server : headers cleanup

.devops/vulkan.Dockerfile

@@ -1,7 +1,9 @@
-ARG UBUNTU_VERSION=26.04
+ARG UBUNTU_VERSION=25.10
 
 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
 
@@ -50,7 +52,6 @@ WORKDIR /app
 
 RUN apt-get update \
     && apt-get install -y \
-    build-essential \
     git \
     python3 \
     python3-pip \

.github/workflows/build.yml

@@ -69,6 +69,13 @@ 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: |
@@ -76,8 +83,6 @@ 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 \
@@ -105,6 +110,13 @@ 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: |
@@ -114,8 +126,6 @@ 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
@@ -141,6 +151,13 @@ 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: |
@@ -200,7 +217,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install -y --no-install-recommends \
             python3 python3-pip python3-dev \
-            libjpeg-dev build-essential libssl-dev \
+            libjpeg-dev build-essential libcurl4-openssl-dev \
             git-lfs
 
       - name: Python Dependencies
@@ -221,8 +238,6 @@ 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)
@@ -279,15 +294,13 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          sudo apt-get install build-essential libcurl4-openssl-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 }}
@@ -298,8 +311,6 @@ 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 }} \
@@ -324,7 +335,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -332,8 +343,6 @@ 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)
@@ -364,14 +373,12 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          sudo apt-get install build-essential libcurl4-openssl-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)
 
@@ -398,14 +405,12 @@ jobs:
       - name: Dependencies
         id: depends
         run: |
-          sudo apt-get install -y glslc libvulkan-dev libssl-dev
+          sudo apt-get install -y glslc libvulkan-dev libcurl4-openssl-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 \
@@ -435,7 +440,7 @@ 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 libssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -461,8 +466,6 @@ 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)
 
@@ -494,7 +497,7 @@ 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 libssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -534,10 +537,7 @@ jobs:
         id: cmake_build
         run: |
           export Dawn_DIR=dawn/lib64/cmake/Dawn
-          cmake -B build \
-            -DLLAMA_CURL=OFF \
-            -DLLAMA_OPENSSL=ON \
-            -DGGML_WEBGPU=ON
+          cmake -B build -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 libssl-dev rocwmma-dev
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev rocwmma-dev
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -572,8 +572,6 @@ 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
@@ -592,7 +590,7 @@ jobs:
         id: depends
         run: |
           apt-get update
-          apt-get install -y build-essential git cmake libssl-dev
+          apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -604,8 +602,6 @@ 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)
 
@@ -630,7 +626,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
 
       - name: install oneAPI MKL library
         shell: bash
@@ -652,8 +648,6 @@ 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
@@ -680,7 +674,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
 
       - name: install oneAPI MKL library
         shell: bash
@@ -702,8 +696,6 @@ 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 \
@@ -730,6 +722,12 @@ 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: |
@@ -761,6 +759,12 @@ 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: |
@@ -786,6 +790,12 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
       - name: Build
         id: cmake_build
         run: |
@@ -828,6 +838,12 @@ 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: |
@@ -979,12 +995,21 @@ 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 ${{ matrix.defines }} `
-            -DLLAMA_CURL=OFF -DLLAMA_BUILD_BORINGSSL=ON
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           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
@@ -1028,7 +1053,7 @@ jobs:
             DEBIAN_FRONTEND: noninteractive
           run: |
               apt update
-              apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
+              apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
 
         - name: ccache
           uses: ggml-org/ccache-action@v1.2.16
@@ -1039,12 +1064,10 @@ 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
@@ -1078,20 +1101,25 @@ 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
+            -DGGML_RPC=ON ^
+            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include"
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
@@ -1123,7 +1151,7 @@ jobs:
         run:  |
           scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
-      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
+      # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
 
       - name: Build
         id: cmake_build
@@ -1180,8 +1208,14 @@ 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}"
@@ -1190,12 +1224,11 @@ 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="${env:HIP_PATH}" `
             -DGGML_HIP=ON `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON `
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
   ios-xcode-build:
@@ -1357,10 +1390,14 @@ jobs:
     strategy:
       matrix:
         arch: [x86, aarch64]
-        chip_type: ['910b', '310p']
-        build: ['Release']
+        cann:
+          - '8.3.rc1.alpha001-910b-openeuler22.03-py3.11'
+        chip_type:
+          - '910b'
+        build:
+          - 'Release'
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.2.rc1-310p-openeuler22.03-py3.11' }}
+    container: ascendai/cann:${{ matrix.cann }}
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/release.yml

@@ -698,9 +698,10 @@ jobs:
       matrix:
         arch: [x86, aarch64]
         chip_type: ['910b', '310p']
-        build: ['Release']
+        build:
+          - 'Release'
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.2.rc1-310p-openeuler22.03-py3.11' }}
+    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.3.rc1.alpha001-310p-openeuler22.03-py3.11' }}
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -736,7 +737,7 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.zip
-          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.zip
+          name: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

.github/workflows/server.yml

@@ -56,7 +56,7 @@ jobs:
             curl \
             wget \
             language-pack-en \
-            libssl-dev
+            libcurl4-openssl-dev
 
       - name: Clone
         id: checkout
@@ -242,7 +242,7 @@ jobs:
             curl \
             wget \
             language-pack-en \
-            libssl-dev
+            libcurl4-openssl-dev
 
       - name: Clone
         id: checkout
@@ -283,8 +283,6 @@ 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 \
@@ -297,8 +295,6 @@ 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 ;
@@ -310,8 +306,6 @@ 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
@@ -351,10 +345,16 @@ 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: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -B build -DLLAMA_CURL=OFF -DLLAMA_BUILD_BORINGSSL=ON
+          cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
@@ -368,6 +368,13 @@ jobs:
         run: |
           pip install -r tools/server/tests/requirements.txt
 
+      - name: Copy Libcurl
+        id: prepare_libcurl
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
+
       - name: Tests
         id: server_integration_tests
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}

.gitignore

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

CODEOWNERS

@@ -2,8 +2,10 @@
 # multiplie collaborators per item can be specified
 
 /.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @CISC
+/.github/actions/                       @slaren @CISC
 /.github/workflows/                     @CISC
+/.github/workflows/release.yml          @slaren
+/.github/workflows/winget.yml           @slaren
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
 /common/CMakeLists.txt                  @ggerganov
@@ -38,14 +40,21 @@
 /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
-/ggml/src/ggml-common.h                 @ggerganov
-/ggml/src/ggml-cpu/                     @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/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
@@ -53,19 +62,19 @@
 /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
+/ggml/src/ggml-impl.h                   @ggerganov @slaren
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
 /ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov
-/ggml/src/ggml-threading.*              @ggerganov
+/ggml/src/ggml-threading.*              @ggerganov @slaren
 /ggml/src/ggml-vulkan/                  @0cc4m
 /ggml/src/ggml-webgpu/                  @reeselevine
 /ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml.c                        @ggerganov
-/ggml/src/ggml.cpp                      @ggerganov
+/ggml/src/ggml.c                        @ggerganov @slaren
+/ggml/src/ggml.cpp                      @ggerganov @slaren
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
 /gguf-py/                               @CISC
 /media/                                 @ggerganov
@@ -77,11 +86,15 @@
 /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
@@ -93,6 +106,8 @@
 /tools/tokenize/                        @ggerganov
 /tools/tts/                             @ggerganov
 /vendor/                                @ggerganov
+/.clang-format                          @slaren
+/.clang-tidy                            @slaren
 /AUTHORS                                @ggerganov
 /CMakeLists.txt                         @ggerganov
 /CONTRIBUTING.md                        @ggerganov

README.md

@@ -242,7 +242,6 @@ 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>
 

ci/run.sh

@@ -45,7 +45,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_SCHED_NO_REALLOC=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
@@ -428,10 +428,10 @@ function gg_run_qwen3_0_6b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -523,8 +523,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
@@ -564,7 +564,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

common/CMakeLists.txt

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

common/arg.cpp

@@ -694,12 +694,6 @@ 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();
 
@@ -980,7 +974,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.kv_unified = true;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED"));
+    ).set_env("LLAMA_ARG_KV_SPLIT"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1238,7 +1232,6 @@ 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(
@@ -1268,7 +1261,6 @@ 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(
@@ -1276,7 +1268,6 @@ 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(
@@ -1284,7 +1275,6 @@ 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(
@@ -1292,7 +1282,6 @@ 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(
@@ -1307,7 +1296,6 @@ 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(
@@ -1315,7 +1303,6 @@ 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(
@@ -1334,7 +1321,6 @@ 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(
@@ -1342,7 +1328,6 @@ 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(
@@ -1440,7 +1425,6 @@ 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(
@@ -1448,7 +1432,6 @@ 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(
@@ -1456,7 +1439,6 @@ 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(
@@ -2494,18 +2476,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--jinja"},
-        string_format("use jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
+        "use jinja template for chat (default: 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/chat-parser-xml-toolcall.cpp

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

common/chat-parser-xml-toolcall.h

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

common/chat-parser.cpp

@@ -13,120 +13,6 @@
 
 using json = nlohmann::ordered_json;
 
