_nelements(op->src[0]);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
+
+    const int top_k = ne0;
+
+    ggml_metal_kargs_argsort args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ std::min(nth, top_k), // for each block, keep just the top_k indices
+    };
+
+    if (npr > 1) {
+        args.ne0 = (npr - 1)*args.top_k + std::min(ne00 - (npr - 1)*nth, args.top_k);
+    }
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    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_dst,  2);
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_top_k_merge(lib, op);
+
+    int len = args.top_k;
+
+    while (len < args.ne0) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        // merges per row
+        const int nm = (args.ne0 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ args.ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ nm == 1 ? top_k : args.ne0, // the final merge outputs top_k elements
+            /*.len   =*/ len,
+        };
+
+        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;
+}
+
 int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

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

@@ -81,6 +81,7 @@ int ggml_metal_op_arange            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argmax            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argsort           (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_top_k             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_leaky_relu        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_adamw    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_sgd      (ggml_metal_op_t ctx, int idx);

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

@@ -202,6 +202,10 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
             {
                 res *= 2;
             } break;
+        case GGML_OP_TOP_K:
+            {
+                res = 2*sizeof(int32_t)*ggml_nelements(tensor->src[0]);
+            } break;
         default:
             break;
     }

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

@@ -4670,11 +4670,12 @@ kernel void kernel_argsort_f32_i32(
         ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
     const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
 
-    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];
+    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);
 
@@ -4710,9 +4711,11 @@ kernel void kernel_argsort_f32_i32(
         }
     }
 
+    const int64_t i0 = ib*args.top_k;
+
     // copy the result to dst without the padding
-    if (i00 + col < args.ne00) {
-        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
 
         dst[col] = shmem_i32[col];
     }
@@ -4747,22 +4750,22 @@ kernel void kernel_argsort_merge_f32_i32(
 
     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 len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (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;
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*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;
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
 
     device const float * src0_row = (device const float *)(src0
         + args.nb01*i01
@@ -4776,7 +4779,11 @@ kernel void kernel_argsort_merge_f32_i32(
     const int chunk = (total + ntg.x - 1) / ntg.x;
 
     const int k0 = tpitg.x * chunk;
-    const int k1 = min(k0 + chunk, total);
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
 
     if (k0 >= total) {
         return;

ggml/src/ggml-opencl/CMakeLists.txt

@@ -70,6 +70,7 @@ set(GGML_OPENCL_KERNELS
     group_norm
     im2col_f32
     im2col_f16
+    mean
     mul_mat_Ab_Bi_8x4
     mul_mv_f16_f16
     mul_mv_f16_f32_1row
@@ -109,6 +110,9 @@ set(GGML_OPENCL_KERNELS
     softmax_4_f16
     softmax_f32
     softmax_f16
+    sqr
+    sqrt
+    ssm_conv
     sub
     sum_rows
     transpose

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

@@ -449,6 +449,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
     cl_kernel kernel_add_id;
     cl_kernel kernel_scale;
+    cl_kernel kernel_sqr_cont_f32, kernel_sqr_cont_f32_4, kernel_sqr_cont_f16, kernel_sqr_cont_f16_4;
+    cl_kernel kernel_sqrt_cont_f32, kernel_sqrt_cont_f32_4, kernel_sqrt_cont_f16, kernel_sqrt_cont_f16_4;
+    cl_kernel kernel_mean_f32;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
     cl_kernel kernel_gelu_erf, kernel_gelu_erf_4;
@@ -509,6 +512,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_conv_2d_f16;
     cl_kernel kernel_conv_2d_f32;
     cl_kernel kernel_conv_2d_f16_f32;
+    cl_kernel kernel_ssm_conv_f32_f32, kernel_ssm_conv_f32_f32_4;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_gemv_moe_mxfp4_f32, kernel_gemm_moe_mxfp4_f32;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
@@ -1552,6 +1556,66 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // sqr
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sqr.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sqr.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f32     = clCreateKernel(prog, "kernel_sqr_cont_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f32_4   = clCreateKernel(prog, "kernel_sqr_cont_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f16     = clCreateKernel(prog, "kernel_sqr_cont_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f16_4   = clCreateKernel(prog, "kernel_sqr_cont_f16_4", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // sqrt
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sqrt.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sqrt.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f32     = clCreateKernel(prog, "kernel_sqrt_cont_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f32_4   = clCreateKernel(prog, "kernel_sqrt_cont_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f16     = clCreateKernel(prog, "kernel_sqrt_cont_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f16_4   = clCreateKernel(prog, "kernel_sqrt_cont_f16_4", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mean
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mean.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mean.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mean_f32 = clCreateKernel(prog, "kernel_mean_f32", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // sub
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1825,6 +1889,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
                 }
     }
 
+    // ssm_conv
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "ssm_conv.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("ssm_conv.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_ssm_conv_f32_f32   = clCreateKernel(prog, "kernel_ssm_conv_f32_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_ssm_conv_f32_f32_4 = clCreateKernel(prog, "kernel_ssm_conv_f32_f32_4", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_id_q4_0_f32_8x_flat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2959,6 +3041,10 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                    (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
         case GGML_OP_ADD_ID:
             return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                    ggml_is_contiguous(op->src[0]);
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
@@ -3007,6 +3093,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16) ||
                    (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
                    (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_SSM_CONV:
+            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
         case GGML_OP_CONCAT:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3075,6 +3163,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return cols <= max_workgroup_size && op->src[0]->type == GGML_TYPE_F32;
         }
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
         case GGML_OP_FLASH_ATTN_EXT:
             {
@@ -5193,6 +5282,224 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
     }
 }
 
+static void ggml_cl_sqr(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    UNUSED(src1);
+
+    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 * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    // Currently assumes src0 is contiguous
+    int n = ggml_nelements(dst);
+    if (n % 4 == 0) {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqr_cont_f32_4;
+        } else {
+            kernel = backend_ctx->kernel_sqr_cont_f16_4;
+        }
+        n /= 4;
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqr_cont_f32;
+        } else {
+            kernel = backend_ctx->kernel_sqr_cont_f16;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_sqrt(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    UNUSED(src1);
+
+    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 * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    // Currently assumes src0 is contiguous
+    int n = ggml_nelements(dst);
+    if (n % 4 == 0) {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqrt_cont_f32_4;
+        } else {
+            kernel = backend_ctx->kernel_sqrt_cont_f16_4;
+        }
+        n /= 4;
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqrt_cont_f32;
+        } else {
+            kernel = backend_ctx->kernel_sqrt_cont_f16;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_mean(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_UNUSED(src1);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    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 * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_mean_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb3));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_ssm_conv(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * 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;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    int ne01 = src0->ne[1];
+    cl_ulong nb00 = src0->nb[0];
+    cl_ulong nb01 = src0->nb[1];
+    cl_ulong nb02 = src0->nb[2];
+
+    int ne10 = src1->ne[0];
+    cl_ulong nb11 = src1->nb[1];
+
+    int ne1  = dst->ne[1];
+    int ne2  = dst->ne[2];
+    cl_ulong nb0 = dst->nb[0];
+    cl_ulong nb1 = dst->nb[1];
+    cl_ulong nb2 = dst->nb[2];
+
+    cl_kernel kernel = backend_ctx->kernel_ssm_conv_f32_f32;
+
+    if (ne10 % 4 == 0) {
+        kernel = backend_ctx->kernel_ssm_conv_f32_f32_4;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb2));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne1, (size_t)ne2};
+    size_t local_work_size[]  = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (ne01 % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
 static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -6895,9 +7202,23 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     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){
-            ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
-            return;
+        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;
+            }
         }
     }
 
@@ -9077,6 +9398,24 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_sub;
             break;
+        case GGML_OP_SQR:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sqr;
+            break;
+        case GGML_OP_SQRT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sqrt;
+            break;
+        case GGML_OP_MEAN:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mean;
+            break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(tensor)) {
                 case GGML_UNARY_OP_GELU:
@@ -9178,6 +9517,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_conv_2d;
             break;
+        case GGML_OP_SSM_CONV:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_ssm_conv;
+            break;
         case GGML_OP_CONCAT:
             if (!any_on_device) {
                 return false;

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

@@ -0,0 +1,39 @@
+
+kernel void kernel_mean_f32(
+    global float *  src0,
+    ulong           offset0,
+    global float *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int i3 = get_global_id(2);
+    int i2 = get_global_id(1);
+    int i1 = get_global_id(0);
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float * dst_row = (global float *) ((global char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum / ne00;
+}

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

@@ -0,0 +1,53 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_sqr_cont_f32(
+    global float * src0,
+    ulong          offset0,
+    global float * dst,
+    ulong          offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f32_4(
+    global float4 * src0,
+    ulong           offset0,
+    global float4 * dst,
+    ulong           offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f16(
+    global half * src0,
+    ulong         offset0,
+    global half * dst,
+    ulong         offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f16_4(
+    global half4 * src0,
+    ulong          offset0,
+    global half4 * dst,
+    ulong          offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}

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

@@ -0,0 +1,53 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_sqrt_cont_f32(
+    global float * src0,
+    ulong          offset0,
+    global float * dst,
+    ulong          offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = sqrt(src0[gid]);
+}
+
+kernel void kernel_sqrt_cont_f32_4(
+    global float4 * src0,
+    ulong           offset0,
+    global float4 * dst,
+    ulong           offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = sqrt(src0[gid]);
+}
+
+kernel void kernel_sqrt_cont_f16(
+    global half * src0,
+    ulong         offset0,
+    global half * dst,
+    ulong         offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = convert_half(sqrt(convert_float(src0[gid])));
+}
+
+kernel void kernel_sqrt_cont_f16_4(
+    global half4 * src0,
+    ulong          offset0,
+    global half4 * dst,
+    ulong          offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = convert_half4(sqrt(convert_float4(src0[gid])));
+}

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

@@ -0,0 +1,77 @@
+kernel void kernel_ssm_conv_f32_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb00,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne10,
+    ulong         nb11,
+    ulong         nb0,
+    ulong         nb1,
+    ulong         nb2
+){
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ir = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    int nc  = ne10;
+
+    global float * s = (global float *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    global float * c = (global float *) (src1 + ir*nb11);
+    global float * d = (global float *) (dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    d[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb00,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne10,
+    ulong         nb11,
+    ulong         nb0,
+    ulong         nb1,
+    ulong         nb2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ir = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    int nc = ne10;
+
+    global float4 * s = (global float4 *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    global float4 * c = (global float4 *) (src1 + ir*nb11);
+    global float  * d = (global float  *) (dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    d[0] = sumf;
+}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -409,6 +409,7 @@ enum shader_reduction_mode {
 // argsort pipelines for up to 1<<10 invocations per workgroup
 static constexpr uint32_t num_argsort_pipelines = 11;
 static constexpr uint32_t num_topk_moe_pipelines = 10;
+static constexpr uint32_t num_topk_pipelines = 11;
 
 static constexpr std::initializer_list<ggml_op> topk_moe_early_softmax_norm{ GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
                                                                              GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
@@ -513,7 +514,9 @@ struct vk_device_struct {
     vk_queue compute_queue;
     vk_queue transfer_queue;
     bool single_queue;
+    bool support_async;
     uint32_t subgroup_size;
+    uint32_t subgroup_size_log2;
     uint32_t shader_core_count;
     bool uma;
     bool prefer_host_memory;
@@ -703,7 +706,9 @@ struct vk_device_struct {
     vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
     vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
     vk_pipeline pipeline_argsort_large_f32[num_argsort_pipelines];
+    vk_pipeline pipeline_topk_f32[num_topk_pipelines];
     vk_pipeline pipeline_sum_rows_f32;
+    vk_pipeline pipeline_cumsum_f32;
     vk_pipeline pipeline_argmax_f32;
     vk_pipeline pipeline_count_equal_i32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
@@ -1203,6 +1208,15 @@ struct vk_op_argsort_push_constants {
     uint32_t inner_end;
 };
 
+struct vk_op_topk_push_constants {
+    uint32_t orig_ncols;
+    uint32_t ncols_input;
+    uint32_t ncols_output;
+    uint32_t nrows;
+    uint32_t first_pass;
+    uint32_t last_pass;
+};
+
 struct vk_op_im2col_push_constants {
     uint64_t dst_addr;
     uint32_t batch_offset; uint32_t offset_delta;
@@ -1628,6 +1642,22 @@ class vk_perf_logger {
             timings[name].push_back(time);
             return;
         }
+        if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+            const ggml_tensor * dst = node;
+            const ggml_tensor * q = node->src[0];
+            const ggml_tensor * k = node->src[1];
+            const ggml_tensor * v = node->src[2];
+            const ggml_tensor * m = node->src[3];
+            std::stringstream name;
+            name << ggml_op_name(node->op) <<
+                " dst(" << dst->ne[0] << "," << dst->ne[1] << "," << dst->ne[2] << "," << dst->ne[3] << "), " <<
+                " q(" << q->ne[0] << "," << q->ne[1] << "," << q->ne[2] << "," << q->ne[3] << "), " <<
+                " k(" << k->ne[0] << "," << k->ne[1] << "," << k->ne[2] << "," << k->ne[3] << "), " <<
+                " v(" << v->ne[0] << "," << v->ne[1] << "," << v->ne[2] << "," << v->ne[3] << "), " <<
+                " m(" << (m?m->ne[0]:0) << "," << (m?m->ne[1]:0) << "," << (m?m->ne[2]:0) << "," << (m?m->ne[3]:0) << ")";
+            timings[name.str()].push_back(time);
+            return;
+        }
         timings[ggml_op_name(node->op)].push_back(time);
     }
   private:
@@ -2484,9 +2514,11 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
 