-static void parse_prefixed_json_tool_call_array(common_chat_msg_parser & builder,
-                                                const common_regex &     prefix,
-                                                size_t                   rstrip_prefix = 0) {
-    static const std::vector<std::vector<std::string>> args_paths = { { "arguments" } };
-    if (auto res = builder.try_find_regex(prefix)) {
-        builder.move_back(rstrip_prefix);
-        auto tool_calls = builder.consume_json_with_dumped_args(args_paths);
-        if (!builder.add_tool_calls(tool_calls.value) || tool_calls.is_partial) {
-            throw common_chat_msg_partial_exception("incomplete tool call array");
-        }
-    } else {
-        builder.add_content(builder.consume_rest());
-    }
-}
-
-static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
-    std::string arguments;
-    if (builder.is_partial()) {
-        arguments = (json{
-                         { "code", code + builder.healing_marker() }
-        })
-                        .dump();
-        auto idx = arguments.find(builder.healing_marker());
-        if (idx != std::string::npos) {
-            arguments.resize(idx);
-        }
-    } else {
-        arguments = (json{
-                         { "code", code }
-        })
-                        .dump();
-    }
-    return arguments;
-}
-
-/**
- * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between.
- * Aggregates the prefix, suffix and in-between text into the content.
- */
-static void parse_json_tool_calls(
-    common_chat_msg_parser &            builder,
-    const std::optional<common_regex> & block_open,
-    const std::optional<common_regex> & function_regex_start_only,
-    const std::optional<common_regex> & function_regex,
-    const common_regex &                close_regex,
-    const std::optional<common_regex> & block_close,
-    bool                                allow_raw_python = false,
-    const std::function<std::string(const common_chat_msg_parser::find_regex_result & fres)> & get_function_name =
-        nullptr) {
-    auto parse_tool_calls = [&]() {
-        size_t from  = std::string::npos;
-        auto   first = true;
-        while (true) {
-            auto start_pos = builder.pos();
-            auto res = function_regex_start_only && first ? builder.try_consume_regex(*function_regex_start_only) :
-                       function_regex                     ? builder.try_find_regex(*function_regex, from) :
-                                                            std::nullopt;
-
-            if (res) {
-                std::string name;
-                if (get_function_name) {
-                    name = get_function_name(*res);
-                } else {
-                    GGML_ASSERT(res->groups.size() == 2);
-                    name = builder.str(res->groups[1]);
-                }
-                first = false;
-                if (name.empty()) {
-                    // get_function_name signalled us that we should skip this match and treat it as content.
-                    from = res->groups[0].begin + 1;
-                    continue;
-                }
-                from = std::string::npos;
-
-                auto maybe_raw_python = name == "python" && allow_raw_python;
-                if (builder.input()[builder.pos()] == '{' || !maybe_raw_python) {
-                    if (auto arguments = builder.try_consume_json_with_dumped_args({ {} })) {
-                        if (!builder.add_tool_call(name, "", arguments->value) || arguments->is_partial) {
-                            throw common_chat_msg_partial_exception("incomplete tool call");
-                        }
-                        builder.consume_regex(close_regex);
-                    }
-                    continue;
-                }
-                if (maybe_raw_python) {
-                    auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
-                    if (!builder.add_tool_call(name, "", arguments)) {
-                        throw common_chat_msg_partial_exception("incomplete tool call");
-                    }
-                    return;
-                }
-                throw common_chat_msg_partial_exception("incomplete tool call");
-            } else {
-                builder.move_to(start_pos);
-            }
-            break;
-        }
-        if (block_close) {
-            builder.consume_regex(*block_close);
-        }
-        builder.consume_spaces();
-        builder.add_content(builder.consume_rest());
-    };
-    if (block_open) {
-        if (auto res = builder.try_find_regex(*block_open)) {
-            parse_tool_calls();
-        } else {
-            builder.add_content(builder.consume_rest());
-        }
-    } else {
-        parse_tool_calls();
-    }
-}
-
 common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
     : input_(input), is_partial_(is_partial), syntax_(syntax)
 {
@@ -646,857 +532,3 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
 void common_chat_msg_parser::clear_tools() {
     result_.tool_calls.clear();
 }
-
-/**
- * All common_chat_parse_* moved from chat.cpp to chat-parser.cpp below
- * to reduce incremental compile time for parser changes.
- */
-static void common_chat_parse_generic(common_chat_msg_parser & builder) {
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-    static const std::vector<std::vector<std::string>> content_paths = {
-        {"response"},
-    };
-    static const std::vector<std::vector<std::string>> args_paths = {
-        {"tool_call", "arguments"},
-        {"tool_calls", "arguments"},
-    };
-    auto data = builder.consume_json_with_dumped_args(args_paths, content_paths);
-    if (data.value.contains("tool_calls")) {
-        if (!builder.add_tool_calls(data.value.at("tool_calls")) || data.is_partial) {
-            throw common_chat_msg_partial_exception("incomplete tool calls");
-        }
-    } else if (data.value.contains("tool_call")) {
-        if (!builder.add_tool_call(data.value.at("tool_call")) || data.is_partial) {
-            throw common_chat_msg_partial_exception("incomplete tool call");
-        }
-    } else if (data.value.contains("response")) {
-        const auto & response = data.value.at("response");
-        builder.add_content(response.is_string() ? response.template get<std::string>() : response.dump(2));
-        if (data.is_partial) {
-            throw common_chat_msg_partial_exception("incomplete response");
-        }
-    } else {
-        throw common_chat_msg_partial_exception("Expected 'tool_call', 'tool_calls' or 'response' in JSON");
-    }
-}
-
-static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
-    parse_prefixed_json_tool_call_array(builder, prefix);
-}
-
-static void common_chat_parse_magistral(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("[THINK]", "[/THINK]");
-
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
-    parse_prefixed_json_tool_call_array(builder, prefix);
-}
-
-static void common_chat_parse_command_r7b(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("<|START_THINKING|>", "<|END_THINKING|>");
-
-    static const common_regex start_action_regex("<\\|START_ACTION\\|>");
-    static const common_regex end_action_regex("<\\|END_ACTION\\|>");
-    static const common_regex start_response_regex("<\\|START_RESPONSE\\|>");
-    static const common_regex end_response_regex("<\\|END_RESPONSE\\|>");
-
-    if (auto res = builder.try_find_regex(start_action_regex)) {
-        // If we didn't extract thoughts, prelude includes them.
-        auto tool_calls = builder.consume_json_with_dumped_args({{"parameters"}});
-        for (const auto & tool_call : tool_calls.value) {
-            std::string name = tool_call.contains("tool_name") ? tool_call.at("tool_name") : "";
-            std::string id = tool_call.contains("tool_call_id") ? tool_call.at("tool_call_id") : "";
-            std::string arguments = tool_call.contains("parameters") ? tool_call.at("parameters") : "";
-            if (!builder.add_tool_call(name, id, arguments) || tool_calls.is_partial) {
-                throw common_chat_msg_partial_exception("incomplete tool call");
-            }
-        }
-        if (tool_calls.is_partial) {
-            throw common_chat_msg_partial_exception("incomplete tool call");
-        }
-        builder.consume_regex(end_action_regex);
-    } else if (auto res = builder.try_find_regex(start_response_regex)) {
-        if (!builder.try_find_regex(end_response_regex)) {
-            builder.add_content(builder.consume_rest());
-            throw common_chat_msg_partial_exception(end_response_regex.str());
-        }
-    } else {
-        builder.add_content(builder.consume_rest());
-    }
-}
-
-static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
-    builder.try_parse_reasoning("<think>", "</think>");
-
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    static const common_regex function_regex(
-        "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: ");
-    static const common_regex close_regex("\\}\\s*");
-
-    static const common_regex function_name_regex("\\s*(\\w+)\\s*\\.\\s*call\\(");
-    static const common_regex arg_name_regex("\\s*(\\w+)\\s*=\\s*");
-
-    if (with_builtin_tools) {
-        static const common_regex builtin_call_regex("<\\|python_tag\\|>");
-        if (auto res = builder.try_find_regex(builtin_call_regex)) {
-            auto fun_res = builder.consume_regex(function_name_regex);
-            auto function_name = builder.str(fun_res.groups[1]);
-
-            common_healing_marker healing_marker;
-            json args = json::object();
-            while (true) {
-                if (auto arg_res = builder.try_consume_regex(arg_name_regex)) {
-                    auto arg_name = builder.str(arg_res->groups[1]);
-                    auto partial = builder.consume_json();
-                    args[arg_name] = partial.json;
-                    healing_marker.marker = partial.healing_marker.marker;
-                    healing_marker.json_dump_marker = partial.healing_marker.json_dump_marker;
-                    builder.consume_spaces();
-                    if (!builder.try_consume_literal(",")) {
-                        break;
-                    }
-                } else {
-                    break;
-                }
-            }
-            builder.consume_literal(")");
-            builder.consume_spaces();
-
-            auto arguments = args.dump();
-            if (!builder.add_tool_call(function_name, "", arguments)) {
-                throw common_chat_msg_partial_exception("Incomplete tool call");
-            }
-            return;
-        }
-    }
-    parse_json_tool_calls(
-        builder,
-        /* block_open= */ std::nullopt,
-        /* function_regex_start_only= */ function_regex,
-        /* function_regex= */ std::nullopt,
-        close_regex,
-        std::nullopt);
-
-}
-
-static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("<think>", "</think>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
-    static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
-    static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?function<｜tool▁sep｜>([^\n]+)\n```json\n");
-    static const common_regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
-
-    parse_json_tool_calls(
-        builder,
-        /* block_open= */ tool_calls_begin,
-        /* function_regex_start_only= */ std::nullopt,
-        function_regex,
-        close_regex,
-        tool_calls_end);
-}
-
-static void common_chat_parse_deepseek_v3_1_content(common_chat_msg_parser & builder) {
-    static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?([^\\n<]+)(?:<｜tool▁sep｜>)");
-
-    static const common_regex close_regex("(?:[\\s]*)?<｜tool▁call▁end｜>");
-    static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
-    static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
-
-    if (!builder.syntax().parse_tool_calls) {
-        LOG_DBG("%s: not parse_tool_calls\n", __func__);
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    LOG_DBG("%s: parse_tool_calls\n", __func__);
-
-    parse_json_tool_calls(
-        builder,
-        /* block_open= */ tool_calls_begin,
-        /* function_regex_start_only= */ std::nullopt,
-        function_regex,
-        close_regex,
-        tool_calls_end);
-}
-
-static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
-    // DeepSeek V3.1 outputs reasoning content between "<think>" and "</think>" tags, followed by regular content
-    // First try to parse using the standard reasoning parsing method
-    LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str());
-
-    auto start_pos = builder.pos();
-    auto found_end_think = builder.try_find_literal("</think>");
-    builder.move_to(start_pos);
-
-    if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) {
-        LOG_DBG("%s: no end_think, not partial, adding content\n", __func__);
-        common_chat_parse_deepseek_v3_1_content(builder);
-    } else if (builder.try_parse_reasoning("<think>", "</think>")) {
-        // If reasoning was parsed successfully, the remaining content is regular content
-        LOG_DBG("%s: parsed reasoning, adding content\n", __func__);
-        // </think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\n```json\nJSON\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>
-        common_chat_parse_deepseek_v3_1_content(builder);
-    } else {
-        if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) {
-          LOG_DBG("%s: reasoning_format none, adding content\n", __func__);
-          common_chat_parse_deepseek_v3_1_content(builder);
-          return;
-        }
-        // If no reasoning tags found, check if we should treat everything as reasoning
-        if (builder.syntax().thinking_forced_open) {
-            // If thinking is forced open but no tags found, treat everything as reasoning
-            LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__);
-            builder.add_reasoning_content(builder.consume_rest());
-        } else {
-            LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__);
-            // <｜tool▁call▁begin｜>NAME<｜tool▁sep｜>JSON<｜tool▁call▁end｜>
-            common_chat_parse_deepseek_v3_1_content(builder);
-        }
-    }
-}
-
-static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form {
-        /* form.scope_start = */ "<minimax:tool_call>",
-        /* form.tool_start  = */ "<invoke name=\"",
-        /* form.tool_sep    = */ "\">",
-        /* form.key_start   = */ "<parameter name=\"",
-        /* form.key_val_sep = */ "\">",
-        /* form.val_end     = */ "</parameter>",
-        /* form.tool_end    = */ "</invoke>",
-        /* form.scope_end   = */ "</minimax:tool_call>",
-    };
-    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
-}
-
-static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "<tool_call>";
-        form.tool_start  = "<function=";
-        form.tool_sep    = ">";
-        form.key_start   = "<parameter=";
-        form.key_val_sep = ">";
-        form.val_end     = "</parameter>";
-        form.tool_end    = "</function>";
-        form.scope_end   = "</tool_call>";
-        form.trim_raw_argval = true;
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form);
-}
-
-static void common_chat_parse_kimi_k2(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "<|tool_calls_section_begin|>";
-        form.tool_start  = "<|tool_call_begin|>";
-        form.tool_sep    = "<|tool_call_argument_begin|>{";
-        form.key_start   = "\"";
-        form.key_val_sep = "\": ";
-        form.val_end     = ", ";
-        form.tool_end    = "}<|tool_call_end|>";
-        form.scope_end   = "<|tool_calls_section_end|>";
-        form.raw_argval  = false;
-        form.last_val_end = "";
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
-}
-
-static void common_chat_parse_apriel_1_5(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "<tool_calls>[";
-        form.tool_start  = "{\"name\": \"";
-        form.tool_sep    = "\", \"arguments\": {";
-        form.key_start   = "\"";
-        form.key_val_sep = "\": ";
-        form.val_end     = ", ";
-        form.tool_end    = "}, ";
-        form.scope_end   = "]</tool_calls>";
-        form.raw_argval  = false;
-        form.last_val_end = "";
-        form.last_tool_end = "}";
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form, "<thinking>", "</thinking>");
-}
-
-static void common_chat_parse_xiaomi_mimo(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "";
-        form.tool_start  = "<tool_call>\n{\"name\": \"";
-        form.tool_sep    = "\", \"arguments\": {";
-        form.key_start   = "\"";
-        form.key_val_sep = "\": ";
-        form.val_end     = ", ";
-        form.tool_end    = "}\n</tool_call>";
-        form.scope_end   = "";
-        form.raw_argval  = false;
-        form.last_val_end = "";
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form);
-}
-
-static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
-    static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))";
-    static const std::string recipient("(?: to=functions\\.([^<\\s]+))");
-
-    static const common_regex start_regex("<\\|start\\|>assistant");
-    static const common_regex analysis_regex("<\\|channel\\|>analysis");
-    static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?");
-    static const common_regex preamble_regex("<\\|channel\\|>commentary");
-    static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?");
-    static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?");
-
-    auto consume_end = [&](bool include_end = false) {
-        if (auto res = builder.try_find_literal("<|end|>")) {
-            return res->prelude + (include_end ? builder.str(res->groups[0]) : "");
-        }
-        return builder.consume_rest();
-    };
-
-    auto handle_tool_call = [&](const std::string & name) {
-        if (auto args = builder.try_consume_json_with_dumped_args({{}})) {
-            if (builder.syntax().parse_tool_calls) {
-                if (!builder.add_tool_call(name, "", args->value) || args->is_partial) {
-                    throw common_chat_msg_partial_exception("incomplete tool call");
-                }
-            } else if (args->is_partial) {
-                throw common_chat_msg_partial_exception("incomplete tool call");
-            }
-        }
-    };
-
-    auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional<common_regex_match> {
-        auto match = regex.search(input, 0, true);
-        if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) {
-            return match;
-        }
-        return std::nullopt;
-    };
-
-    do {
-        auto header_start_pos = builder.pos();
-        auto content_start = builder.try_find_literal("<|message|>");
-        if (!content_start) {
-            throw common_chat_msg_partial_exception("incomplete header");
-        }
-
-        auto header = content_start->prelude;
-
-        if (auto match = regex_match(tool_call1_regex, header)) {
-            auto group = match->groups[1];
-            auto name = header.substr(group.begin, group.end - group.begin);
-            handle_tool_call(name);
-            continue;
-        }
-
-        if (auto match = regex_match(tool_call2_regex, header)) {
-            auto group = match->groups[2];
-            auto name = header.substr(group.begin, group.end - group.begin);
-            handle_tool_call(name);
-            continue;
-        }
-
-        if (regex_match(analysis_regex, header)) {
-            builder.move_to(header_start_pos);
-            if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
-                builder.add_content(consume_end(true));
-            } else {
-                builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>");
-            }
-            continue;
-        }
-
-        if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) {
-            builder.add_content(consume_end());
-            continue;
-        }
-
-        // Possibly a malformed message, attempt to recover by rolling
-        // back to pick up the next <|start|>
-        LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str());
-        builder.move_to(header_start_pos);
-    } while (builder.try_find_regex(start_regex, std::string::npos, false));
-
-    auto remaining = builder.consume_rest();
-    if (!remaining.empty()) {
-        LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str());
-    }
-}
-
-static void common_chat_parse_glm_4_5(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form {
-        /* form.scope_start  = */ "",
-        /* form.tool_start   = */ "<tool_call>",
-        /* form.tool_sep     = */ "",
-        /* form.key_start    = */ "<arg_key>",
-        /* form.key_val_sep  = */ "</arg_key>",
-        /* form.val_end      = */ "</arg_value>",
-        /* form.tool_end     = */ "</tool_call>",
-        /* form.scope_end    = */ "",
-        /* form.key_val_sep2 = */ "<arg_value>",
-    };
-    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
-}
-
-static void common_chat_parse_firefunction_v2(common_chat_msg_parser & builder) {
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-    static const common_regex prefix(regex_escape(" functools["));
-    parse_prefixed_json_tool_call_array(builder, prefix, /* rstrip_prefix= */ 1);
-}
-
-static void common_chat_parse_functionary_v3_2(common_chat_msg_parser & builder) {
-    static const common_regex function_regex_start_only(R"((\w+\n\{|python\n|all\n))");
-    static const common_regex function_regex(R"(>>>(\w+\n\{|python\n|all\n))");
-    static const common_regex close_regex(R"(\s*)");
-
-    parse_json_tool_calls(
-        builder,
-        std::nullopt,
-        function_regex_start_only,
-        function_regex,
-        close_regex,
-        std::nullopt,
-        /* allow_raw_python= */ true,
-        /* get_function_name= */ [&](const auto & res) -> std::string {
-            auto at_start = res.groups[0].begin == 0;
-            auto name = builder.str(res.groups[1]);
-            if (!name.empty() && name.back() == '{') {
-                // Unconsume the opening brace '{' to ensure the JSON parsing goes well.
-                builder.move_back(1);
-            }
-            auto idx = name.find_last_not_of("\n{");
-            name = name.substr(0, idx + 1);
-            if (at_start && name == "all") {
-                return "";
-            }
-            return name;
-        });
-}
-
-static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser & builder) {
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-    // This version of Functionary still supports the llama 3.1 tool call format for the python tool.
-    static const common_regex python_tag_regex(regex_escape("<|python_tag|>"));
-
-    static const common_regex function_regex(R"(<function=(\w+)>)");
-    static const common_regex close_regex(R"(</function>)");
-
-    parse_json_tool_calls(
-        builder,
-        /* block_open= */ std::nullopt,
-        /* function_regex_start_only= */ std::nullopt,
-        function_regex,
-        close_regex,
-        std::nullopt);
-
-    if (auto res = builder.try_find_regex(python_tag_regex)) {
-        auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
-        builder.add_tool_call("python", "", arguments);
-        return;
-    }
-}
-
-static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("<think>", "</think>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    static const common_regex open_regex(
-        "(?:"
-            "(```(?:xml|json)?\\n\\s*)?" // match 1 (block_start)
-            "("                          // match 2 (open_tag)
-                "<tool_call>"
-                "|<function_call>"
-                "|<tool>"
-                "|<tools>"
-                "|<response>"
-                "|<json>"
-                "|<xml>"
-                "|<JSON>"
-            ")?"
-            "(\\s*\\{\\s*\"name\")" // match 3 (named tool call)
-        ")"
-        "|<function=([^>]+)>"            // match 4 (function name)
-        "|<function name=\"([^\"]+)\">"  // match 5 (function name again)
-    );
-
-    while (auto res = builder.try_find_regex(open_regex)) {
-        const auto & block_start = res->groups[1];
-        std::string block_end = block_start.empty() ? "" : "```";
-
-        const auto & open_tag = res->groups[2];
-        std::string close_tag;
-
-        if (!res->groups[3].empty()) {
-            builder.move_to(res->groups[3].begin);
-            close_tag = open_tag.empty() ? "" : "</" + builder.str(open_tag).substr(1);
-
-            if (auto tool_call = builder.try_consume_json_with_dumped_args({{"arguments"}})) {
-                if (!builder.add_tool_call(tool_call->value) || tool_call->is_partial) {
-                    throw common_chat_msg_partial_exception("incomplete tool call");
-                }
-                builder.consume_spaces();
-                builder.consume_literal(close_tag);
-                builder.consume_spaces();
-                if (!block_end.empty()) {
-                    builder.consume_literal(block_end);