-static uint32_t get_fa_scalar_num_large_rows(uint32_t hsv) {
+static uint32_t get_fa_scalar_num_large_rows(uint32_t hsk, uint32_t hsv) {
     if (hsv >= 192) {
         return 2;
+    } else if ((hsv | hsk) & 8) {
+        return 4;
     } else {
         return 8;
     }
@@ -2518,9 +2550,9 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
             if ((hsv | hsk) & 8) {
                 // HSV/HSK not being a multiple of 16 makes D_split smaller, which makes cols_per_iter
                 // larger, and Bc needs to be >= cols_per_thread. 64 is large enough, 32 is not.
-                return {get_fa_scalar_num_large_rows(hsv), 64};
+                return {get_fa_scalar_num_large_rows(hsk, hsv), 64};
             } else {
-                return {get_fa_scalar_num_large_rows(hsv), 32};
+                return {get_fa_scalar_num_large_rows(hsk, hsv), 32};
             }
         }
     }
@@ -3945,10 +3977,29 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline2(device, device->pipeline_argsort_large_f32[i], "argsort_large_f32_"+std::to_string(i), argsort_large_f32_len, argsort_large_f32_data, "main", 3, sizeof(vk_op_argsort_push_constants), {BLOCK_SIZE * WG_UNROLL_FACTOR, 1, 1}, {BLOCK_SIZE, WG_UNROLL_FACTOR}, 1, true);
     }
 
+    for (uint32_t i = 0; i < num_topk_pipelines; ++i) {
+        const uint32_t BLOCK_SIZE = 1u << i;
+        const uint32_t NCOLS_PADDED_LOG2 = i;
+        if (i <= device->max_workgroup_size_log2) {
+            uint32_t nary_shmem = 2 * sizeof(int) * BLOCK_SIZE +
+                                  sizeof(int) * device->subgroup_size +
+                                  2 * sizeof(int) +
+                                  (BLOCK_SIZE / device->subgroup_size) * sizeof(int);
+            if (device->subgroup_arithmetic && device->subgroup_require_full_support && device->subgroup_shuffle && device->subgroup_ballot &&
+                nary_shmem <= device->properties.limits.maxComputeSharedMemorySize) {
+                ggml_vk_create_pipeline2(device, device->pipeline_topk_f32[i], "topk_f32_"+std::to_string(i), topk_nary_search_f32_len, topk_nary_search_f32_data, "main", 2, sizeof(vk_op_topk_push_constants), {BLOCK_SIZE, 1, 1}, {BLOCK_SIZE, device->subgroup_size, device->subgroup_size_log2}, 1, true, true, device->subgroup_size);
+            } else if (2 * sizeof(int) * BLOCK_SIZE <= device->properties.limits.maxComputeSharedMemorySize) {
+                ggml_vk_create_pipeline2(device, device->pipeline_topk_f32[i], "topk_f32_"+std::to_string(i), topk_argsort_f32_len, topk_argsort_f32_data, "main", 2, sizeof(vk_op_topk_push_constants), {BLOCK_SIZE, 1, 1}, {BLOCK_SIZE, NCOLS_PADDED_LOG2}, 1, true);
+            }
+        }
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 128, device->subgroup_size }, 1, true, true, device->subgroup_size);
+
     ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
 
 #define IM2COL(bda) \
@@ -4273,6 +4324,16 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->vendor_id = device->properties.vendorID;
         device->driver_id = driver_props.driverID;
 
+        // Implementing the async backend interfaces seems broken on older Intel HW,
+        // see https://github.com/ggml-org/llama.cpp/issues/17302.
+        device->support_async = (device->vendor_id != VK_VENDOR_ID_INTEL ||
+                                 std::string(device->properties.deviceName.data()).find("(DG1)") == std::string::npos) &&
+                                getenv("GGML_VK_DISABLE_ASYNC") == nullptr;
+
+        if (!device->support_async) {
+            GGML_LOG_DEBUG("ggml_vulkan: WARNING: Async execution disabled on certain Intel devices.\n");
+        }
+
         const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
 
         if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) {
@@ -4304,6 +4365,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size);
 
         device->subgroup_size = subgroup_props.subgroupSize;
+        device->subgroup_size_log2 = uint32_t(log2f(float(device->subgroup_size)));
         device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
         if (sm_builtins) {
             device->shader_core_count = sm_props.shaderSMCount;
@@ -5227,7 +5289,8 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->prealloc_size_x = 0;
     ctx->prealloc_size_y = 0;
     ctx->prealloc_size_split_k = 0;
-    ctx->prealloc_size_add_rms_partials = 0;
+    // Fixed size of 1KB, for deterministic behavior
+    ctx->prealloc_size_add_rms_partials = 1024;
 
     ctx->fence = ctx->device->device.createFence({});
     ctx->almost_ready_fence = ctx->device->device.createFence({});
@@ -7713,7 +7776,7 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
     // Needs to be kept up to date on shader changes
     GGML_UNUSED(hsv);
     const uint32_t wg_size = scalar_flash_attention_workgroup_size;
-    const uint32_t Br = get_fa_scalar_num_large_rows(hsv);
+    const uint32_t Br = get_fa_scalar_num_large_rows(hsk, hsv);
     const uint32_t Bc = scalar_flash_attention_Bc;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
@@ -7844,7 +7907,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     case FA_SCALAR:
     case FA_COOPMAT1:
         // We may switch from coopmat1 to scalar, so use the scalar limit for both
-        max_gqa = get_fa_scalar_num_large_rows(HSV);
+        max_gqa = get_fa_scalar_num_large_rows(HSK, HSV);
         break;
     case FA_COOPMAT2:
         max_gqa = get_fa_num_small_rows(FA_COOPMAT2);
@@ -8428,6 +8491,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_sum_rows_f32;
         }
         return nullptr;
+    case GGML_OP_CUMSUM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_cumsum_f32;
+        }
+        return nullptr;
     case GGML_OP_ARGMAX:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
             return ctx->device->pipeline_argmax_f32;
@@ -8619,41 +8687,6 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     GGML_UNUSED(src2);
 }
 
-static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
-    switch (op) {
-    case GGML_OP_CPY:
-    case GGML_OP_GET_ROWS:
-    case GGML_OP_ADD:
-    case GGML_OP_SUB:
-    case GGML_OP_MUL:
-    case GGML_OP_DIV:
-    case GGML_OP_ADD_ID:
-    case GGML_OP_CONCAT:
-    case GGML_OP_UPSCALE:
-    case GGML_OP_SQR:
-    case GGML_OP_SQRT:
-    case GGML_OP_SIN:
-    case GGML_OP_COS:
-    case GGML_OP_LOG:
-    case GGML_OP_CLAMP:
-    case GGML_OP_PAD:
-    case GGML_OP_REPEAT:
-    case GGML_OP_REPEAT_BACK:
-    case GGML_OP_ROPE:
-    case GGML_OP_RMS_NORM:
-    case GGML_OP_CONV_2D_DW:
-    case GGML_OP_IM2COL:
-    case GGML_OP_IM2COL_3D:
-    case GGML_OP_SET_ROWS:
-    case GGML_OP_SUM:
-    case GGML_OP_SUM_ROWS:
-    case GGML_OP_MEAN:
-        return true;
-    default:
-        return false;
-    }
-}
-
 template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
@@ -8738,7 +8771,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << ggml_op_name(op) << ")");
     GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
-    GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0));  // NOLINT
     GGML_ASSERT(dst->buffer != nullptr);
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
@@ -8769,22 +8801,17 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
 
     ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
-    const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op);
-
-    vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0, op_supports_incontiguous);
-    vk_subbuffer src1_buf = use_src1 ? ggml_vk_tensor_subbuffer(ctx, src1, op_supports_incontiguous) : vk_subbuffer{};
-    vk_subbuffer src2_buf = use_src2 ? ggml_vk_tensor_subbuffer(ctx, src2, op_supports_incontiguous) : vk_subbuffer{};
-    vk_subbuffer src3_buf = use_src3 ? ggml_vk_tensor_subbuffer(ctx, src3, op_supports_incontiguous) : vk_subbuffer{};
-    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, op_supports_incontiguous);
+    vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0, true);
+    vk_subbuffer src1_buf = use_src1 ? ggml_vk_tensor_subbuffer(ctx, src1, true) : vk_subbuffer{};
+    vk_subbuffer src2_buf = use_src2 ? ggml_vk_tensor_subbuffer(ctx, src2, true) : vk_subbuffer{};
+    vk_subbuffer src3_buf = use_src3 ? ggml_vk_tensor_subbuffer(ctx, src3, true) : vk_subbuffer{};
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, true);
 
     // Compute misalignment offset for descriptors and store it in in push constants.
     init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, src3, dst);
 
     std::array<uint32_t, 3> elements;
 
-    // Single call if dimension 2 is contiguous
-    GGML_ASSERT(op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1))));
-
     switch (op) {
     case GGML_OP_NORM:
     case GGML_OP_RMS_NORM_BACK:
@@ -8792,6 +8819,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_SOFT_MAX:
     case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_CUMSUM:
     case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
         {
@@ -10105,6 +10133,104 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
     }
 }
 
+static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    uint32_t ncols = src0->ne[0];
+    uint32_t nrows = ggml_nrows(src0);
+    uint32_t k = dst->ne[0];
+
+    vk_op_topk_push_constants pc { ncols, ncols, k, nrows, 0, 0 };
+
+    // Reserve space for ivec2 per element, double buffered
+    const size_t dbl_buf_size = size_t{ncols} * nrows * 2 * sizeof(int);
+    const size_t x_sz = dbl_buf_size * 2;
+    uint32_t dbl_buf_index = 0;
+
+    if (ctx->prealloc_size_x < x_sz) {
+        ctx->prealloc_size_x = x_sz;
+        ggml_vk_preallocate_buffers(ctx, subctx);
+    }
+    if (ctx->prealloc_x_need_sync) {
+        ggml_vk_sync_buffers(ctx, subctx);
+    }
+
+    std::array<uint32_t, 3> elements;
+    elements[1] = std::min(nrows, ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+    elements[2] = 1;
+
+    uint32_t num_elements = ncols;
+
+    // Each iteration reduces a workgroup's worth of elements down to the K
+    // largest elements. Repeat until we have the top K elements.
+    // Need to do at least one iteration to write out the results.
+    bool done_one_iter = false;
+    while (num_elements > k || !done_one_iter) {
+        done_one_iter = true;
+
+        // Prefer going as small as num_topk_pipelines - 3 for perf reasons.
+        // But if K is larger, then we need a larger workgroup
+        uint32_t max_pipeline = num_topk_pipelines - 3;
+        uint32_t min_pipeline = (uint32_t)log2f(float(k)) + 1;
+        // require full subgroup
+        min_pipeline = std::max(min_pipeline, ctx->device->subgroup_size_log2);
+
+        uint32_t pipeline_idx = (uint32_t)ceilf(log2f(float(num_elements)));
+        pipeline_idx = std::min(pipeline_idx, max_pipeline);
+        pipeline_idx = std::max(pipeline_idx, min_pipeline);
+
+        if (num_elements > (1u << pipeline_idx)) {
+            // If we could finish on this loop iteration (i.e. a single workgroup)
+            // then do so. It's better than the overhead of another pass.
+            for (uint32_t i = pipeline_idx; i < num_topk_pipelines; ++i) {
+                if (num_elements <= (1u << i)) {
+                    pipeline_idx = i;
+                    break;
+                }
+            }
+        }
+
+        vk_pipeline pipeline = ctx->device->pipeline_topk_f32[pipeline_idx];
+        // If the device doesn't support a pipeline this large, use smaller
+        while (!pipeline) {
+            pipeline_idx--;
+            GGML_ASSERT(pipeline_idx >= min_pipeline);
+            pipeline = ctx->device->pipeline_topk_f32[pipeline_idx];
+        }
+
+        vk_op_topk_push_constants pc2 = pc;
+        pc2.ncols_input = num_elements;
+
+        // Number of elements remaining after this pass
+        uint32_t num_dst_elements = (num_elements / pipeline->wg_denoms[0]) * k + std::min(k, num_elements % pipeline->wg_denoms[0]);
+
+        vk_subbuffer src_buf;
+        vk_subbuffer dst_buf;
+
+        if (num_elements == ncols) {
+            pc2.first_pass = 1;
+            src_buf = ggml_vk_tensor_subbuffer(ctx, src0);
+        } else {
+            src_buf = { ctx->prealloc_x, dbl_buf_index * dbl_buf_size, dbl_buf_size };
+        }
+        if (num_dst_elements == k) {
+            pc2.last_pass = 1;
+            dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+        } else {
+            dst_buf = { ctx->prealloc_x, (dbl_buf_index ^ 1) * dbl_buf_size, dbl_buf_size };
+        }
+
+        elements[0] = num_elements;
+
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src_buf, dst_buf }, pc2, elements);
+        num_elements = num_dst_elements;
+        dbl_buf_index ^= 1;
+        if (num_elements > k) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    ctx->prealloc_x_need_sync = true;
+}
+
 static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, ggml_nelements(src0));
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM, p);
@@ -10121,6 +10247,11 @@ static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_MEAN, p);
 }
 