-                    builder.consume_spaces();
-                }
-            } else {
-                throw common_chat_msg_partial_exception("failed to parse tool call");
-            }
-        } else {
-            auto function_name = builder.str(res->groups[4]);
-            if (function_name.empty()) {
-                function_name = builder.str(res->groups[5]);
-            }
-            GGML_ASSERT(!function_name.empty());
-
-            close_tag = "</function>";
-
-            if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) {
-                if (!builder.add_tool_call(function_name, "", arguments->value) || arguments->is_partial) {
-                    throw common_chat_msg_partial_exception("incomplete tool call");
-                }
-                builder.consume_spaces();
-                builder.consume_literal(close_tag);
-                builder.consume_spaces();
-                if (!block_end.empty()) {
-                    builder.consume_literal(block_end);
-                    builder.consume_spaces();
-                }
-            }
-        }
-    }
-
-    builder.add_content(builder.consume_rest());
-}
-
-static void common_chat_parse_granite(common_chat_msg_parser & builder) {
-    // Parse thinking tags
-    static const common_regex start_think_regex(regex_escape("<think>"));
-    static const common_regex end_think_regex(regex_escape("</think>"));
-    // Granite models output partial tokens such as "<" and "<think".
-    // By leveraging try_consume_regex()/try_find_regex() throwing
-    // common_chat_msg_partial_exception for these partial tokens,
-    // processing is interrupted and the tokens are not passed to add_content().
-    if (auto res = builder.try_consume_regex(start_think_regex)) {
-        // Restore position for try_parse_reasoning()
-        builder.move_to(res->groups[0].begin);
-        builder.try_find_regex(end_think_regex, std::string::npos, false);
-        // Restore position for try_parse_reasoning()
-        builder.move_to(res->groups[0].begin);
-    }
-    builder.try_parse_reasoning("<think>", "</think>");
-
-    // Parse response tags
-    static const common_regex start_response_regex(regex_escape("<response>"));
-    static const common_regex end_response_regex(regex_escape("</response>"));
-    // Granite models output partial tokens such as "<" and "<response".
-    // Same hack as reasoning parsing.
-    if (builder.try_consume_regex(start_response_regex)) {
-        builder.try_find_regex(end_response_regex);
-    }
-
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // Look for tool calls
-    static const common_regex tool_call_regex(regex_escape("<|tool_call|>"));
-    if (auto res = builder.try_find_regex(tool_call_regex)) {
-        builder.move_to(res->groups[0].end);
-
-        // Expect JSON array of tool calls
-        if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) {
-            if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) {
-                throw common_chat_msg_partial_exception("incomplete tool call");
-            }
-        }
-    } else {
-        builder.add_content(builder.consume_rest());
-    }
-}
-
-static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) {
-    // Parse thinking tags
-    builder.try_parse_reasoning("<think>", "</think>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // Look for tool calls
-    static const common_regex tool_call_regex(regex_escape("<TOOLCALL>"));
-    if (auto res = builder.try_find_regex(tool_call_regex)) {
-        builder.move_to(res->groups[0].end);
-
-        // Expect JSON array of tool calls
-        auto tool_calls_data = builder.consume_json();
-        if (tool_calls_data.json.is_array()) {
-            if (!builder.try_consume_literal("</TOOLCALL>")) {
-                throw common_chat_msg_partial_exception("Incomplete tool call");
-            }
-            builder.add_tool_calls(tool_calls_data.json);
-        } else {
-            throw common_chat_msg_partial_exception("Incomplete tool call");
-        }
-    }
-    builder.add_content(builder.consume_rest());
-}
-
-static void common_chat_parse_apertus(common_chat_msg_parser & builder) {
-    // Parse thinking tags
-    builder.try_parse_reasoning("<|inner_prefix|>", "<|inner_suffix|>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // Look for tool calls
-    static const common_regex tool_call_regex(regex_escape("<|tools_prefix|>"));
-    if (auto res = builder.try_find_regex(tool_call_regex)) {
-        builder.move_to(res->groups[0].end);
-
-        auto tool_calls_data = builder.consume_json();
-        if (tool_calls_data.json.is_array()) {
-            builder.consume_spaces();
-            if (!builder.try_consume_literal("<|tools_suffix|>")) {
-                throw common_chat_msg_partial_exception("Incomplete tool call");
-            }
-            for (const auto & value : tool_calls_data.json) {
-                if (value.is_object()) {
-                    builder.add_tool_call_short_form(value);
-                }
-            }
-        } else {
-            throw common_chat_msg_partial_exception("Incomplete tool call");
-        }
-    }
-    builder.add_content(builder.consume_rest());
-}
-
-
-static void common_chat_parse_lfm2(common_chat_msg_parser & builder) {
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
-    }
-
-    // LFM2 format: <|tool_call_start|>[{"name": "get_current_time", "arguments": {"location": "Paris"}}]<|tool_call_end|>
-    static const common_regex tool_call_start_regex(regex_escape("<|tool_call_start|>"));
-    static const common_regex tool_call_end_regex(regex_escape("<|tool_call_end|>"));
-
-    // Loop through all tool calls
-    while (auto res = builder.try_find_regex(tool_call_start_regex, std::string::npos, /* add_prelude_to_content= */ true)) {
-        builder.move_to(res->groups[0].end);
-
-        // Parse JSON array format: [{"name": "...", "arguments": {...}}]
-        auto tool_calls_data = builder.consume_json();
-
-        // Consume end marker
-        builder.consume_spaces();
-        if (!builder.try_consume_regex(tool_call_end_regex)) {
-            throw common_chat_msg_partial_exception("Expected <|tool_call_end|>");
-        }
-
-        // Process each tool call in the array
-        if (tool_calls_data.json.is_array()) {
-            for (const auto & tool_call : tool_calls_data.json) {
-                if (!tool_call.is_object()) {
-                    throw common_chat_msg_partial_exception("Tool call must be an object");
-                }
-
-                if (!tool_call.contains("name")) {
-                    throw common_chat_msg_partial_exception("Tool call missing 'name' field");
-                }
-
-                std::string function_name = tool_call.at("name");
-                std::string arguments = "{}";
-
-                if (tool_call.contains("arguments")) {
-                    if (tool_call.at("arguments").is_object()) {
-                        arguments = tool_call.at("arguments").dump();
-                    } else if (tool_call.at("arguments").is_string()) {
-                        arguments = tool_call.at("arguments");
-                    }
-                }
-
-                if (!builder.add_tool_call(function_name, "", arguments)) {
-                    throw common_chat_msg_partial_exception("Incomplete tool call");
-                }
-            }
-        } else {
-            throw common_chat_msg_partial_exception("Expected JSON array for tool calls");
-        }
-
-        // Consume any trailing whitespace after this tool call
-        builder.consume_spaces();
-    }
-
-    // Consume any remaining content after all tool calls
-    auto remaining = builder.consume_rest();
-    if (!string_strip(remaining).empty()) {
-        builder.add_content(remaining);
-    }
-}
-
-static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form {
-        /* form.scope_start = */ "<seed:tool_call>",
-        /* form.tool_start  = */ "<function=",
-        /* form.tool_sep    = */ ">",
-        /* form.key_start   = */ "<parameter=",
-        /* form.key_val_sep = */ ">",
-        /* form.val_end     = */ "</parameter>",
-        /* form.tool_end    = */ "</function>",
-        /* form.scope_end   = */ "</seed:tool_call>",
-    };
-    builder.consume_reasoning_with_xml_tool_calls(form, "<seed:think>", "</seed:think>");
-}
-
-static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
-    builder.try_parse_reasoning("<think>", "</think>");
-    builder.add_content(builder.consume_rest());
-}
-
-static void common_chat_parse(common_chat_msg_parser & builder) {
-    LOG_DBG("Parsing input with format %s: %s\n", common_chat_format_name(builder.syntax().format), builder.input().c_str());
-
-    switch (builder.syntax().format) {
-        case COMMON_CHAT_FORMAT_CONTENT_ONLY:
-            common_chat_parse_content_only(builder);
-            break;
-        case COMMON_CHAT_FORMAT_GENERIC:
-            common_chat_parse_generic(builder);
-            break;
-        case COMMON_CHAT_FORMAT_MISTRAL_NEMO:
-            common_chat_parse_mistral_nemo(builder);
-            break;
-        case COMMON_CHAT_FORMAT_MAGISTRAL:
-            common_chat_parse_magistral(builder);
-            break;
-        case COMMON_CHAT_FORMAT_LLAMA_3_X:
-            common_chat_parse_llama_3_1(builder);
-            break;
-        case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS:
-            common_chat_parse_llama_3_1(builder, /* with_builtin_tools= */ true);
-            break;
-        case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
-            common_chat_parse_deepseek_r1(builder);
-            break;
-        case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1:
-            common_chat_parse_deepseek_v3_1(builder);
-            break;
-        case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
-            common_chat_parse_functionary_v3_2(builder);
-            break;
-        case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1:
-            common_chat_parse_functionary_v3_1_llama_3_1(builder);
-            break;
-        case COMMON_CHAT_FORMAT_HERMES_2_PRO:
-            common_chat_parse_hermes_2_pro(builder);
-            break;
-        case COMMON_CHAT_FORMAT_FIREFUNCTION_V2:
-            common_chat_parse_firefunction_v2(builder);
-            break;
-        case COMMON_CHAT_FORMAT_COMMAND_R7B:
-            common_chat_parse_command_r7b(builder);
-            break;
-        case COMMON_CHAT_FORMAT_GRANITE:
-            common_chat_parse_granite(builder);
-            break;
-        case COMMON_CHAT_FORMAT_GPT_OSS:
-            common_chat_parse_gpt_oss(builder);
-            break;
-        case COMMON_CHAT_FORMAT_SEED_OSS:
-            common_chat_parse_seed_oss(builder);
-            break;
-        case COMMON_CHAT_FORMAT_NEMOTRON_V2:
-            common_chat_parse_nemotron_v2(builder);
-            break;
-        case COMMON_CHAT_FORMAT_APERTUS:
-            common_chat_parse_apertus(builder);
-            break;
-        case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS:
-            common_chat_parse_lfm2(builder);
-            break;
-        case COMMON_CHAT_FORMAT_MINIMAX_M2:
-            common_chat_parse_minimax_m2(builder);
-            break;
-        case COMMON_CHAT_FORMAT_GLM_4_5:
-            common_chat_parse_glm_4_5(builder);
-            break;
-        case COMMON_CHAT_FORMAT_KIMI_K2:
-            common_chat_parse_kimi_k2(builder);
-            break;
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML:
-            common_chat_parse_qwen3_coder_xml(builder);
-            break;
-        case COMMON_CHAT_FORMAT_APRIEL_1_5:
-            common_chat_parse_apriel_1_5(builder);
-            break;
-        case COMMON_CHAT_FORMAT_XIAOMI_MIMO:
-            common_chat_parse_xiaomi_mimo(builder);
-            break;
-        default:
-            throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
-    }
-    builder.finish();
-}
-
-common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
-    common_chat_msg_parser builder(input, is_partial, syntax);
-    try {
-        common_chat_parse(builder);
-    } catch (const common_chat_msg_partial_exception & ex) {
-        LOG_DBG("Partial parse: %s\n", ex.what());
-        if (!is_partial) {
-            builder.clear_tools();
-            builder.move_to(0);
-            common_chat_parse_content_only(builder);
-        }
-    }
-    auto msg = builder.result();
-    if (!is_partial) {
-        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
-    }
-    return msg;
-}

common/chat-parser.h

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

common/chat.h

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

common/common.cpp

@@ -8,7 +8,6 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -27,6 +26,7 @@
 #include <sstream>
 #include <string>
 #include <thread>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -60,14 +60,6 @@
 #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
 //
@@ -950,58 +942,6 @@ 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);
@@ -1013,8 +953,6 @@ 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

@@ -2,15 +2,17 @@
 
 #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 '\\'
@@ -28,15 +30,6 @@
 
 #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;
@@ -140,22 +133,6 @@ 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
@@ -188,8 +165,6 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
-    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
-
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
 

common/json-partial.cpp

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

common/json-schema-to-grammar.cpp

@@ -268,10 +268,10 @@ static bool is_reserved_name(const std::string & name) {
 }
 
 std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
 std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
 };
 
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
@@ -303,8 +303,6 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
-std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
-
 class SchemaConverter {
 private:
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);

common/json-schema-to-grammar.h

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

common/sampling.cpp

@@ -3,10 +3,9 @@
 #include "common.h"
 #include "log.h"
 
-#include <algorithm>
 #include <cmath>
-#include <cstring>
 #include <unordered_map>
+#include <algorithm>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -113,13 +112,6 @@ 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);
 
@@ -136,12 +128,6 @@ 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 {
@@ -312,8 +298,6 @@ 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);
     }
@@ -324,7 +308,9 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
 }
 
 void common_sampler_reset(struct common_sampler * gsmpl) {
-    gsmpl->reset();
+    llama_sampler_reset(gsmpl->grmr);
+
+    llama_sampler_reset(gsmpl->chain);
 }
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
@@ -341,54 +327,16 @@ 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) {
-        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);
+        llama_perf_sampler_print(gsmpl->chain);
     }
-
     if (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_perf_context_print(ctx);
         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;
@@ -480,8 +428,6 @@ 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 new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
+                    or not new_name.endswith(".weight")
                 ):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
@@ -1673,9 +1673,11 @@ class GPTNeoXModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GPTNEOX
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_dimension_count(
             int(self.hparams["rotary_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])),
@@ -1733,7 +1735,7 @@ class BloomModel(TextModel):
         self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
         self.gguf_writer.add_embedding_length(n_embed)
         self.gguf_writer.add_feed_forward_length(4 * n_embed)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -1796,9 +1798,10 @@ class MPTModel(TextModel):
             self.gguf_writer.add_unk_token_id(0)
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["n_layers"]
         self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["d_model"])
         self.gguf_writer.add_head_count(self.hparams["n_heads"])
         if kv_n_heads := self.hparams["attn_config"].get("kv_n_heads"):
@@ -1831,6 +1834,7 @@ class OrionModel(TextModel):
         self._set_vocab_sentencepiece()
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -1848,7 +1852,7 @@ class OrionModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
@@ -1865,6 +1869,7 @@ class BaichuanModel(TextModel):
         self._set_vocab_sentencepiece()
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -1881,7 +1886,7 @@ class BaichuanModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(head_count)
@@ -1988,6 +1993,7 @@ class XverseModel(TextModel):
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -2004,7 +2010,7 @@ class XverseModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(head_count)
@@ -2047,6 +2053,10 @@ class FalconModel(TextModel):
     model_arch = gguf.MODEL_ARCH.FALCON
 
     def set_gguf_parameters(self):
+        block_count = self.hparams.get("num_hidden_layers")
+        if block_count is None:
+            block_count = self.hparams["n_layer"]  # old name
+
         n_head = self.hparams.get("num_attention_heads")
         if n_head is None:
             n_head = self.hparams["n_head"]  # old name
@@ -2059,7 +2069,7 @@ class FalconModel(TextModel):
         self.gguf_writer.add_tensor_data_layout("jploski")  # qkv tensor transform
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -2097,10 +2107,12 @@ class StarCoderModel(TextModel):
     model_arch = gguf.MODEL_ARCH.STARCODER
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["n_layer"]
+
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
         self.gguf_writer.add_head_count_kv(1)
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -2130,12 +2142,14 @@ class RefactModel(TextModel):
         multiple_of = 256
         ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
 
+        block_count = self.hparams["n_layer"]
+
         # refact uses Alibi. So this is from config.json which might be used by training.
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
 
         self.gguf_writer.add_feed_forward_length(ff_dim)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
         self.gguf_writer.add_head_count_kv(1)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
@@ -2182,10 +2196,11 @@ class StableLMModel(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
@@ -3136,7 +3151,7 @@ class DbrxModel(TextModel):
     def set_gguf_parameters(self):
         ffn_config = self.hparams["ffn_config"]
         attn_config = self.hparams["attn_config"]
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["n_layers"])
 
         self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
@@ -3338,7 +3353,7 @@ class QwenModel(TextModel):
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
@@ -4183,51 +4198,6 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
-@ModelBase.register("Qwen3NextForCausalLM")
-class Qwen3NextModel(Qwen2MoeModel):
-    model_arch = gguf.MODEL_ARCH.QWEN3NEXT
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_ssm_conv_kernel(self.hparams["linear_conv_kernel_dim"])
-        self.gguf_writer.add_ssm_state_size(self.hparams["linear_key_head_dim"])
-        self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
-        self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
-        self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
-        if (rope_dim := self.hparams.get("head_dim")) is None:
-            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
-        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if name.startswith("mtp"):
-            return [] # ignore MTP layers for now
-        if name.endswith(".A_log"):
-            data_torch = -torch.exp(data_torch)
-        elif name.endswith(".dt_bias"):
-            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
-        elif "conv1d" in name:
-            data_torch = data_torch.squeeze()
-        elif name.endswith("norm.weight") and not name.endswith("linear_attn.norm.weight"):
-            data_torch = data_torch + 1
-
-        yield from super().modify_tensors(data_torch, name, bid)
-
-
-@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):
@@ -4414,7 +4384,7 @@ class GPT2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GPT2
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
         self.gguf_writer.add_context_length(self.hparams["n_ctx"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
@@ -4446,6 +4416,8 @@ class Phi2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI2
 
     def set_gguf_parameters(self):
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
+
         rot_pct = self.find_hparam(["partial_rotary_factor"])
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
@@ -4454,7 +4426,7 @@ class Phi2Model(TextModel):
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
@@ -4572,6 +4544,8 @@ class Phi3MiniModel(TextModel):
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
+
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
@@ -4585,7 +4559,7 @@ class Phi3MiniModel(TextModel):
         self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
@@ -4705,11 +4679,12 @@ class PlamoModel(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(4096)  # not in config.json
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(5)  # hparams["num_key_value_heads"]) is wrong
         self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
@@ -4832,6 +4807,7 @@ class Plamo2Model(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
         self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
 
         # Which layers are Mamba layers
@@ -4843,10 +4819,10 @@ class Plamo2Model(TextModel):
         num_attention_heads = []
 
         if mamba_enabled:
-            for i in range(self.block_count):
-                if self.block_count <= (mamba_step // 2):
+            for i in range(block_count):
+                if block_count <= (mamba_step // 2):
                     # use attention in last layer
-                    is_mamba = (i != self.block_count - 1)
+                    is_mamba = (i != block_count - 1)
                 else:
                     is_mamba = (i % mamba_step) != (mamba_step // 2)
                 if is_mamba:
@@ -4864,7 +4840,7 @@ class Plamo2Model(TextModel):
         self.gguf_writer.add_embedding_length(hparams.get("hidden_size", 4096))
         self.gguf_writer.add_key_length(hparams.get("hidden_size_per_head", 128))
         self.gguf_writer.add_value_length(hparams.get("hidden_size_per_head", 128))
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 10000))
 
@@ -4921,10 +4897,12 @@ class CodeShellModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CODESHELL
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["n_layer"]
+
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_query_groups"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
@@ -5066,7 +5044,7 @@ class InternLM2Model(TextModel):
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
@@ -5687,10 +5665,11 @@ class GemmaModel(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
@@ -5726,10 +5705,11 @@ class Gemma2Model(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
 