+static void ggml_vk_cumsum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CUMSUM, p);
+}
+
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f });
 }
@@ -11370,13 +11501,13 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
     }
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool almost_ready);
+static void ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool almost_ready);
 
 // Returns true if node has enqueued work into the queue, false otherwise
 // If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution.
 static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool last_node, bool almost_ready, bool submit){
     ggml_tensor * node = cgraph->nodes[node_idx];
-    if (ggml_is_empty(node) || !node->buffer) {
+    if (ggml_is_empty(node) || ggml_op_is_empty(node->op) || !node->buffer) {
         return false;
     }
 
@@ -11388,132 +11519,19 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     ggml_tensor * src2 = node->src[2];
     ggml_tensor * src3 = node->src[3];
 
-    switch (node->op) {
-    // Return on empty ops to avoid generating a compute_ctx and setting exit_tensor
-    case GGML_OP_RESHAPE:
-    case GGML_OP_VIEW:
-    case GGML_OP_PERMUTE:
-    case GGML_OP_TRANSPOSE:
-    case GGML_OP_NONE:
-        return false;
-    case GGML_OP_UNARY:
-        switch (ggml_get_unary_op(node)) {
-        case GGML_UNARY_OP_EXP:
-        case GGML_UNARY_OP_SILU:
-        case GGML_UNARY_OP_GELU:
-        case GGML_UNARY_OP_GELU_ERF:
-        case GGML_UNARY_OP_GELU_QUICK:
-        case GGML_UNARY_OP_RELU:
-        case GGML_UNARY_OP_NEG:
-        case GGML_UNARY_OP_TANH:
-        case GGML_UNARY_OP_SIGMOID:
-        case GGML_UNARY_OP_HARDSIGMOID:
-        case GGML_UNARY_OP_HARDSWISH:
-        case GGML_UNARY_OP_ABS:
-        case GGML_UNARY_OP_SOFTPLUS:
-        case GGML_UNARY_OP_STEP:
-        case GGML_UNARY_OP_ROUND:
-        case GGML_UNARY_OP_CEIL:
-        case GGML_UNARY_OP_FLOOR:
-        case GGML_UNARY_OP_TRUNC:
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_GLU:
-        switch (ggml_get_glu_op(node)) {
-        case GGML_GLU_OP_GEGLU:
-        case GGML_GLU_OP_REGLU:
-        case GGML_GLU_OP_SWIGLU:
-        case GGML_GLU_OP_SWIGLU_OAI:
-        case GGML_GLU_OP_GEGLU_ERF:
-        case GGML_GLU_OP_GEGLU_QUICK:
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_ADD:
-        {
-            int next_node_idx = node_idx + 1 + ctx->num_additional_fused_ops;
-            if (next_node_idx < cgraph->n_nodes &&
-                cgraph->nodes[next_node_idx]->op == GGML_OP_RMS_NORM &&
-                cgraph->nodes[next_node_idx]->src[0] == cgraph->nodes[next_node_idx - 1] &&
-                ggml_nrows(cgraph->nodes[next_node_idx]) == 1 &&
-                ctx->device->add_rms_fusion) {
-                uint32_t size = ggml_vk_rms_partials_size(ctx, cgraph->nodes[node_idx]);
-                ctx->do_add_rms_partials_offset_calculation = true;
-                if (ctx->prealloc_size_add_rms_partials_offset + size <= ctx->prealloc_size_add_rms_partials) {
-                    ctx->do_add_rms_partials = true;
-                }
+    if (node->op == GGML_OP_ADD) {
+        int next_node_idx = node_idx + 1 + ctx->num_additional_fused_ops;
+        if (next_node_idx < cgraph->n_nodes &&
+            cgraph->nodes[next_node_idx]->op == GGML_OP_RMS_NORM &&
+            cgraph->nodes[next_node_idx]->src[0] == cgraph->nodes[next_node_idx - 1] &&
+            ggml_nrows(cgraph->nodes[next_node_idx]) == 1 &&
+            ctx->device->add_rms_fusion) {
+            uint32_t size = ggml_vk_rms_partials_size(ctx, cgraph->nodes[node_idx]);
+            ctx->do_add_rms_partials_offset_calculation = true;
+            if (ctx->prealloc_size_add_rms_partials_offset + size <= ctx->prealloc_size_add_rms_partials) {
+                ctx->do_add_rms_partials = true;
             }
-        } break;
-    case GGML_OP_REPEAT:
-    case GGML_OP_REPEAT_BACK:
-    case GGML_OP_GET_ROWS:
-    case GGML_OP_ADD_ID:
-    case GGML_OP_ACC:
-    case GGML_OP_SUB:
-    case GGML_OP_MUL:
-    case GGML_OP_DIV:
-    case GGML_OP_ADD1:
-    case GGML_OP_ARANGE:
-    case GGML_OP_FILL:
-    case GGML_OP_CONCAT:
-    case GGML_OP_UPSCALE:
-    case GGML_OP_SCALE:
-    case GGML_OP_SQR:
-    case GGML_OP_SQRT:
-    case GGML_OP_SIN:
-    case GGML_OP_COS:
-    case GGML_OP_LOG:
-    case GGML_OP_CLAMP:
-    case GGML_OP_PAD:
-    case GGML_OP_ROLL:
-    case GGML_OP_CPY:
-    case GGML_OP_SET_ROWS:
-    case GGML_OP_CONT:
-    case GGML_OP_DUP:
-    case GGML_OP_SILU_BACK:
-    case GGML_OP_NORM:
-    case GGML_OP_GROUP_NORM:
-    case GGML_OP_RMS_NORM:
-    case GGML_OP_RMS_NORM_BACK:
-    case GGML_OP_L2_NORM:
-    case GGML_OP_DIAG_MASK_INF:
-    case GGML_OP_SOFT_MAX:
-    case GGML_OP_SOFT_MAX_BACK:
-    case GGML_OP_ROPE:
-    case GGML_OP_ROPE_BACK:
-    case GGML_OP_MUL_MAT:
-    case GGML_OP_MUL_MAT_ID:
-    case GGML_OP_ARGSORT:
-    case GGML_OP_SUM:
-    case GGML_OP_SUM_ROWS:
-    case GGML_OP_MEAN:
-    case GGML_OP_ARGMAX:
-    case GGML_OP_COUNT_EQUAL:
-    case GGML_OP_IM2COL:
-    case GGML_OP_IM2COL_3D:
-    case GGML_OP_TIMESTEP_EMBEDDING:
-    case GGML_OP_CONV_TRANSPOSE_1D:
-    case GGML_OP_POOL_2D:
-    case GGML_OP_CONV_2D:
-    case GGML_OP_CONV_TRANSPOSE_2D:
-    case GGML_OP_CONV_2D_DW:
-    case GGML_OP_RWKV_WKV6:
-    case GGML_OP_RWKV_WKV7:
-    case GGML_OP_SSM_SCAN:
-    case GGML_OP_SSM_CONV:
-    case GGML_OP_LEAKY_RELU:
-    case GGML_OP_FLASH_ATTN_EXT:
-    case GGML_OP_OPT_STEP_ADAMW:
-    case GGML_OP_OPT_STEP_SGD:
-        break;
-    default:
-        std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
-        GGML_ABORT("fatal error");
+        }
     }
 
     vk_context compute_ctx;
@@ -11825,6 +11843,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
             ggml_vk_argsort(ctx, compute_ctx, src0, node);
         }
 
+        break;
+    case GGML_OP_TOP_K:
+        ggml_vk_topk(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_SUM:
         ggml_vk_sum(ctx, compute_ctx, src0, node);
@@ -11833,6 +11855,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_SUM_ROWS:
         ggml_vk_sum_rows(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_CUMSUM:
+        ggml_vk_cumsum(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_MEAN:
         ggml_vk_mean(ctx, compute_ctx, src0, node);
@@ -11950,145 +11976,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         ctx->compute_ctx.reset();
 
-        bool ok = ggml_vk_compute_forward(ctx, cgraph, node_begin, node_idx_begin, almost_ready);
-        if (!ok) {
-            if (node->op == GGML_OP_UNARY) {
-                std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
-            } else if (node->op == GGML_OP_GLU) {
-                std::cerr << __func__ << ": error: op not supported GLU " << node->name << " (" << ggml_glu_op_name(static_cast<ggml_glu_op>(node->op_params[0])) << ")" << std::endl;
-            } else {
-                std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
-            }
-        }
-
+        ggml_vk_compute_forward(ctx, cgraph, node_begin, node_idx_begin, almost_ready);
     }
     return true;
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, ggml_tensor * tensor, int tensor_idx, bool almost_ready = false) {
+static void ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, ggml_tensor * tensor, int tensor_idx, bool almost_ready = false) {
     GGML_UNUSED(cgraph);
-    ggml_backend_buffer * buf = nullptr;
-
-    switch (tensor->op) {
-    case GGML_OP_ADD:
-    case GGML_OP_ACC:
-    case GGML_OP_GET_ROWS:
-    case GGML_OP_SUB:
-    case GGML_OP_MUL:
-    case GGML_OP_DIV:
-    case GGML_OP_ADD1:
-    case GGML_OP_ARANGE:
-    case GGML_OP_FILL:
-    case GGML_OP_ADD_ID:
-    case GGML_OP_CONCAT:
-    case GGML_OP_UPSCALE:
-    case GGML_OP_SCALE:
-    case GGML_OP_SQR:
-    case GGML_OP_SQRT:
-    case GGML_OP_SIN:
-    case GGML_OP_COS:
-    case GGML_OP_LOG:
-    case GGML_OP_CLAMP:
-    case GGML_OP_PAD:
-    case GGML_OP_ROLL:
-    case GGML_OP_CPY:
-    case GGML_OP_SET_ROWS:
-    case GGML_OP_CONT:
-    case GGML_OP_DUP:
-    case GGML_OP_SILU_BACK:
-    case GGML_OP_NORM:
-    case GGML_OP_GROUP_NORM:
-    case GGML_OP_RMS_NORM:
-    case GGML_OP_RMS_NORM_BACK:
-    case GGML_OP_L2_NORM:
-    case GGML_OP_DIAG_MASK_INF:
-    case GGML_OP_SOFT_MAX:
-    case GGML_OP_SOFT_MAX_BACK:
-    case GGML_OP_ROPE:
-    case GGML_OP_ROPE_BACK:
-    case GGML_OP_RESHAPE:
-    case GGML_OP_VIEW:
-    case GGML_OP_PERMUTE:
-    case GGML_OP_TRANSPOSE:
-    case GGML_OP_NONE:
-    case GGML_OP_ARGSORT:
-    case GGML_OP_SUM:
-    case GGML_OP_SUM_ROWS:
-    case GGML_OP_MEAN:
-    case GGML_OP_ARGMAX:
-    case GGML_OP_COUNT_EQUAL:
-    case GGML_OP_IM2COL:
-    case GGML_OP_IM2COL_3D:
-    case GGML_OP_TIMESTEP_EMBEDDING:
-    case GGML_OP_CONV_TRANSPOSE_1D:
-    case GGML_OP_POOL_2D:
-    case GGML_OP_CONV_2D:
-    case GGML_OP_CONV_TRANSPOSE_2D:
-    case GGML_OP_CONV_2D_DW:
-    case GGML_OP_RWKV_WKV6:
-    case GGML_OP_RWKV_WKV7:
-    case GGML_OP_SSM_SCAN:
-    case GGML_OP_SSM_CONV:
-    case GGML_OP_LEAKY_RELU:
-    case GGML_OP_REPEAT:
-    case GGML_OP_REPEAT_BACK:
-    case GGML_OP_OPT_STEP_ADAMW:
-    case GGML_OP_OPT_STEP_SGD:
-        buf = tensor->buffer;
-        break;
-    case GGML_OP_UNARY:
-        switch (ggml_get_unary_op(tensor)) {
-        case GGML_UNARY_OP_EXP:
-        case GGML_UNARY_OP_SILU:
-        case GGML_UNARY_OP_GELU:
-        case GGML_UNARY_OP_GELU_ERF:
-        case GGML_UNARY_OP_GELU_QUICK:
-        case GGML_UNARY_OP_RELU:
-        case GGML_UNARY_OP_NEG:
-        case GGML_UNARY_OP_TANH:
-        case GGML_UNARY_OP_SIGMOID:
-        case GGML_UNARY_OP_HARDSIGMOID:
-        case GGML_UNARY_OP_HARDSWISH:
-        case GGML_UNARY_OP_ABS:
-        case GGML_UNARY_OP_SOFTPLUS:
-        case GGML_UNARY_OP_STEP:
-        case GGML_UNARY_OP_ROUND:
-        case GGML_UNARY_OP_CEIL:
-        case GGML_UNARY_OP_FLOOR:
-        case GGML_UNARY_OP_TRUNC:
-            buf = tensor->buffer;
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_GLU:
-        switch (ggml_get_glu_op(tensor)) {
-        case GGML_GLU_OP_GEGLU:
-        case GGML_GLU_OP_REGLU:
-        case GGML_GLU_OP_SWIGLU:
-        case GGML_GLU_OP_SWIGLU_OAI:
-        case GGML_GLU_OP_GEGLU_ERF:
-        case GGML_GLU_OP_GEGLU_QUICK:
-            buf = tensor->buffer;
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_MUL_MAT:
-    case GGML_OP_MUL_MAT_ID:
-    case GGML_OP_FLASH_ATTN_EXT:
-        buf = tensor->buffer;
-
-        break;
-    default:
-        return false;
-    }
-
-    if (buf == nullptr) {
-        return false;
-    }
+    GGML_UNUSED(tensor);
 
     VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")");
 
@@ -12132,8 +12027,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
         subctx->out_memcpys.clear();
         subctx->memsets.clear();
     }
-
-    return true;
 }
 
 // Clean up after graph processing is done
@@ -13162,7 +13055,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->fused_ops_write_mask = 0;
     }
 
-    ctx->prealloc_size_add_rms_partials = std::max(ctx->prealloc_size_add_rms_partials, ctx->prealloc_size_add_rms_partials_offset);
     ctx->last_total_mul_mat_bytes = total_mul_mat_bytes;
 
     if (vk_perf_logger_enabled) {
@@ -13187,6 +13079,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->device->perf_logger->print_timings();
     }
 
+    if (!ctx->device->support_async) {
+        ggml_vk_synchronize(ctx);
+    }
+
     return GGML_STATUS_SUCCESS;
 
     UNUSED(backend);
@@ -13221,24 +13117,6 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
         return false;
     };
 
-    // This function tries to reorder the graph to allow nodes to run in parallel.
-    // This helps with small batches, but for large batches its a slowdown, probably
-    // due to cache contention. So only reorder if the majority of nodes have few rows.
-    int num_small_nodes = 0;
-    int num_counted_nodes = 0;
-    for (int i = 0; i < graph->n_nodes; ++i) {
-        if (!is_empty(graph->nodes[i]) &&
-            graph->nodes[i]->op != GGML_OP_SET_ROWS) {
-            if (ggml_nrows(graph->nodes[i]) <= 8) {
-                num_small_nodes++;
-            }
-            num_counted_nodes++;
-        }
-    }
-    if (num_small_nodes < num_counted_nodes / 2) {
-        return;
-    }
-
     std::vector<ggml_tensor *> new_order;
     std::vector<bool> used(graph->n_nodes, false);
     std::set<ggml_tensor *> used_node_set;
@@ -13480,6 +13358,10 @@ ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
         /* .context = */ ctx,
     };
 
+    if (!ctx->device->support_async) {
+        vk_backend->iface.get_tensor_async = nullptr;
+    }
+
     return vk_backend;
 }
 
@@ -13953,15 +13835,17 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                    op->type == GGML_TYPE_F32;
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM_BACK:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
-            return op->src[0]->type == GGML_TYPE_F32;
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LOG:
             return op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_ARGSORT:
@@ -13978,33 +13862,82 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     return op->ne[0] <= (1 << device->max_workgroup_size_log2);
                 }
             }
+        case GGML_OP_TOP_K:
+            {
+                if (!ggml_is_contiguous(op) || !ggml_is_contiguous(op->src[0])) {
+                    return false;
+                }
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                auto device = ggml_vk_get_device(ctx->device);
+                // We could potentially support larger, using argsort to sort the
+                // whole thing. Not clear if this is needed.
+                uint32_t min_pipeline = (uint32_t)log2f(float(op->ne[0])) + 1;
+                if (min_pipeline >= num_topk_pipelines ||
+                    !device->pipeline_topk_f32[min_pipeline]) {
+                    return false;
+                }
+            }
+            return true;
         case GGML_OP_UPSCALE:
         case GGML_OP_ACC:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONCAT:
+            return ggml_type_size(op->src[0]->type) == ggml_type_size(GGML_TYPE_F32);
         case GGML_OP_ADD1:
+            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32)
+                || (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32)
+                || (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16);
         case GGML_OP_ARANGE:
         case GGML_OP_FILL:
+            return op->type == GGML_TYPE_F32;
         case GGML_OP_SCALE:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
         case GGML_OP_ROLL:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32
+                && (!op->src[1] || (op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16));
         case GGML_OP_SOFT_MAX_BACK:
-            return true;
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32
+                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
+        case GGML_OP_CUMSUM:
+            {
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                auto device = ggml_vk_get_device(ctx->device);
+                if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
+                    return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
+                }
+                return false;
+            }
         case GGML_OP_ARGMAX:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_COUNT_EQUAL:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_I32
+                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_I32;
         case GGML_OP_IM2COL:
+            return ggml_is_contiguous(op->src[1])
+                && op->src[1]->type == GGML_TYPE_F32
+                && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
         case GGML_OP_IM2COL_3D:
+            return op->src[1]->type == GGML_TYPE_F32
+                && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
         case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_2D_DW:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16)
+                && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_POOL_2D:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
-            return true;
+            return true; // all inputs are contiguous, see ggml.c
         case GGML_OP_SSM_SCAN:
             {
                 for (int i = 0; i < 6; i++) {
@@ -14045,7 +13978,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 return true;
             }
         case GGML_OP_SSM_CONV:
-            return true;
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_1D:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_2D:
@@ -14641,10 +14574,14 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             tensor_clone = ggml_get_rows(ggml_ctx, src_clone[0], src_clone[1]);
         } else if (tensor->op == GGML_OP_ARGSORT) {
             tensor_clone = ggml_argsort(ggml_ctx, src_clone[0], (ggml_sort_order) *(int *)tensor->op_params);
+        } else if (tensor->op == GGML_OP_TOP_K) {
+            tensor_clone = ggml_top_k(ggml_ctx, src_clone[0], tensor->ne[0]);
         } else if (tensor->op == GGML_OP_SUM) {
             tensor_clone = ggml_sum(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_SUM_ROWS) {
             tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_CUMSUM) {
+            tensor_clone = ggml_cumsum(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_MEAN) {
             tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_ARGMAX) {

ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp

@@ -0,0 +1,69 @@
+#version 450
+
+#include "types.glsl"
+#include "sum_rows.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 128;
+layout (constant_id = 1) const uint SUBGROUP_SIZE = 32;
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+shared FLOAT_TYPE partial[BLOCK_SIZE / SUBGROUP_SIZE];
+shared FLOAT_TYPE last_sum;
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
+
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
+
+    uint subgroup_id = tid / SUBGROUP_SIZE;
+
+    if (tid == 0) {
+        last_sum = 0;
+    }
+
+    uint col = tid;
+    uint num_iter = CEIL_DIV(p.n_cols, BLOCK_SIZE);
+    for (int i = 0; i < num_iter; ++i) {
+        FLOAT_TYPE v = 0;
+        if (col < p.n_cols) {
+            v = FLOAT_TYPE(data_a[src_idx + col]);
+        }
+        v = subgroupInclusiveAdd(v);
+
+        // Store the largest partial sum for each subgroup, then add the partials for all
+        // lower subgroups and the final partial sum from the previous iteration.
+        if (gl_SubgroupInvocationID == SUBGROUP_SIZE - 1) {
+            partial[subgroup_id] = v;
+        }
+        barrier();
+        for (int j = 0; j < subgroup_id; ++j) {
+            v += partial[j];
+        }
+        v += last_sum;
+        barrier();
+        if (tid == BLOCK_SIZE - 1) {
+            last_sum = v;
+        }
+        if (col < p.n_cols) {
+            data_d[dst_idx + col] = D_TYPE(v);
+        }
+        col += BLOCK_SIZE;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp

@@ -1,6 +1,7 @@
 #version 450
 
 #include "types.glsl"
+#include "sum_rows.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 
@@ -11,30 +12,6 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout (push_constant) uniform parameter
-{
-    uint n_cols;
-    uint ne01, ne02;
-    uint nb01, nb02, nb03;
-    uint nb11, nb12, nb13;
-    float weight;
-    uint misalign_offsets;
-    uint ne0_12mp, ne0_12L;
-    uint ne0_1mp, ne0_1L;
-} p;
-
-uint get_aoffset() { return p.misalign_offsets >> 16; }
-uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
-
-// see init_fastdiv_values in ggml-vulkan.cpp
-uint fastdiv(uint n, uint mp, uint L) {
-    uint msbs, lsbs;
-    // msbs = mulhi(n, mp)
-    umulExtended(n, mp, msbs, lsbs);
-    return (msbs + n) >> L;
-}
-
-
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {

ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl

@@ -0,0 +1,25 @@
+
+// vk_op_sum_rows_push_constants
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+

ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp

@@ -0,0 +1,113 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int NCOLS_PADDED_LOG2 = 10;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// Input can either be the source (A) or intermediate values (S).
+// Similarly, output can be either destination (D) or intermediate values (S).
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 0) readonly buffer S {ivec2 data_s[];};
+layout (binding = 1) writeonly buffer D {int data_d[];};
+layout (binding = 1) writeonly buffer T {ivec2 data_t[];};
+
+layout (push_constant) uniform parameter {
+    uint orig_ncols;
+    uint ncols_input;
+    uint ncols_output;
+    uint nrows;
+    uint first_pass;
+    uint last_pass;
+} p;
+
+// pairs of (gid, value)
+shared ivec2 dst_row[BLOCK_SIZE];
+
+void topk(bool needs_bounds_check, const uint row) {
+    const int col = int(gl_LocalInvocationID.x);
+
+    // initialize indices
+    if (gl_GlobalInvocationID.x < p.ncols_input) {
+        if (p.first_pass != 0) {
+            const uint row_offset = row * p.ncols_input;
+            dst_row[col] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
+        } else {
+            const uint row_offset = row * p.orig_ncols;
+            dst_row[col] = data_s[row_offset + gl_GlobalInvocationID.x];
+        }
+    } else {
+        dst_row[col] = ivec2(p.orig_ncols, 0);
+    }
+    barrier();
+
+    if (p.ncols_output == 1) {
+        // Fast path for single output - just do a max reduction
+        [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
+            if (col < s) {
+                ivec2 a = dst_row[col];
+                ivec2 b = dst_row[col + s];
+                if (a.x >= p.orig_ncols ||
+                    b.x < p.orig_ncols && b.y > a.y) {
+                    dst_row[col] = b;
+                }
+            }
+            barrier();
+        }
+    } else {
+        // bitonic sort on this group of elements
+        uint num_outer_loop_iters = NCOLS_PADDED_LOG2;
+        for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
+            uint num_inner_loop_iters = outer_idx + 1;
+            for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
+                const int ixj = int(col ^ j);
+
+                int idx_0 = (col & k) == 0 ? col : ixj;
+                int idx_1 = (col & k) == 0 ? ixj : col;
+
+                ivec2 sh_idx_0 = dst_row[idx_0];
+                ivec2 sh_idx_1 = dst_row[idx_1];
+                bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.orig_ncols : false;
+                bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.orig_ncols : false;
+
+                if ((idx_0_oob ||
+                    (!idx_1_oob && intBitsToFloat(sh_idx_0.y) < intBitsToFloat(sh_idx_1.y))) && (ixj > col)) {
+                    dst_row[idx_0] = sh_idx_1;
+                    dst_row[idx_1] = sh_idx_0;
+                }
+
+                barrier();
+            }
+        }
+    }
+
+    if (col < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+        if (p.last_pass != 0) {
+            const uint row_offset = row * p.ncols_output;
+            data_d[row_offset + col] = dst_row[col].x;
+        } else {
+            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
+            data_t[row_offset + col] = dst_row[col];
+        }
+    }
+}
+
+void main() {
+    // Fast path for fully occupied workgroups
+    if ((p.ncols_input % BLOCK_SIZE) == 0) {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            topk(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    } else {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            topk(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp

@@ -0,0 +1,199 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_debug_printf : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_shuffle : enable
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int SUBGROUP_SIZE = 32;
+layout(constant_id = 2) const int SUBGROUP_SIZE_LOG2 = 5;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// Input can either be the source (A) or intermediate values (S).
+// Similarly, output can be either destination (D) or intermediate values (S).
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 0) readonly buffer S {ivec2 data_s[];};
+layout (binding = 1) writeonly buffer D {int data_d[];};
+layout (binding = 1) writeonly buffer T {ivec2 data_t[];};
+
+layout (push_constant) uniform parameter {
+    uint orig_ncols;
+    uint ncols_input;
+    uint ncols_output;
+    uint nrows;
+    uint first_pass;
+    uint last_pass;
+} p;
+
+// pairs of (gid, value)
+shared ivec2 dst_row[BLOCK_SIZE];
+
+shared int counts[SUBGROUP_SIZE];
+shared int sh_min_idx;
+shared uint sh_total;
+shared uint offset_partials[BLOCK_SIZE / SUBGROUP_SIZE];
+
+// Map float values to uint such that comparisons still work.
+// Positive values set the high bit, negative values are inverted.
+// +0.0 -> 0x80000000, -0.0 -> 0x7FFFFFFF are in the correct places.
+uint f2ui(float x) {
+    uint y = floatBitsToUint(x);
+    if ((y & 0x80000000) != 0) {
+        y ^= ~0;
+    } else {
+        y |= 0x80000000;
+    }
+    return y;
+}
+
+void topk(const uint row) {
+    const int tid = int(gl_LocalInvocationID.x);
+
+    // initialize indices
+    if (gl_GlobalInvocationID.x < p.ncols_input) {
+        if (p.first_pass != 0) {
+            const uint row_offset = row * p.ncols_input;
+            dst_row[tid] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
+        } else {
+            const uint row_offset = row * p.orig_ncols;
+            dst_row[tid] = data_s[row_offset + gl_GlobalInvocationID.x];
+        }
+    } else {
+        dst_row[tid] = ivec2(p.orig_ncols, 0xFF800000); // -inf
+    }
+    barrier();
+
+    if (p.ncols_output == 1) {
+        // Fast path for single output - just do a max reduction
+        [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
+            if (tid < s) {
+                ivec2 a = dst_row[tid];
+                ivec2 b = dst_row[tid + s];
+                if (a.x >= p.orig_ncols ||
+                    b.x < p.orig_ncols && b.y > a.y) {
+                    dst_row[tid] = b;
+                }
+            }
+            barrier();
+        }
+    } else {
+        // Do an N-ary search to find the K-th largest value.
+        // We remap the float values to be comparable as unsigned integers,
+        // and split the range into 2^N smaller ranges where N is the
+        // subgroup size. Count how many values are in each range, if the K-th
+        // largest value is in the middle of one of thee ranges then repeat
+        // and split again.
+
+        // Mask is the current set of bits we're searching. Shift is the LSB index.
+        int shift = 32 - SUBGROUP_SIZE_LOG2;
+        uint mask = ((1 << SUBGROUP_SIZE_LOG2) - 1) << shift;
+
+        // The current range.
+        uint range_min = 0;
+        uint range_max = 0xFF800000;
+        // How many are above the current range, and how many we need to find.
+        uint total = 0;
+        uint limit = min(p.ncols_output, p.ncols_input - gl_WorkGroupID.x * BLOCK_SIZE);
+
+        while (mask != 0) {
+            barrier();
+            // Initialize bucket counts to zero.
+            if (tid < SUBGROUP_SIZE) {
+                counts[tid] = 0;
+            }
+            barrier();
+            // Count how many values are in each bucket.
+            if (tid < p.ncols_input) {
+                float y = intBitsToFloat(dst_row[tid].y);
+                uint fy = f2ui(y);
+                if (fy >= range_min && fy < range_max) {
+                    uint bucket = (fy & mask) >> shift;
+                    atomicAdd(counts[bucket], 1);
+                }
+            }
+            barrier();
+
+            // On the first subgroup, do a scan to count (from the top down) how
+            // many elements are in the top N buckets. Find the index of the first
+            // that is over the limit. Copy it to the other invocations through
+            // shared memory.
+            if (tid < SUBGROUP_SIZE) {
+                uint partial_sum = counts[SUBGROUP_SIZE - 1 - tid];
+                partial_sum = subgroupInclusiveAdd(partial_sum) + total;
+                uint t = subgroupBallotFindLSB(subgroupBallot(partial_sum >= limit));
+                if (tid == t) {
+                    sh_min_idx = int(SUBGROUP_SIZE - 1 - t);
+                    sh_total = partial_sum;
+                }
+            }
+            barrier();
+            int min_idx = sh_min_idx;
+            total = sh_total;
+
+            // Update the range, and break if we've found the K-th largest.
+            range_max = range_min + ((min_idx + 1) << shift);
+            range_min = range_min + (min_idx << shift);
+
+            if (total == p.ncols_output) {
+                break;
+            }
+            total -= counts[min_idx];
+            mask >>= SUBGROUP_SIZE_LOG2;
+            shift -= SUBGROUP_SIZE_LOG2;
+            if (shift < 0) {
+                shift = 0;
+            }
+        }
+
+        ivec2 v = dst_row[tid];
+
+        // We need to compact these values to the start of the dst_row array.
+        // Have each subgroup count how many items it'll store, so other
+        // subgroups can compute their base offset.
+        bool top = f2ui(intBitsToFloat(v.y)) >= range_min;
+        uvec4 b = subgroupBallot(top);
+        uint bit_count = subgroupBallotBitCount(b);
+        if ((tid % SUBGROUP_SIZE) == 0) {
+            offset_partials[tid / SUBGROUP_SIZE] = bit_count;
+        }
+        barrier();
+
+        uint out_idx = 0;
+        [[unroll]] for (int i = 0; i < BLOCK_SIZE / SUBGROUP_SIZE; ++i) {
+            if (i < tid / SUBGROUP_SIZE) {
+                out_idx += offset_partials[i];
+            }
+        }
+
+        uint bit_count_ex = subgroupBallotExclusiveBitCount(b);
+        if (top) {
+            // TODO: Copy directly to the output?
+            dst_row[out_idx + bit_count_ex] = v;
+        }
+
+        barrier();
+    }
+
+    if (tid < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+        if (p.last_pass != 0) {
+            const uint row_offset = row * p.ncols_output;
+            data_d[row_offset + tid] = dst_row[tid].x;
+        } else {
+            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
+            data_t[row_offset + tid] = dst_row[tid];
+        }
+    }
+}
+
+void main() {
+    uint row = gl_WorkGroupID.y;
+    while (row < p.nrows) {
+        topk(row);
+        row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -913,9 +913,13 @@ void process_shaders() {
     string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
     string_to_spv("argsort_large_f32", "argsort_large.comp", {{"A_TYPE", "float"}});
 
+    string_to_spv("topk_argsort_f32", "topk_argsort.comp", {{"A_TYPE", "float"}});
+    string_to_spv("topk_nary_search_f32", "topk_nary_search.comp", {{"A_TYPE", "float"}});
+
     string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
     string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
+    string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     for (std::string dim_str : {"", "_3d"}) {
         for (bool bda : {false, true}) {

ggml/src/ggml.c

@@ -990,6 +990,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
+    "TOP_K",
     "LEAKY_RELU",
     "TRI",
     "FILL",
@@ -1023,7 +1024,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
+static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1098,6 +1099,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
+    "top_k(x)",
     "leaky_relu(x)",
     "tri(x)",
     "fill(x, c)",
@@ -1131,7 +1133,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
+static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -5036,28 +5038,6 @@ struct ggml_tensor * ggml_roll(
     return result;
 }
 
-// ggml_arange
-
-struct ggml_tensor * ggml_arange(
-        struct ggml_context * ctx,
-        float                 start,
-        float                 stop,
-        float                 step) {
-    GGML_ASSERT(stop > start);
-
-    const int64_t steps = (int64_t) ceilf((stop - start) / step);
-
-    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
-
-    ggml_set_op_params_f32(result, 0, start);
-    ggml_set_op_params_f32(result, 1, stop);
-    ggml_set_op_params_f32(result, 2, step);
-
-    result->op = GGML_OP_ARANGE;
-
-    return result;
-}
-
 // ggml_timestep_embedding
 
 struct ggml_tensor * ggml_timestep_embedding(
@@ -5139,6 +5119,7 @@ struct ggml_tensor * ggml_argsort(
         struct ggml_tensor   * a,
         enum ggml_sort_order   order) {
     GGML_ASSERT(a->ne[0] <= INT32_MAX);
+
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -5149,6 +5130,24 @@ struct ggml_tensor * ggml_argsort(
     return result;
 }
 
+// ggml_argsort_top_k
+
+struct ggml_tensor * ggml_argsort_top_k(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   k) {
+    GGML_ASSERT(a->ne[0] >= k);
+
+    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
+
+    result = ggml_view_4d(ctx, result,
+                k, result->ne[1], result->ne[2], result->ne[3],
+                   result->nb[1], result->nb[2], result->nb[3],
+                0);
+
+    return result;
+}
+
 // ggml_top_k
 
 struct ggml_tensor * ggml_top_k(
@@ -5157,12 +5156,32 @@ struct ggml_tensor * ggml_top_k(
         int                   k) {
     GGML_ASSERT(a->ne[0] >= k);
 
-    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
 
-    result = ggml_view_4d(ctx, result,
-                k, result->ne[1], result->ne[2], result->ne[3],
-                   result->nb[1], result->nb[2], result->nb[3],
-                0);
+    result->op     = GGML_OP_TOP_K;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_arange
+
+struct ggml_tensor * ggml_arange(
+        struct ggml_context * ctx,
+        float                 start,
+        float                 stop,
+        float                 step) {
+    GGML_ASSERT(stop > start);
+
+    const int64_t steps = (int64_t) ceilf((stop - start) / step);
+
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
+
+    ggml_set_op_params_f32(result, 0, start);
+    ggml_set_op_params_f32(result, 1, stop);
+    ggml_set_op_params_f32(result, 2, step);
+
+    result->op = GGML_OP_ARANGE;
 
     return result;
 }

gguf-py/gguf/constants.py

@@ -25,6 +25,20 @@ class Keys:
         ALIGNMENT                  = "general.alignment"
         FILE_TYPE                  = "general.file_type"
 
+        # Recommended Sampler Parameters
+        SAMPLING_SEQUENCE           = "general.sampling.sequence"
+        SAMPLING_TOP_K              = "general.sampling.top_k"
+        SAMPLING_TOP_P              = "general.sampling.top_p"
+        SAMPLING_MIN_P              = "general.sampling.min_p"
+        SAMPLING_XTC_PROBABILITY    = "general.sampling.xtc_probability"
+        SAMPLING_XTC_THRESHOLD      = "general.sampling.xtc_threshold"
+        SAMPLING_TEMP               = "general.sampling.temp"
+        SAMPLING_PENALTY_LAST_N     = "general.sampling.penalty_last_n"
+        SAMPLING_PENALTY_REPEAT     = "general.sampling.penalty_repeat"
+        SAMPLING_MIROSTAT           = "general.sampling.mirostat"
+        SAMPLING_MIROSTAT_TAU       = "general.sampling.mirostat_tau"
+        SAMPLING_MIROSTAT_ETA       = "general.sampling.mirostat_eta"
+
         # Authorship Metadata
         NAME                       = "general.name"
         AUTHOR                     = "general.author"
@@ -427,6 +441,7 @@ class MODEL_ARCH(IntEnum):
     APERTUS          = auto()
     COGVLM           = auto()
     MINIMAXM2        = auto()
+    RND1             = auto()
     PANGU_EMBED      = auto()
 
 
@@ -797,6 +812,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.APERTUS:          "apertus",
     MODEL_ARCH.MINIMAXM2:        "minimax-m2",
     MODEL_ARCH.COGVLM:           "cogvlm",
+    MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
 }
 
@@ -2991,6 +3007,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.VISEXP_UP,
         MODEL_TENSOR.VISEXP_DOWN,
     ],
+    MODEL_ARCH.RND1: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     MODEL_ARCH.PANGU_EMBED: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/gguf_writer.py

@@ -4,6 +4,7 @@ import logging
 import os
 import shutil
 import struct
+import sys
 import tempfile
 from dataclasses import dataclass
 from enum import Enum, auto
@@ -372,8 +373,10 @@ class GGUFWriter:
         self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
         raw_dtype: GGMLQuantizationType | None = None,
     ) -> None:
-        if self.endianess == GGUFEndian.BIG:
-            tensor.byteswap(inplace=True)
+        if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \
+                (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'):
+            # Don't byteswap inplace since lazy copies cannot handle it
+            tensor = tensor.byteswap(inplace=False)
         if self.use_temp_file and self.temp_file is None:
             fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256 * 1024 * 1024)
             fp.seek(0)
@@ -399,8 +402,10 @@ class GGUFWriter:
             raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}')
         assert self.fout is not None
 
-        if self.endianess == GGUFEndian.BIG:
-            tensor.byteswap(inplace=True)
+        if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \
+                (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'):
+            # Don't byteswap inplace since lazy copies cannot handle it
+            tensor = tensor.byteswap(inplace=False)
 
         file_id = -1
         for i, tensors in enumerate(self.tensors):
@@ -496,6 +501,42 @@ class GGUFWriter:
     def add_file_type(self, ftype: int) -> None:
         self.add_uint32(Keys.General.FILE_TYPE, ftype)
 
+    def add_sampling_sequence(self, sequence: str) -> None:
+        self.add_string(Keys.General.SAMPLING_SEQUENCE, sequence)
+
+    def add_sampling_top_k(self, top_k: int) -> None:
+        self.add_int32(Keys.General.SAMPLING_TOP_K, top_k)
+
+    def add_sampling_top_p(self, top_p: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_TOP_P, top_p)
+
+    def add_sampling_min_p(self, min_p: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_MIN_P, min_p)
+
+    def add_sampling_xtc_probability(self, xtc_probability: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_XTC_PROBABILITY, xtc_probability)
+
+    def add_sampling_xtc_threshold(self, xtc_threshold: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_XTC_THRESHOLD, xtc_threshold)
+
+    def add_sampling_temp(self, temp: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_TEMP, temp)
+
+    def add_sampling_penalty_last_n(self, penalty_last_n: int) -> None:
+        self.add_int32(Keys.General.SAMPLING_PENALTY_LAST_N, penalty_last_n)
+
+    def add_sampling_penalty_repeat(self, penalty_repeat: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_PENALTY_REPEAT, penalty_repeat)
+
+    def add_sampling_mirostat(self, mirostat: int) -> None:
+        self.add_int32(Keys.General.SAMPLING_MIROSTAT, mirostat)
+
+    def add_sampling_mirostat_tau(self, mirostat_tau: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_MIROSTAT_TAU, mirostat_tau)
+
+    def add_sampling_mirostat_eta(self, mirostat_eta: float) -> None:
+        self.add_float32(Keys.General.SAMPLING_MIROSTAT_ETA, mirostat_eta)
+
     def add_name(self, name: str) -> None:
         self.add_string(Keys.General.NAME, name)
 

gguf-py/gguf/metadata.py

@@ -17,6 +17,20 @@ logger = logging.getLogger("metadata")
 