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
@@ -5773,11 +5753,12 @@ class Gemma3Model(TextModel):
 
     def set_gguf_parameters(self):
         hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
 
         # some default values are not specified in the hparams
         self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 131072))
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
         self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 8))
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6))
@@ -6053,6 +6034,7 @@ class Rwkv6Model(TextModel):
         self._set_vocab_rwkv_world()
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         head_size = self.hparams["head_size"]
         hidden_size = self.hparams["hidden_size"]
         layer_norm_eps = self.hparams["layer_norm_epsilon"]
@@ -6064,7 +6046,7 @@ class Rwkv6Model(TextModel):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
         self.gguf_writer.add_rescale_every_n_layers(rescale_every_n_layers)
         self.gguf_writer.add_wkv_head_size(head_size)
@@ -6128,6 +6110,7 @@ class RWKV6Qwen2Model(Rwkv6Model):
             self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         num_attention_heads = self.hparams["num_attention_heads"]
         num_key_value_heads = self.hparams["num_key_value_heads"]
         hidden_size = self.hparams["hidden_size"]
@@ -6140,7 +6123,7 @@ class RWKV6Qwen2Model(Rwkv6Model):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_wkv_head_size(head_size)
         self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
         self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
@@ -6181,6 +6164,7 @@ class Rwkv7Model(TextModel):
         return max(1, round(hidden_size ** exponent * multiplier / 32)) * 32
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         try:
             head_size = self.hparams["head_size"]
             layer_norm_eps = self.hparams["layer_norm_epsilon"]
@@ -6205,7 +6189,7 @@ class Rwkv7Model(TextModel):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
         self.gguf_writer.add_wkv_head_size(head_size)
         self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
@@ -6299,6 +6283,7 @@ class ARwkv7Model(Rwkv7Model):
             self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
         hidden_size = self.hparams["hidden_size"]
         head_size = self.hparams["head_size"]
         rms_norm_eps = self.hparams["rms_norm_eps"]
@@ -6315,7 +6300,7 @@ class ARwkv7Model(Rwkv7Model):
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
         self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
         self.gguf_writer.add_wkv_head_size(head_size)
         self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
@@ -7539,7 +7524,7 @@ class T5Model(TextModel):
         self.gguf_writer.add_context_length(n_ctx)
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
         if (dec_n_layer := self.hparams.get("num_decoder_layers")) is not None:
             self.gguf_writer.add_decoder_block_count(dec_n_layer)
         self.gguf_writer.add_head_count(self.hparams["num_heads"])
@@ -7678,7 +7663,7 @@ class T5EncoderModel(TextModel):
         self.gguf_writer.add_context_length(n_ctx)
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
         self.gguf_writer.add_head_count(self.hparams["num_heads"])
         self.gguf_writer.add_key_length(self.hparams["d_kv"])
         self.gguf_writer.add_value_length(self.hparams["d_kv"])
@@ -7741,7 +7726,7 @@ class JaisModel(TextModel):
         self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"])
@@ -8083,7 +8068,7 @@ class ChatGLMModel(TextModel):
         self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
         self.gguf_writer.add_embedding_length(n_embed)
         self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", self.hparams.get("intermediate_size", 4 * n_embed)))
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(self.hparams.get("num_layers", self.hparams["num_hidden_layers"]))
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon",1e-5))
@@ -8165,6 +8150,7 @@ class ExaoneModel(TextModel):
         num_kv_heads = hparams.get("num_key_value_heads", num_heads)
         layer_norm_eps = hparams["layer_norm_epsilon"]
         intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim
+        num_layers = hparams["num_layers"]
         # ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0
         # attention_dropout_rate = hparams["attention_dropout"]
         # ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0
@@ -8175,7 +8161,7 @@ class ExaoneModel(TextModel):
         self.gguf_writer.add_context_length(max_position_embeddings)
         self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps)
         self.gguf_writer.add_feed_forward_length(intermediate_size)
-        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_block_count(num_layers)
         self.gguf_writer.add_file_type(self.ftype)
 
         if (rope_theta := self.hparams.get("rope_theta")) is not None:
@@ -10091,25 +10077,6 @@ 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
@@ -10153,14 +10120,8 @@ 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]
-            numpy_dtype = cls._dtype_byteswap_map[dtype]
-            return torch.from_numpy(byteswap_tensor(tensor.mmap_bytes(), numpy_dtype)).view(dtype).reshape(tensor.shape)
+            return torch.from_numpy(tensor.mmap_bytes()).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))
@@ -10168,16 +10129,10 @@ 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.from_numpy(byteswap_tensor(np.frombuffer(r.data(), dtype=numpy_dtype), numpy_dtype)).view(dtype).reshape(shape))
+        lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.frombuffer(r.data(), dtype=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="f32",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
         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(
@@ -277,15 +277,10 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
-def load_hparams_from_hf(hf_model_id: str) -> tuple[dict[str, Any], Path | None]:
-    from huggingface_hub import try_to_load_from_cache
-
+def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
     # normally, adapter does not come with base model config, we need to load it from AutoConfig
     config = AutoConfig.from_pretrained(hf_model_id)
-    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
+    return config.to_dict()
 
 
 if __name__ == '__main__':
@@ -330,13 +325,13 @@ 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, dir_base_model = load_hparams_from_hf(base_model_id)
+        hparams = 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, dir_base_model = load_hparams_from_hf(model_id)
+                hparams = 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")
@@ -485,7 +480,6 @@ if __name__ == '__main__':
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
             hparams=hparams,
-            remote_hf_model_id=base_model_id,
         )
 
         logger.info("Exporting model...")

docs/backend/SYCL.md

@@ -42,9 +42,6 @@ The following releases are verified and recommended:
 
 ## News
 
-- 2025.11
-  - Support malloc memory on device more than 4GB.
-
 - 2025.2
   - Optimize MUL_MAT Q4_0 on Intel GPU for all dGPUs and built-in GPUs since MTL. Increase the performance of LLM (llama-2-7b.Q4_0.gguf) 21%-87% on Intel GPUs (MTL, ARL-H, Arc, Flex, PVC).
     |GPU|Base tokens/s|Increased tokens/s|Percent|
@@ -792,8 +789,6 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
-| UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Support malloc device memory more than 4GB.|
-
 
 
 ## Known Issues
@@ -840,14 +835,6 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   | The default context is too big. It leads to excessive memory usage.|Set `-c 8192` or a smaller value.|
   | The model is too big and requires more memory than what is available.|Choose a smaller model or change to a smaller quantization, like Q5 -> Q4;<br>Alternatively, use more than one device to load model.|
 
-- `ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 5000000000 Bytes of memory on device`
-
-  You need to enable to support 4GB memory malloc by:
-  ```
-    export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-    set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-  ```
-
 ### **GitHub contribution**:
 Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
 

docs/ops.md

@@ -17,12 +17,12 @@ Legend:
 |                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
@@ -43,9 +43,9 @@ Legend:
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                             FILL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -87,7 +87,7 @@ Legend:
 |                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -99,7 +99,7 @@ Legend:
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
 |                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -107,7 +107,7 @@ Legend:
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
 |                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ |
@@ -116,6 +116,6 @@ Legend:
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
 |                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

docs/ops/Vulkan.csv

@@ -5,8 +5,8 @@
 "Vulkan0","SGN","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","NEG","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
 "Vulkan0","NEG","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","STEP","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","STEP","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
+"Vulkan0","STEP","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","STEP","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","TANH","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
 "Vulkan0","TANH","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
 "Vulkan0","ELU","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
@@ -29,18 +29,18 @@
 "Vulkan0","EXP","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
 "Vulkan0","EXPM1","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
 "Vulkan0","EXPM1","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
-"Vulkan0","SOFTPLUS","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","SOFTPLUS","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
+"Vulkan0","SOFTPLUS","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","SOFTPLUS","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
 "Vulkan0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
+"Vulkan0","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","ABS","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","Vulkan"
 "Vulkan0","ABS","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","Vulkan"
 "Vulkan0","SGN","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","Vulkan"
@@ -89,8 +89,8 @@
 "Vulkan0","SGN","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","NEG","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
 "Vulkan0","NEG","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","STEP","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","STEP","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
+"Vulkan0","STEP","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","STEP","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","TANH","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
 "Vulkan0","TANH","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
 "Vulkan0","ELU","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
@@ -113,18 +113,18 @@
 "Vulkan0","EXP","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
 "Vulkan0","EXPM1","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
 "Vulkan0","EXPM1","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
-"Vulkan0","SOFTPLUS","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","SOFTPLUS","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
+"Vulkan0","SOFTPLUS","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","SOFTPLUS","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
 "Vulkan0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","Vulkan"
+"Vulkan0","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","Vulkan"
 "Vulkan0","ABS","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","Vulkan"
 "Vulkan0","ABS","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","Vulkan"
 "Vulkan0","SGN","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","Vulkan"
@@ -5654,7 +5654,7 @@
 "Vulkan0","SUB","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","Vulkan"
 "Vulkan0","MUL","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","Vulkan"
 "Vulkan0","DIV","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","Vulkan"
-"Vulkan0","ADD1","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","ADD1","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
 "Vulkan0","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=0.000000,inplace=0","support","1","yes","Vulkan"
 "Vulkan0","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000,inplace=0","support","1","yes","Vulkan"
 "Vulkan0","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000,inplace=1","support","1","yes","Vulkan"
@@ -8632,10 +8632,10 @@
 "Vulkan0","COS","type=f16,ne=[10,2,2,2]","support","0","no","Vulkan"
 "Vulkan0","CLAMP","type=f16,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","0","no","Vulkan"
 "Vulkan0","LEAKY_RELU","type=f16,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","Vulkan"
-"Vulkan0","FLOOR","type=f16,ne=[10,2,2,2]","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f16,ne=[10,2,2,2]","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f16,ne=[10,2,2,2]","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f16,ne=[10,2,2,2]","support","1","yes","Vulkan"
+"Vulkan0","FLOOR","type=f16,ne=[10,2,2,2]","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f16,ne=[10,2,2,2]","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f16,ne=[10,2,2,2]","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f16,ne=[10,2,2,2]","support","0","no","Vulkan"
 "Vulkan0","SQR","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
 "Vulkan0","SQRT","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
 "Vulkan0","LOG","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
@@ -8643,10 +8643,10 @@
 "Vulkan0","COS","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
 "Vulkan0","CLAMP","type=f16,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","0","no","Vulkan"
 "Vulkan0","LEAKY_RELU","type=f16,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","Vulkan"
-"Vulkan0","FLOOR","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","FLOOR","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f16,ne=[7,1,5,3]","support","0","no","Vulkan"
 "Vulkan0","SQR","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
 "Vulkan0","SQRT","type=f32,ne=[10,3,3,2]","support","1","yes","Vulkan"
 "Vulkan0","LOG","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
@@ -8654,10 +8654,10 @@
 "Vulkan0","COS","type=f32,ne=[10,2,2,2]","support","1","yes","Vulkan"
 "Vulkan0","CLAMP","type=f32,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","Vulkan"
 "Vulkan0","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","1","yes","Vulkan"
-"Vulkan0","FLOOR","type=f32,ne=[10,2,2,2]","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f32,ne=[10,2,2,2]","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f32,ne=[10,2,2,2]","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f32,ne=[10,2,2,2]","support","1","yes","Vulkan"
+"Vulkan0","FLOOR","type=f32,ne=[10,2,2,2]","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f32,ne=[10,2,2,2]","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f32,ne=[10,2,2,2]","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f32,ne=[10,2,2,2]","support","0","no","Vulkan"
 "Vulkan0","SQR","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
 "Vulkan0","SQRT","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
 "Vulkan0","LOG","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
@@ -8665,10 +8665,10 @@
 "Vulkan0","COS","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
 "Vulkan0","CLAMP","type=f32,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","Vulkan"
 "Vulkan0","LEAKY_RELU","type=f32,ne_a=[7,1,5,3],negative_slope=0.100000","support","1","yes","Vulkan"
-"Vulkan0","FLOOR","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
-"Vulkan0","CEIL","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
-"Vulkan0","ROUND","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
-"Vulkan0","TRUNC","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","FLOOR","type=f32,ne=[7,1,5,3]","support","0","no","Vulkan"
+"Vulkan0","CEIL","type=f32,ne=[7,1,5,3]","support","0","no","Vulkan"
+"Vulkan0","ROUND","type=f32,ne=[7,1,5,3]","support","0","no","Vulkan"
+"Vulkan0","TRUNC","type=f32,ne=[7,1,5,3]","support","0","no","Vulkan"
 "Vulkan0","DIAG_MASK_INF","type=f32,ne=[10,10,1,1],n_past=5","support","1","yes","Vulkan"
 "Vulkan0","DIAG_MASK_INF","type=f32,ne=[10,10,3,1],n_past=5","support","1","yes","Vulkan"
 "Vulkan0","DIAG_MASK_INF","type=f32,ne=[10,10,3,2],n_past=5","support","1","yes","Vulkan"
@@ -9478,7 +9478,7 @@
 "Vulkan0","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","no","Vulkan"
 "Vulkan0","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","0","no","Vulkan"
 "Vulkan0","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","1","yes","Vulkan"
-"Vulkan0","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","1","yes","Vulkan"
+"Vulkan0","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","0","no","Vulkan"
 "Vulkan0","TIMESTEP_EMBEDDING","type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000","support","1","yes","Vulkan"
 "Vulkan0","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","1","yes","Vulkan"
 "Vulkan0","CUMSUM","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
@@ -9487,9 +9487,9 @@
 "Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","0","no","Vulkan"
 "Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","0","no","Vulkan"
 "Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=0","support","0","no","Vulkan"
-"Vulkan0","FILL","type=f32,ne=[10,10,4,3],c=0.000000","support","1","yes","Vulkan"
-"Vulkan0","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","1","yes","Vulkan"
-"Vulkan0","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","1","yes","Vulkan"
+"Vulkan0","FILL","type=f32,ne=[10,10,4,3],c=0.000000","support","0","no","Vulkan"
+"Vulkan0","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","0","no","Vulkan"
+"Vulkan0","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","0","no","Vulkan"
 "Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","0","no","Vulkan"
 "Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","0","no","Vulkan"
 "Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","0","no","Vulkan"

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 --kv-unified
+./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
 
 ...
 

examples/diffusion/README.md

@@ -6,54 +6,8 @@ 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:
 
-### 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 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`
 
-### Scheduling Parameters
-Choose one of the following scheduling methods:
+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`
 
-**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/embedding/embedding.cpp

@@ -104,16 +104,12 @@ int main(int argc, char ** argv) {
 
     params.embedding = true;
 
-    // get max number of sequences per batch
-    const int n_seq_max = llama_max_parallel_sequences();
-
     // if the number of prompts that would be encoded is known in advance, it's more efficient to specify the
     //   --parallel argument accordingly. for convenience, if not specified, we fallback to unified KV cache
     //   in order to support any number of prompts
     if (params.n_parallel == 1) {
         LOG_INF("%s: n_parallel == 1 -> unified KV cache is enabled\n", __func__);
         params.kv_unified = true;
-        params.n_parallel = n_seq_max;
     }
 
     // utilize the full context
@@ -127,6 +123,9 @@ int main(int argc, char ** argv) {
         params.n_ubatch = params.n_batch;
     }
 
+    // get max number of sequences per batch
+    const int n_seq_max = llama_max_parallel_sequences();
+
     llama_backend_init();
     llama_numa_init(params.numa);
 

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(const 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(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(*(const ggml_fp16_t *) &data[i]);
+        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
     } else if (type == GGML_TYPE_F32) {
-        v = *(const float *) &data[i];
+        v = *(float *) &data[i];
     } else if (type == GGML_TYPE_I64) {
-        v = (float) *(const int64_t *) &data[i];
+        v = (float) *(int64_t *) &data[i];
     } else if (type == GGML_TYPE_I32) {
-        v = (float) *(const int32_t *) &data[i];
+        v = (float) *(int32_t *) &data[i];
     } else if (type == GGML_TYPE_I16) {
-        v = (float) *(const int16_t *) &data[i];
+        v = (float) *(int16_t *) &data[i];
     } else if (type == GGML_TYPE_I8) {
-        v = (float) *(const int8_t *) &data[i];
+        v = (float) *(int8_t *) &data[i];
     } else if (type == GGML_TYPE_BF16) {
-        v = ggml_compute_bf16_to_fp32(*(const ggml_bf16_t *) &data[i]);
+        v = ggml_compute_bf16_to_fp32(*(ggml_bf16_t *) &data[i]);
     } else {
         GGML_ABORT("fatal error");
     }

examples/json_schema_to_grammar.py

@@ -231,9 +231,9 @@ DOT = '[^\\x0A\\x0D]'
 RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
 
 INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
-GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\\]')
+GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
 GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]', '\\': '\\\\'}
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
 
 NON_LITERAL_SET = set('|.()[]{}*+?')
 ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('^$.[]()|{}*+?')

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -4,11 +4,6 @@ set -e
 
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
-MODEL_TESTING_PROMPT="${2:-"$MODEL_TESTING_PROMPT"}"
-
-if [ -z "$MODEL_TESTING_PROMPT"]; then
-    MODEL_TESTING_PROMPT="Hello, my name is"
-fi
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -19,8 +14,7 @@ if [ -z "$CONVERTED_MODEL" ]; then
 fi
 
 echo $CONVERTED_MODEL
-echo $MODEL_TESTING_PROMPT
 
 cmake --build ../../build --target llama-logits -j8
 
-../../build/bin/llama-logits -m "$CONVERTED_MODEL" "$MODEL_TESTING_PROMPT"
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"

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

@@ -184,12 +184,8 @@ model_name = os.path.basename(model_path)
 # of using AutoModelForCausalLM.
 print(f"Model class: {model.__class__.__name__}")
 
-device = next(model.parameters()).device
-if os.getenv("MODEL_TESTING_PROMPT"):
-    prompt = os.getenv("MODEL_TESTING_PROMPT")
-else:
-    prompt = "Hello, my name is"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+prompt = "Hello, my name is"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
 
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")

examples/sycl/run-llama2.sh

@@ -15,9 +15,6 @@ MODEL_FILE=models/llama-2-7b.Q4_0.gguf
 NGL=99
 CONTEXT=4096
 
-#support malloc device memory more than 4GB.
-export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"

examples/sycl/run-llama3.sh

@@ -6,7 +6,7 @@
 
 # If you want more control, DPC++ Allows selecting a specific device through the
 # following environment variable
-export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+#export ONEAPI_DEVICE_SELECTOR="level_zero:0"
 source /opt/intel/oneapi/setvars.sh
 
 #export GGML_SYCL_DEBUG=1
@@ -18,14 +18,11 @@ MODEL_FILE=models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