 @dataclass
 class Metadata:
+    # Recommended Sampler Parameters to be written to GGUF KV Store
+    sampling_sequence: Optional[str] = None
+    sampling_top_k: Optional[int] = None
+    sampling_top_p: Optional[float] = None
+    sampling_min_p: Optional[float] = None
+    sampling_xtc_probability: Optional[float] = None
+    sampling_xtc_threshold: Optional[float] = None
+    sampling_temp: Optional[float] = None
+    sampling_penalty_last_n: Optional[int] = None
+    sampling_penalty_repeat: Optional[float] = None
+    sampling_mirostat: Optional[int] = None
+    sampling_mirostat_tau: Optional[float] = None
+    sampling_mirostat_eta: Optional[float] = None
+
     # Authorship Metadata to be written to GGUF KV Store
     name: Optional[str] = None
     author: Optional[str] = None
@@ -54,15 +68,43 @@ class Metadata:
 
         model_card = Metadata.load_model_card(model_path)
         hf_params = Metadata.load_hf_parameters(model_path)
+        gen_config = Metadata.load_generation_config(model_path)
         # TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter
 
         # heuristics
         metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params)
 
+        if gen_config:
+            metadata.sampling_sequence        = gen_config.get("sequence",        metadata.sampling_sequence)
+            metadata.sampling_top_k           = gen_config.get("top_k",           metadata.sampling_top_k)
+            metadata.sampling_top_p           = gen_config.get("top_p",           metadata.sampling_top_p)
+            metadata.sampling_min_p           = gen_config.get("min_p",           metadata.sampling_min_p)
+            metadata.sampling_xtc_probability = gen_config.get("xtc_probability", metadata.sampling_xtc_probability)
+            metadata.sampling_xtc_threshold   = gen_config.get("xtc_threshold",   metadata.sampling_xtc_threshold)
+            metadata.sampling_temp            = gen_config.get("temperature",     metadata.sampling_temp)
+            metadata.sampling_penalty_last_n  = gen_config.get("penalty_last_n",  metadata.sampling_penalty_last_n)
+            metadata.sampling_penalty_repeat  = gen_config.get("penalty_repeat",  metadata.sampling_penalty_repeat)
+            metadata.sampling_mirostat        = gen_config.get("mirostat",        metadata.sampling_mirostat)
+            metadata.sampling_mirostat_tau    = gen_config.get("mirostat_tau",    metadata.sampling_mirostat_tau)
+            metadata.sampling_mirostat_eta    = gen_config.get("mirostat_eta",    metadata.sampling_mirostat_eta)
+
         # Metadata Override File Provided
         # This is based on LLM_KV_NAMES mapping in llama.cpp
         metadata_override = Metadata.load_metadata_override(metadata_override_path)
 
+        metadata.sampling_sequence        = metadata_override.get(Keys.General.SAMPLING_SEQUENCE,        metadata.sampling_sequence)
+        metadata.sampling_top_k           = metadata_override.get(Keys.General.SAMPLING_TOP_K,           metadata.sampling_top_k)
+        metadata.sampling_top_p           = metadata_override.get(Keys.General.SAMPLING_TOP_P,           metadata.sampling_top_p)
+        metadata.sampling_min_p           = metadata_override.get(Keys.General.SAMPLING_MIN_P,           metadata.sampling_min_p)
+        metadata.sampling_xtc_probability = metadata_override.get(Keys.General.SAMPLING_XTC_PROBABILITY, metadata.sampling_xtc_probability)
+        metadata.sampling_xtc_threshold   = metadata_override.get(Keys.General.SAMPLING_XTC_THRESHOLD,   metadata.sampling_xtc_threshold)
+        metadata.sampling_temp            = metadata_override.get(Keys.General.SAMPLING_TEMP,            metadata.sampling_temp)
+        metadata.sampling_penalty_last_n  = metadata_override.get(Keys.General.SAMPLING_PENALTY_LAST_N,  metadata.sampling_penalty_last_n)
+        metadata.sampling_penalty_repeat  = metadata_override.get(Keys.General.SAMPLING_PENALTY_REPEAT,  metadata.sampling_penalty_repeat)
+        metadata.sampling_mirostat        = metadata_override.get(Keys.General.SAMPLING_MIROSTAT,        metadata.sampling_mirostat)
+        metadata.sampling_mirostat_tau    = metadata_override.get(Keys.General.SAMPLING_MIROSTAT_TAU,    metadata.sampling_mirostat_tau)
+        metadata.sampling_mirostat_eta    = metadata_override.get(Keys.General.SAMPLING_MIROSTAT_ETA,    metadata.sampling_mirostat_eta)
+
         metadata.name            = metadata_override.get(Keys.General.NAME,            metadata.name)
         metadata.author          = metadata_override.get(Keys.General.AUTHOR,          metadata.author)
         metadata.version         = metadata_override.get(Keys.General.VERSION,         metadata.version)
@@ -172,6 +214,23 @@ class Metadata:
         with open(config_path, "r", encoding="utf-8") as f:
             return json.load(f)
 
+    @staticmethod
+    def load_generation_config(model_path: Optional[Path] = None) -> dict[str, Any]:
+        if model_path is None or not model_path.is_dir():
+            return {}
+
+        generation_config_path = model_path / "generation_config.json"
+
+        if not generation_config_path.is_file():
+            return {}
+
+        try:
+            with open(generation_config_path, "r", encoding="utf-8") as f:
+                return json.load(f)
+        except (json.JSONDecodeError, IOError):
+            # not all models have valid generation_config.json
+            return {}
+
     @staticmethod
     def id_to_title(string):
         # Convert capitalization into title form unless acronym or version number
@@ -546,6 +605,32 @@ class Metadata:
 
     def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter):
         assert self.name is not None
+
+        if self.sampling_sequence is not None:
+            gguf_writer.add_sampling_sequence(self.sampling_sequence)
+        if self.sampling_top_k is not None:
+            gguf_writer.add_sampling_top_k(self.sampling_top_k)
+        if self.sampling_top_p is not None:
+            gguf_writer.add_sampling_top_p(self.sampling_top_p)
+        if self.sampling_min_p is not None:
+            gguf_writer.add_sampling_min_p(self.sampling_min_p)
+        if self.sampling_xtc_probability is not None:
+            gguf_writer.add_sampling_xtc_probability(self.sampling_xtc_probability)
+        if self.sampling_xtc_threshold is not None:
+            gguf_writer.add_sampling_xtc_threshold(self.sampling_xtc_threshold)
+        if self.sampling_temp is not None:
+            gguf_writer.add_sampling_temp(self.sampling_temp)
+        if self.sampling_penalty_last_n is not None:
+            gguf_writer.add_sampling_penalty_last_n(self.sampling_penalty_last_n)
+        if self.sampling_penalty_repeat is not None:
+            gguf_writer.add_sampling_penalty_repeat(self.sampling_penalty_repeat)
+        if self.sampling_mirostat is not None:
+            gguf_writer.add_sampling_mirostat(self.sampling_mirostat)
+        if self.sampling_mirostat_tau is not None:
+            gguf_writer.add_sampling_mirostat_tau(self.sampling_mirostat_tau)
+        if self.sampling_mirostat_eta is not None:
+            gguf_writer.add_sampling_mirostat_eta(self.sampling_mirostat_eta)
+
         gguf_writer.add_name(self.name)
 
         if self.author is not None:

gguf-py/gguf/scripts/gguf_convert_endian.py

@@ -19,6 +19,11 @@ import gguf
 logger = logging.getLogger("gguf-convert-endian")
 
 
+def byteswap_noop(tensor, block_offs):
+    # this function is used when byteswapping is not needed
+    pass
+
+
 def byteswap_q4_0(tensor, block_offs):
     # Each block_q4_0 consists of an f16 delta (scaling factor) followed by 16 int8 quantizations.
 
@@ -55,22 +60,11 @@ def byteswap_q6_k(tensor, block_offs):
 
 
 byteswap_tensors = {
-    gguf.GGMLQuantizationType.Q4_0: {
-        "block_size": 18, # 18 bytes = <f16 delta scaling factor> + 16 * <int8 quant>
-        "byteswap_func": byteswap_q4_0,
-    },
-    gguf.GGMLQuantizationType.Q8_0: {
-        "block_size": 34, # 34 bytes = <f16 delta scaling factor> + 32 * <int8 quant>
-        "byteswap_func": byteswap_q8_0,
-    },
-    gguf.GGMLQuantizationType.Q4_K: {
-        "block_size": 144, # 144 bytes = 2 * <f16 delta scaling factor> + 140 * <int8 quant>
-        "byteswap_func": byteswap_q4_k,
-    },
-    gguf.GGMLQuantizationType.Q6_K: {
-        "block_size": 210, # 210 bytes = <f16 delta scaling factor> + 208 * <int8 quant>
-        "byteswap_func": byteswap_q6_k,
-    },
+    gguf.GGMLQuantizationType.Q4_0:  byteswap_q4_0,
+    gguf.GGMLQuantizationType.Q8_0:  byteswap_q8_0,
+    gguf.GGMLQuantizationType.Q4_K:  byteswap_q4_k,
+    gguf.GGMLQuantizationType.Q6_K:  byteswap_q6_k,
+    gguf.GGMLQuantizationType.MXFP4: byteswap_noop,
 }
 
 
@@ -135,8 +129,8 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None
 
             tensor.data.resize(newshape)
 
-            block_size    = byteswap_tensors[tensor.tensor_type]["block_size"]
-            byteswap_func = byteswap_tensors[tensor.tensor_type]["byteswap_func"]
+            block_size    = gguf.constants.GGML_QUANT_SIZES[tensor.tensor_type][1]
+            byteswap_func = byteswap_tensors[tensor.tensor_type]
 
             n_blocks = len(tensor.data) // block_size
             for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)):

include/llama.h

@@ -246,6 +246,21 @@ extern "C" {
         LLAMA_KV_OVERRIDE_TYPE_STR,
     };
 
+    enum llama_model_meta_key {
+        LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_K,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIN_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD,
+        LLAMA_MODEL_META_KEY_SAMPLING_TEMP,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA,
+    };
+
     struct llama_model_kv_override {
         enum llama_model_kv_override_type tag;
 
@@ -518,6 +533,9 @@ extern "C" {
     // Get the number of metadata key/value pairs
     LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
 
+    // Get sampling metadata key name. Returns nullptr if the key is invalid
+    LLAMA_API const char * llama_model_meta_key_str(enum llama_model_meta_key key);
+
     // Get metadata key name by index
     LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
 

scripts/sync-ggml.last

@@ -1 +1 @@
-7b6abb2b92fcef35cb01c6ce6ada9bd85306522d
+55bc9320a4aae82af18e23eefd5de319a755d7b9

scripts/sync_vendor.py

@@ -16,7 +16,7 @@ vendor = {
     # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
     "https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",
 
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.27.0/httplib.h": "vendor/cpp-httplib/httplib.h",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.28.0/httplib.h": "vendor/cpp-httplib/httplib.h",
 }
 
 for url, filename in vendor.items():

src/CMakeLists.txt

@@ -115,6 +115,7 @@ add_library(llama
             models/qwen3vl-moe.cpp
             models/qwen3moe.cpp
             models/refact.cpp
+            models/rnd1.cpp
             models/rwkv6-base.cpp
             models/rwkv6.cpp
             models/rwkv6qwen2.cpp

src/llama-arch.cpp

@@ -108,24 +108,37 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_APERTUS,          "apertus"          },
     { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
     { LLM_ARCH_COGVLM,           "cogvlm"           },
+    { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-    { LLM_KV_GENERAL_TYPE,                 "general.type"                          },
-    { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
-    { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
-    { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
-    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
-    { LLM_KV_GENERAL_NAME,                 "general.name"                          },
-    { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
-    { LLM_KV_GENERAL_VERSION,              "general.version"                       },
-    { LLM_KV_GENERAL_URL,                  "general.url"                           },
-    { LLM_KV_GENERAL_DESCRIPTION,          "general.description"                   },
-    { LLM_KV_GENERAL_LICENSE,              "general.license"                       },
-    { LLM_KV_GENERAL_SOURCE_URL,           "general.source.url"                    },
-    { LLM_KV_GENERAL_SOURCE_HF_REPO,       "general.source.huggingface.repository" },
+    { LLM_KV_GENERAL_TYPE,                     "general.type"                          },
+    { LLM_KV_GENERAL_ARCHITECTURE,             "general.architecture"                  },
+    { LLM_KV_GENERAL_QUANTIZATION_VERSION,     "general.quantization_version"          },
+    { LLM_KV_GENERAL_ALIGNMENT,                "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,                "general.file_type"                     },
+    { LLM_KV_GENERAL_SAMPLING_SEQUENCE,        "general.sampling.sequence"             },
+    { LLM_KV_GENERAL_SAMPLING_TOP_K,           "general.sampling.top_k"                },
+    { LLM_KV_GENERAL_SAMPLING_TOP_P,           "general.sampling.top_p"                },
+    { LLM_KV_GENERAL_SAMPLING_MIN_P,           "general.sampling.min_p"                },
+    { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability"      },
+    { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,   "general.sampling.xtc_threshold"        },
+    { LLM_KV_GENERAL_SAMPLING_TEMP,            "general.sampling.temp"                 },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,  "general.sampling.penalty_last_n"       },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,  "general.sampling.penalty_repeat"       },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT,        "general.sampling.mirostat"             },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,    "general.sampling.mirostat_tau"         },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,    "general.sampling.mirostat_eta"         },
+    { LLM_KV_GENERAL_NAME,                     "general.name"                          },
+    { LLM_KV_GENERAL_AUTHOR,                   "general.author"                        },
+    { LLM_KV_GENERAL_VERSION,                  "general.version"                       },
+    { LLM_KV_GENERAL_URL,                      "general.url"                           },
+    { LLM_KV_GENERAL_DESCRIPTION,              "general.description"                   },
+    { LLM_KV_GENERAL_LICENSE,                  "general.license"                       },
+    { LLM_KV_GENERAL_SOURCE_URL,               "general.source.url"                    },
+    { LLM_KV_GENERAL_SOURCE_HF_REPO,           "general.source.huggingface.repository" },
 
     { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
     { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
@@ -2446,6 +2459,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_VISEXP_FFN_UP,   "blk.%d.vis_up" },
         },
     },
+    {
+        LLM_ARCH_RND1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2722,6 +2755,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
             return true;
         default:
             return false;

src/llama-arch.h

@@ -112,6 +112,7 @@ enum llm_arch {
     LLM_ARCH_APERTUS,
     LLM_ARCH_MINIMAX_M2,
     LLM_ARCH_COGVLM,
+    LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_UNKNOWN,
 };
@@ -122,6 +123,18 @@ enum llm_kv {
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
     LLM_KV_GENERAL_FILE_TYPE,
+    LLM_KV_GENERAL_SAMPLING_SEQUENCE,
+    LLM_KV_GENERAL_SAMPLING_TOP_K,
+    LLM_KV_GENERAL_SAMPLING_TOP_P,
+    LLM_KV_GENERAL_SAMPLING_MIN_P,
+    LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY,
+    LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,
+    LLM_KV_GENERAL_SAMPLING_TEMP,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,

src/llama-context.cpp

@@ -1248,7 +1248,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
-        if (!sorted_output) {
+        if (!sorted_output && n_outputs > 1) {
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?

src/llama-grammar.cpp

@@ -6,8 +6,10 @@
 
 #include <cmath>
 #include <algorithm>
+#include <cstdint>
 #include <stdexcept>
 
+#define MAX_REPETITION_THRESHOLD 2000
 //
 // helpers
 //
@@ -345,8 +347,10 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-    auto handle_repetitions = [&](int min_times, int max_times) {
-
+    // use UINT64_MAX as the empty value because we aligned to the proper uint64_t type so -1 can't be used
+    // (though it's technically the same as -1 now)
+    auto handle_repetitions = [&](uint64_t min_times, uint64_t max_times) {
+        bool no_max = max_times == UINT64_MAX;
         if (last_sym_start == rule.size()) {
             throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
         }
@@ -373,20 +377,20 @@ const char * llama_grammar_parser::parse_sequence(
             rule.resize(last_sym_start);
         } else {
             // Repeat the previous elements (min_times - 1) times
-            for (int i = 1; i < min_times; i++) {
+            for (uint64_t i = 1; i < min_times; i++) {
                 rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
 
         uint32_t last_rec_rule_id = 0;
-        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+        auto n_opt = no_max ? 1 : max_times - min_times;
 
         llama_grammar_rule rec_rule(prev_rule);
-        for (int i = 0; i < n_opt; i++) {
+        for (uint64_t i = 0; i < n_opt; i++) {
             rec_rule.resize(prev_rule.size());
             uint32_t rec_rule_id = generate_symbol_id( rule_name);
-            if (i > 0 || max_times < 0) {
-                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            if (i > 0 || no_max) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, no_max ? rec_rule_id : last_rec_rule_id});
             }
             rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
             rec_rule.push_back({LLAMA_GRETYPE_END, 0});
@@ -478,10 +482,10 @@ const char * llama_grammar_parser::parse_sequence(
                 throw std::runtime_error(std::string("expecting an int at ") + pos);
             }
             const char * int_end = parse_int(pos);
-            int min_times = std::stoul(std::string(pos, int_end - pos));
+            uint64_t min_times = std::stoul(std::string(pos, int_end - pos));
             pos = parse_space(int_end, is_nested);
 
-            int max_times = -1;
+            uint64_t max_times = UINT64_MAX; // default: no max limit
 
             if (*pos == '}') {
                 max_times = min_times;
@@ -502,6 +506,10 @@ const char * llama_grammar_parser::parse_sequence(
             } else {
                 throw std::runtime_error(std::string("expecting ',' at ") + pos);
             }
+            bool has_max = max_times != UINT64_MAX;
+            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
+            }
             handle_repetitions(min_times, max_times);
         } else {
             break;

src/llama-graph.cpp

@@ -961,14 +961,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         // organize experts into n_expert_groups
         ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
 
-        ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
+        ggml_tensor * group_scores = ggml_argsort_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
         group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
 
         // get top n_group_used expert groups
         group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
         group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
 
-        ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
+        ggml_tensor * expert_groups = ggml_argsort_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
         cb(expert_groups, "ffn_moe_group_topk", il);
 
         // mask out the other groups
@@ -979,7 +979,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     }
 
     // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
+    ggml_tensor * selected_experts = ggml_argsort_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
     cb(selected_experts, "ffn_moe_topk", il);
 

src/llama-impl.cpp

@@ -20,10 +20,10 @@ static llama_logger_state g_logger_state;
 time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
 
 time_meas::~time_meas() {
-        if (t_start_us >= 0) {
-            t_acc += ggml_time_us() - t_start_us;
-        }
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
     }
+}
 
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);

src/llama-model.cpp

@@ -1036,6 +1036,18 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_RND1:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_30B_A3B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+                // Set non-causal attention for diffusion models
+                hparams.causal_attn = false;
+            } break;
         case LLM_ARCH_QWEN2MOE:
             {
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
@@ -1593,7 +1605,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_DEEPSEEK2:
             {
-                bool is_lite = (hparams.n_layer == 27);
+                // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+                bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
                 if (!is_lite) {
@@ -3401,6 +3414,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_QWEN3MOE:
             case LLM_ARCH_QWEN3VLMOE:
+            case LLM_ARCH_RND1:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -4581,7 +4595,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_DEEPSEEK2:
                 {
-                    const bool is_lite = (hparams.n_layer == 27);
+                    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+                    const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
 
                     const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
 
@@ -6718,7 +6733,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
 
-    if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE) {
+    if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE || arch == LLM_ARCH_RND1) {
         LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
     }
 
@@ -6880,6 +6895,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
             {
                 res = nullptr;
             } break;
@@ -7073,6 +7089,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_llada_moe>(*this, params);
             }
             break;
+        case LLM_ARCH_RND1:
+            {
+                llm = std::make_unique<llm_build_rnd1>(*this, params);
+            }
+            break;
         case LLM_ARCH_QWEN2VL:
             {
                 llm = std::make_unique<llm_build_qwen2vl>(*this, params);
@@ -7593,6 +7614,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_QWEN3:
         case LLM_ARCH_QWEN3MOE:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
         case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
@@ -7665,6 +7687,24 @@ int32_t llama_model_meta_count(const llama_model * model) {
     return (int)model->gguf_kv.size();
 }
 
+const char * llama_model_meta_key_str(llama_model_meta_key key) {
+    switch (key) {
+        case LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE:        return "general.sampling.sequence";
+        case LLAMA_MODEL_META_KEY_SAMPLING_TOP_K:           return "general.sampling.top_k";
+        case LLAMA_MODEL_META_KEY_SAMPLING_TOP_P:           return "general.sampling.top_p";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIN_P:           return "general.sampling.min_p";
+        case LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY: return "general.sampling.xtc_probability";
+        case LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD:   return "general.sampling.xtc_threshold";
+        case LLAMA_MODEL_META_KEY_SAMPLING_TEMP:            return "general.sampling.temp";
+        case LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N:  return "general.sampling.penalty_last_n";
+        case LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT:  return "general.sampling.penalty_repeat";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT:        return "general.sampling.mirostat";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU:    return "general.sampling.mirostat_tau";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA:    return "general.sampling.mirostat_eta";
+        default:                                            return nullptr;
+    }
+}
+
 int32_t llama_model_meta_key_by_index(const llama_model * model, int i, char * buf, size_t buf_size) {
     if (i < 0 || i >= (int)model->gguf_kv.size()) {
         if (buf_size > 0) {

src/llama-sampling.cpp

@@ -472,9 +472,6 @@ static void llama_sampler_chain_reset(struct llama_sampler * smpl) {
     for (auto * smpl : chain->samplers) {
         llama_sampler_reset(smpl);
     }
-
-    chain->t_sample_us = 0;
-    chain->n_sample    = 0;
 }
 
 static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) {
@@ -2670,8 +2667,7 @@ struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * c
 void llama_perf_sampler_print(const struct llama_sampler * chain) {
     const auto data = llama_perf_sampler(chain);
 
-    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+    LLAMA_LOG_INFO("%s:    samplers time = %10.2f ms / %5d runs\n", __func__, data.t_sample_ms, data.n_sample);
 }
 
 void llama_perf_sampler_reset(struct llama_sampler * chain) {
@@ -2681,5 +2677,6 @@ void llama_perf_sampler_reset(struct llama_sampler * chain) {
 
     auto * ctx = (struct llama_sampler_chain *) chain->ctx;
 
-    ctx->t_sample_us = ctx->n_sample = 0;
+    ctx->t_sample_us = 0;
+    ctx->n_sample    = 0;
 }

src/models/deepseek2.cpp

@@ -4,7 +4,8 @@
 
 llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    bool is_lite = (hparams.n_layer == 27);
+    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+    bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
 
     const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
 

src/models/models.h

@@ -431,6 +431,10 @@ struct llm_build_refact : public llm_graph_context {
     llm_build_refact(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_rnd1 : public llm_graph_context {
+    llm_build_rnd1(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_rwkv6 : public llm_build_rwkv6_base {
     llm_build_rwkv6(const llama_model & model, const llm_graph_params & params);
 };

src/models/rnd1.cpp

@@ -0,0 +1,126 @@
+#include "models.h"
+
+// RND1 is a Qwen3Moe AR model converted to diffusion model.
+llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Non-causal attention for diffusion
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+        cur = moe_out;
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

tests/test-backend-ops.cpp

@@ -39,6 +39,7 @@
 #include <string_view>
 #include <thread>
 #include <vector>
+#include <unordered_map>
 
 static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
     size_t nels = ggml_nelements(tensor);
@@ -269,6 +270,34 @@ static double nmse(const float * a, const float * b, size_t n) {
     return mse_a_b / mse_a_0;
 }
 
+// difference between 2 integer sets (Jaccard distance, 0 - no difference, 1 - no overlap)
+static double jdst(const int32_t * a, const int32_t * b, size_t n) {
+    std::unordered_map<int32_t, size_t> set_a;
+    std::unordered_map<int32_t, size_t> set_b;
+
+    for (size_t i = 0; i < n; ++i) {
+        set_a[a[i]]++;
+        set_b[b[i]]++;
+    }
+
+    size_t diff = 0;
+
+    for (const auto & p : set_a) {
+        const int64_t na = p.second;
+        const int64_t nb = set_b.find(p.first) != set_b.end() ? set_b.at(p.first) : 0;
+
+        diff += std::abs(na - nb);
+    }
+
+    for (const auto & p : set_b) {
+        if (set_a.find(p.first) == set_a.end()) {
+            diff += p.second;
+        }
+    }
+
+    return (double) diff / (2*n);
+}
+
 // maximum absolute asymmetry between a and b
 // asymmetry: (a - b) / (a + b)
 // This is more stable than relative error if one of the values fluctuates towards zero.
@@ -1051,6 +1080,14 @@ struct test_case {
         return 1e-4;
     }
 
+    virtual double max_err() {
+        return max_nmse_err();
+    }
+
+    virtual double err(const float * a, const float * b, size_t n) {
+        return nmse(a, b, n);
+    }
+
     virtual float grad_eps() {
         return 1e-1f;
     }
@@ -1257,16 +1294,16 @@ struct test_case {
         // compare
         struct callback_userdata {
             bool   ok;
-            double max_err;
+            test_case * tc;
             ggml_backend_t backend1;
             ggml_backend_t backend2;
         };
 
         callback_userdata ud {
             true,
-            max_nmse_err(),
+            this,
             backend1,
-            backend2
+            backend2,
         };
 
         auto callback = [](int index, ggml_tensor * t1, ggml_tensor * t2, void * user_data) -> bool {
@@ -1314,9 +1351,9 @@ struct test_case {
                 }
             }
 
-            double err = nmse(f1.data(), f2.data(), f1.size());
-            if (err > ud->max_err) {
-                printf("[%s] NMSE = %.9f > %.9f ", ggml_op_desc(t1), err, ud->max_err);
+            double err = ud->tc->err(f1.data(), f2.data(), f1.size());
+            if (err > ud->tc->max_err()) {
+                printf("[%s] ERR = %.9f > %.9f ", ggml_op_desc(t1), err, ud->tc->max_err());
                 //for (int i = 0; i < (int) f1.size(); i++) {
                 //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
                 //}
@@ -4943,7 +4980,71 @@ struct test_argsort : public test_case {
     }
 };
 
-struct test_topk_moe: public test_case {
+// GGML_OP_TOP_K
+struct test_top_k : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const int k;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, k);
+    }
+
+    test_top_k(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {16, 10, 10, 10},
+            int k = 4)
+        : type(type), ne(ne), k(k) {}
+
+    double max_err() override {
+        return 0.0;
+    }
+
+    double err(const float * a, const float * b, size_t n) override {
+        std::vector<int32_t> ia(n);
+        std::vector<int32_t> ib(n);
+
+        double diff = 0.0f;
+
+        for (size_t i = 0; i < n; i++) {
+            ia[i] = (int32_t) a[i];
+            ib[i] = (int32_t) b[i];
+
+            // penalize the result if the data is not integer valued
+            diff += std::fabs(a[i] - ia[i]);
+            diff += std::fabs(b[i] - ib[i]);
+        }
+
+        return diff + jdst(ia.data(), ib.data(), n);
+    }
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_top_k(ctx, a, k);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        std::random_device rd;
+        std::default_random_engine rng(rd());
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            // initialize with unique values to avoid ties
+            for (int64_t r = 0; r < ggml_nrows(t); r++) {
+                std::vector<float> data(t->ne[0]);
+                for (int i = 0; i < t->ne[0]; i++) {
+                    data[i] = i;
+                }
+                std::shuffle(data.begin(), data.end(), rng);
+                ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
+            }
+        }
+    }
+};
+
+struct test_topk_moe : public test_case {
     const std::array<int64_t, 4> ne;
     const int n_expert_used;
     const bool with_norm;
@@ -4976,7 +5077,7 @@ struct test_topk_moe: public test_case {
 
         ggml_tensor * logits = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne.data());
         ggml_tensor * probs            = delayed_softmax ? logits : ggml_soft_max(ctx, logits);
-        ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_expert_used); // [n_expert_used, n_tokens]
+        ggml_tensor * selected_experts = ggml_argsort_top_k(ctx, probs, n_expert_used); // [n_expert_used, n_tokens]
 
         ggml_tensor * out = ggml_get_rows(ctx, ggml_reshape_3d(ctx, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
 
@@ -6953,9 +7054,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_I32, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_I32, {256, 1, 4, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 1, 4, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
 
-    for (ggml_type type_dst : { GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16 }) {
+    for (ggml_type type_dst : { GGML_TYPE_F32, GGML_TYPE_I32, GGML_TYPE_F16, GGML_TYPE_BF16 }) {
         for (bool use_view_slice : { true, false }) {
             for (std::array<int64_t, 4> ne : std::initializer_list<std::array<int64_t, 4>>{ {2, 1, 1, 1}, {2, 1, 3, 5},
                 {2, 3, 5, 7}, {1, 4, 4, 1}, {1, 8, 17, 1}, {10, 10, 10, 1} }) {
@@ -7532,6 +7635,31 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {2, 8, 8192, 1}, order)); // bailingmoe2 (group selection)
     }
 
+    for (int i = 0; i < 20; ++i) {
+        for (int k : {1, 2, 3, 7, 15, 100, 500, 1023, 9999}) {
+            if (k <= 1<<i) {
+                test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {(1<<i), 1, 1, 1}, k));
+                test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {(1<<i) + 11, 1, 2, 1}, k));
+            }
+        }
+    }
+    for (int k : {1, 2, 3, 7, 15}) {
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {16, 10, 10, 10}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {60, 10, 10, 10}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {1023, 2, 1, 3}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {1024, 2, 1, 3}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {1025, 2, 1, 3}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {16384, 1, 1, 1}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {2047, 2, 1, 3}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {2048, 2, 1, 3}, k));
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {2049, 2, 1, 3}, k));
+    }
+
+    // exhaustive top_k tests
+    //for (int i = 1; i < 9999; ++i) {
+    //    test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {i, 2, 1, 3}, rand() % i + 1));
+    //}
+
     for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR, GGML_SCALE_MODE_BICUBIC}) {
         test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));
         test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode, true));
@@ -7819,6 +7947,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     for (int bs : {1, 4, 8, 32, 64, 128, 256, 512}) {
         for (ggml_type type_a : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_Q4_K, GGML_TYPE_Q6_K, GGML_TYPE_IQ2_XS}) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {
+                test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, 128, 8, false, 768, bs, 2048));
                 test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 128, 8, false, 768, bs, 2048, 1));
             }
         }
@@ -7827,6 +7956,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     for (int bs : {1, 4, 8, 32, 64, 128, 256, 512}) {
         for (ggml_type type_a : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_Q4_K, GGML_TYPE_Q6_K, GGML_TYPE_IQ2_XS}) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {
+                test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, 32, 4, false, 1792, bs, 2048));
                 test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 32, 4, false, 1792, bs, 2048, 1));
             }
         }
@@ -7837,6 +7967,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     for (int bs : {1, 4, 8, 512}) {
         for (ggml_type type_a : {GGML_TYPE_MXFP4}) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {
+                test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, 32, 4, false, 2880, bs, 2880));
                 test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 32, 4, false, 2880, bs, 2880, 1));
             }
         }
@@ -7854,6 +7985,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    // Qwen3-VL-8B https://github.com/ggml-org/llama.cpp/issues/17012
+    test_cases.emplace_back(new test_flash_attn_ext(72, 72, 16, {1, 1}, 5776, 5776, false, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+
     for (int kv : { 4096, 8192, 16384, }) {
         for (int hs : { 64, 128, }) {
             for (int nr : { 1, 4, }) {
@@ -7906,6 +8040,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     }
 
     test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {65000, 16, 1, 1}));
+    for (auto k : {1, 10, 40}) {
+        for (auto nrows : {1, 16}) {
+            for (auto cols : {k, 1000, 65000, 200000}) {
+                test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {cols, nrows, 1, 1}, k));
+            }
+        }
+    }
 
     return test_cases;
 }

tools/main/main.cpp

@@ -147,11 +147,15 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    auto * mem = llama_get_memory(ctx);
-
+    llama_memory_t mem = llama_get_memory(ctx);
     const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    // note: the time for chat template initialization is not negligible:
     auto chat_templates = common_chat_templates_init(model, params.chat_template);
 
+    // start measuring performance timings from here
+    llama_perf_context_reset(ctx);
+
     LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
 
     auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);

tools/mtmd/clip.cpp

@@ -1175,10 +1175,11 @@ struct clip_graph {
             cb(K, "resampler_K", -1);
             cb(V, "resampler_V", -1);
 
+            float resampler_kq_scale = 1.0f/ sqrtf(float(d_head));
             embeddings = build_attn(
                 model.mm_model_attn_o_w,
                 model.mm_model_attn_o_b,
-                Q, K, V, nullptr, kq_scale, -1);
+                Q, K, V, nullptr, resampler_kq_scale, -1);
             cb(embeddings, "resampler_attn_out", -1);
         }
         // layernorm

tools/server/CMakeLists.txt

@@ -13,9 +13,14 @@ endif()
 
 set(TARGET_SRCS
     server.cpp
-    utils.hpp
     server-http.cpp
     server-http.h
+    server-task.cpp
+    server-task.h
+    server-queue.cpp
+    server-queue.h
+    server-common.cpp
+    server-common.h
 )
 set(PUBLIC_ASSETS
     index.html.gz

tools/server/README.md

@@ -30,9 +30,10 @@ The project is under active development, and we are [looking for feedback and co
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
+| `-cl, --cache-list` | show list of models in cache |
 | `--completion-bash` | print source-able bash completion script for llama.cpp |
 | `--verbose-prompt` | print a verbose prompt before generation (default: false) |
-| `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
+| `-t, --threads N` | number of CPU threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
 | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
 | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
 | `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
@@ -51,7 +52,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
 | `--swa-full` | use full-size SWA cache (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)<br/>(env: LLAMA_ARG_SWA_FULL) |
 | `--kv-unified, -kvu` | use single unified KV buffer for the KV cache of all sequences (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)<br/>(env: LLAMA_ARG_KV_SPLIT) |
-| `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
+| `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `--no-perf` | disable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_NO_PERF) |
 | `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
 | `--no-escape` | do not process escape sequences |
@@ -61,11 +62,12 @@ The project is under active development, and we are [looking for feedback and co
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
 | `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
 | `-nr, --no-repack` | disable weight repacking<br/>(env: LLAMA_ARG_NO_REPACK) |
+| `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
@@ -78,7 +80,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--override-tensor, -ot <tensor name pattern>=<buffer type>,...` | override tensor buffer type |
 | `--cpu-moe, -cmoe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
 | `--n-cpu-moe, -ncmoe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
-| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM (default: -1)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
 | `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
@@ -92,6 +94,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
 | `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)<br/>(env: LLAMA_ARG_MODEL) |
 | `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
+| `-dr, --docker-repo [<repo>/]<model>[:quant]` | Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.<br/>example: gemma3<br/>(default: unused)<br/>(env: LLAMA_ARG_DOCKER_REPO) |
 | `-hf, -hfr, --hf-repo <user>/<model>[:quant]` | Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.<br/>mmproj is also downloaded automatically if available. to disable, add --no-mmproj<br/>example: unsloth/phi-4-GGUF:q4_k_m<br/>(default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
 | `-hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_HFD_REPO) |
 | `-hff, --hf-file FILE` | Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
@@ -100,7 +103,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
 | `--log-disable` | Log disable |
 | `--log-file FNAME` | Log to file |
-| `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |
+| `--log-colors [on\|off\|auto]` | Set colored logging ('on', 'off', or 'auto', default: 'auto')<br/>'auto' enables colors when output is to a terminal<br/>(env: LLAMA_LOG_COLORS) |
 | `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) |
 | `--offline` | Offline mode: forces use of cache, prevents network access<br/>(env: LLAMA_OFFLINE) |
 | `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.<br/>(env: LLAMA_LOG_VERBOSITY) |
@@ -151,7 +154,8 @@ The project is under active development, and we are [looking for feedback and co
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--swa-checkpoints N` | max number of SWA checkpoints per slot to create (default: 3)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_SWA_CHECKPOINTS) |
+| `--ctx-checkpoints, --swa-checkpoints N` | max number of context checkpoints to create per slot (default: 8)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_CTX_CHECKPOINTS) |
+| `--cache-ram, -cram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
 | `--no-context-shift` | disables context shift on infinite text generation (default: enabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
 | `--context-shift` | enables context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
 | `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode<br/> |
@@ -165,6 +169,8 @@ The project is under active development, and we are [looking for feedback and co
 | `--mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
 | `--no-mmproj` | explicitly disable multimodal projector, useful when using -hf<br/>(env: LLAMA_ARG_NO_MMPROJ) |
 | `--no-mmproj-offload` | do not offload multimodal projector to GPU<br/>(env: LLAMA_ARG_NO_MMPROJ_OFFLOAD) |
+| `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) |
+| `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) |
 | `--override-tensor-draft, -otd <tensor name pattern>=<buffer type>,...` | override tensor buffer type for draft model |
 | `--cpu-moe-draft, -cmoed` | keep all Mixture of Experts (MoE) weights in the CPU for the draft model<br/>(env: LLAMA_ARG_CPU_MOE_DRAFT) |
 | `--n-cpu-moe-draft, -ncmoed N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model<br/>(env: LLAMA_ARG_N_CPU_MOE_DRAFT) |
@@ -189,13 +195,14 @@ The project is under active development, and we are [looking for feedback and co
 | `--slots` | enable slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
 | `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
-| `--jinja` | use jinja template for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
-| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: deepseek)<br/>(env: LLAMA_ARG_THINK) |
+| `--jinja` | use jinja template for chat (default: enabled)<br/><br/>(env: LLAMA_ARG_JINJA) |
+| `--no-jinja` | disable jinja template for chat (default: enabled)<br/><br/>(env: LLAMA_ARG_NO_JINJA) |
+| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/><br/>(env: LLAMA_ARG_NO_PREFILL_ASSISTANT) |
-| `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
+| `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
 | `-td, --threads-draft N` | number of threads to use during generation (default: same as --threads) |
 | `-tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |
@@ -209,15 +216,17 @@ The project is under active development, and we are [looking for feedback and co
 | `--spec-replace TARGET DRAFT` | translate the string in TARGET into DRAFT if the draft model and main model are not compatible |
 | `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) |
 | `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall |
-| `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) |
-| `--embd-e5-small-en-default` | use default e5-small-v2 model (note: can download weights from the internet) |
-| `--embd-gte-small-default` | use default gte-small model (note: can download weights from the internet) |
+| `--embd-gemma-default` | use default EmbeddingGemma model (note: can download weights from the internet) |
 | `--fim-qwen-1.5b-default` | use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet) |
 | `--fim-qwen-3b-default` | use default Qwen 2.5 Coder 3B (note: can download weights from the internet) |
 | `--fim-qwen-7b-default` | use default Qwen 2.5 Coder 7B (note: can download weights from the internet) |
 | `--fim-qwen-7b-spec` | use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet) |
 | `--fim-qwen-14b-spec` | use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet) |
 | `--fim-qwen-30b-default` | use default Qwen 3 Coder 30B A3B Instruct (note: can download weights from the internet) |
+| `--gpt-oss-20b-default` | use gpt-oss-20b (note: can download weights from the internet) |
+| `--gpt-oss-120b-default` | use gpt-oss-120b (note: can download weights from the internet) |
+| `--vision-gemma-4b-default` | use Gemma 3 4B QAT (note: can download weights from the internet) |
+| `--vision-gemma-12b-default` | use Gemma 3 12B QAT (note: can download weights from the internet) |
 
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.