 NGL=99 # Layers offloaded to the GPU. If the device runs out of memory, reduce this value according to the model you are using.
 CONTEXT=4096
 
-#support malloc device memory more than 4GB.
-export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "Using $GGML_SYCL_DEVICE as the main GPU"
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT}
 fi

examples/sycl/win-run-llama2.bat

@@ -5,7 +5,5 @@
 set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-:: support malloc device memory more than 4GB.
-set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
 .\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 99 -s 0

examples/sycl/win-run-llama3.bat

@@ -5,7 +5,5 @@
 set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-:: support malloc device memory more than 4GB.
-set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
-.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -s 0 -e -ngl 99
+.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -e -ngl 99

ggml/CMakeLists.txt

@@ -25,17 +25,16 @@ 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)
@@ -183,7 +182,6 @@ endif()
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
 option(GGML_CPU                             "ggml: enable CPU backend"                        ON)
-option(GGML_SCHED_NO_REALLOC                "ggml: disallow reallocations in ggml-alloc (for debugging)" OFF)
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)

ggml/include/ggml-rpc.h

@@ -8,7 +8,7 @@ extern "C" {
 #endif
 
 #define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    5
+#define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 

ggml/include/ggml.h

@@ -530,7 +530,6 @@ 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,
@@ -2259,25 +2258,18 @@ 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

@@ -221,10 +221,6 @@ if (GGML_BACKEND_DL)
     target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL)
 endif()
 
-if (GGML_SCHED_NO_REALLOC)
-    target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
-endif()
-
 add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
@@ -332,14 +328,6 @@ 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()
@@ -414,13 +402,6 @@ 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-alloc.c

@@ -921,15 +921,10 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         }
         if (realloc) {
 #ifndef NDEBUG
-            {
-                size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
-                if (cur_size > 0) {
-                    GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n",
-                        __func__, ggml_backend_buft_name(galloc->bufts[i]),
-                        cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
-                }
-            }
+            size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
+            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+
             ggml_vbuffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
             if (galloc->buffers[i] == NULL) {

ggml/src/ggml-backend.cpp

@@ -1395,20 +1395,14 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
-#ifdef GGML_SCHED_NO_REALLOC
-        GGML_ABORT("%s: failed to allocate graph, but graph re-allocation is disabled by GGML_SCHED_NO_REALLOC\n", __func__);
-#endif
-
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
-#endif
-
         // the re-allocation may cause the split inputs to be moved to a different address
         // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
         for (int i = 0; i < sched->n_backends; i++) {
             ggml_backend_synchronize(sched->backends[i]);
         }
-
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
         ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
         if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
             GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__);

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -42,7 +42,6 @@
 #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>
@@ -2207,120 +2206,78 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context & ctx,
 }
 
 /**
- * @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.
+ * @brief Initializes and caches sine/cosine positional encoding values
+ *        (used in RoPE, Rotary Position Embedding) for attention layers.
  *
- * 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.
+ * 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.
  *
- * 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
+ * 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.
  *
- * @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.
+ * @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).
  */
-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) {
+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) {
     ggml_tensor * src0 = dst->src[0];  // input
     ggml_tensor * src1 = dst->src[1];  // position
     ggml_tensor * src2 = dst->src[2];  // freq_factors
 
-    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)) {
+    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) {
         // use cache.
         return;
     }
 
-    // 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) };
+    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) };
 
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    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_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 cache_ne[] = { theta_scale_length, 1, position_length, 1 };
-    size_t  cache_nb[GGML_MAX_DIMS];
-    cache_nb[0] = sizeof(float);
+    int64_t theta_ne[] = { theta_scale_length, 1, position_length, 1 };
+    size_t  theta_nb[GGML_MAX_DIMS];
+    theta_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        cache_nb[i] = cache_nb[i - 1] * cache_ne[i - 1];
+        theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 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.
+    // theta_scale arange, [0,1,...,ne00/2 - 1]
     acl_tensor_ptr acl_theta_scale_tensor;
-    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;
-            }
-        }
+    // 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;
 
         if (ctx.rope_cache.theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(ctx.rope_cache.theta_scale_cache));
@@ -2328,138 +2285,74 @@ static void aclnn_rope_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);
-    }
 
-    // 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;
+        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);
 
-        // -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_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);
 
-        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());
+            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());
 
-        // 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());
-    }
-
-    // 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());
+            // 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());
         }
-    } else {
-        if (freq_scale != 1 && (ctx.rope_cache.freq_scale != freq_scale || theta_scale_updated)) {
-            theta_scale_updated = true;
+
+        // 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) {
+            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);
         }
-    }
-
-    // Nothing changed, use cache.
-    if (!theta_scale_updated) {
+    } else {
+        // use cache
         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();
@@ -2472,85 +2365,6 @@ static void aclnn_rope_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;
@@ -2567,30 +2381,44 @@ static void aclnn_rope_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), cache_ne, cache_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_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), cache_ne, cache_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_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);
     }
 
-    // Step 5: multiply by attn_factor
+    // 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, dst->ne[2], 1 };
+    int64_t sin_reshape_ne[4] = { src0->ne[0], 1, src0->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++) {
@@ -2601,9 +2429,8 @@ static void aclnn_rope_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);
 
-    // Step 6: repeat
+    // 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);
@@ -2611,15 +2438,17 @@ static void aclnn_rope_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);
     }
 
-    // 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);
+    // 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;
 }
 
 #ifdef __cplusplus
@@ -2645,7 +2474,6 @@ 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];
@@ -2654,13 +2482,12 @@ 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(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
+    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));
 
     // TODO: n_dims <= ne0
     GGML_ASSERT(n_dims == ne0);
@@ -2671,25 +2498,10 @@ 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);
 
-    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;
-    }
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     // init ctx.rope_cos/rope_sin cache
-    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);
+    aclnn_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox);
 
     int64_t sin_reshape_ne[4] = { ne00, 1, ne02, 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
@@ -2845,7 +2657,8 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     return;
 #endif
 
-    int64_t acl_mode = is_neox ? 0 : 1;
+    // ggml_mode = 0 --> aclnn_model = 1
+    int64_t acl_mode = mode == 0 ? 1 : mode;
 
     switch (src0->type) {
         case GGML_TYPE_F32:
@@ -3423,64 +3236,3 @@ 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

@@ -1125,23 +1125,3 @@ 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

@@ -300,92 +300,30 @@ struct ggml_cann_graph_lru_cache {
 
 struct ggml_cann_rope_cache {
     ~ggml_cann_rope_cache() {
-        if (theta_scale_cache) {
+        if (theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
-        if (sin_cache) {
+        if (sin_cache != nullptr) {
             ACL_CHECK(aclrtFree(sin_cache));
         }
-        if (cos_cache) {
+        if (cos_cache != nullptr) {
             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);
-        }
     }
 
-    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;
+    void *  theta_scale_cache  = nullptr;
+    int64_t theta_scale_length = 0;
     // sin/cos cache, used only to accelerate first layer on each device
-    void *  sin_cache                  = nullptr;
-    void *  cos_cache                  = nullptr;
+    void *  sin_cache          = nullptr;
+    void *  cos_cache          = nullptr;
+    int64_t position_length    = 0;
     // Properties to check before reusing the sincos cache
-    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;
+    bool    cached             = false;
+    float   ext_factor         = 0.0f;
+    float   theta_scale        = 0.0f;
+    float   freq_scale         = 0.0f;
+    float   attn_factor        = 0.0f;
+    bool    is_neox            = false;
 };
 
 struct ggml_cann_tensor_cache {

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

@@ -1886,9 +1886,6 @@ 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;
     }
@@ -2249,7 +2246,8 @@ 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
-    if (use_cann_graph && cann_graph_update_required) {  // Begin CANN graph capture
+    ggml_cann_graph * matched_graph = cann_ctx->graph_lru_cache.cache_list.front();
+    if (use_cann_graph && cann_graph_update_required) {
         ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
     }
 #endif  // USE_ACL_GRAPH
@@ -2273,14 +2271,12 @@ 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) {
-        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
+        // Execute graph
         ACL_CHECK(aclmdlRIExecuteAsync(matched_graph->graph, cann_ctx->stream()));
     }
 #endif  // USE_ACL_GRAPH
@@ -2306,9 +2302,9 @@ 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) {
@@ -2339,6 +2335,7 @@ 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);
 
@@ -2480,6 +2477,13 @@ 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;
                 }
@@ -2559,16 +2563,6 @@ 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

@@ -224,8 +224,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
             include(CheckCXXSourceCompiles)
             set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-            string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
-            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
+            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS}")
             foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
                 set(ARM_FEATURE "HAVE_${feature}")
                 check_cxx_source_compiles(
@@ -393,9 +392,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
-                list(APPEND ARCH_FLAGS -mcpu=power10)
+                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
             elseif (EXTRACTED_NUMBER EQUAL 9)
-                list(APPEND ARCH_FLAGS -mcpu=power9)
+                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
             elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
                 list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
             else()
@@ -453,35 +452,22 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ggml-cpu/spacemit/ime_kernels.h
             )
         endif()
-        if(NOT GGML_CPU_ALL_VARIANTS)
-            set(MARCH_STR "rv64gc")
-            if (GGML_RV_ZFH)
-                string(APPEND MARCH_STR "_zfh")
-            endif()
-            if (GGML_XTHEADVECTOR)
-                string(APPEND MARCH_STR "_xtheadvector")
-            elseif (GGML_RVV)
-                string(APPEND MARCH_STR "_v")
-                if (GGML_RV_ZVFH)
-                    string(APPEND MARCH_STR "_zvfh")
-                endif()
-            endif()
-            if (GGML_RV_ZICBOP)
-                string(APPEND MARCH_STR "_zicbop")
-            endif()
-            list(APPEND ARCH_FLAGS "-march=${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 ${ARCH_DEFINITIONS})
+        set(MARCH_STR "rv64gc")
+        if (GGML_RV_ZFH)
+            string(APPEND MARCH_STR "_zfh")
         endif()
+        if (GGML_XTHEADVECTOR)
+            string(APPEND MARCH_STR "_xtheadvector")
+        elseif (GGML_RVV)
+            string(APPEND MARCH_STR "_v")
+            if (GGML_RV_ZVFH)
+                string(APPEND MARCH_STR "_zvfh")
+            endif()
+        endif()
+        if (GGML_RV_ZICBOP)
+            string(APPEND MARCH_STR "_zicbop")
+        endif()
+        list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
         list(APPEND GGML_CPU_SOURCES

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

@@ -33,12 +33,10 @@
 // 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
@@ -46,30 +44,27 @@
 #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
@@ -81,12 +76,10 @@
 // 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
@@ -94,7 +87,6 @@
 #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
@@ -109,12 +101,10 @@
 // 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
@@ -122,7 +112,6 @@
 #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
@@ -145,18 +134,15 @@
 // 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
@@ -177,12 +163,10 @@
 // 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
@@ -190,7 +174,6 @@
 #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
@@ -213,12 +196,10 @@
 // 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
@@ -226,7 +207,6 @@
 #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

@@ -24,29 +24,6 @@
 
 #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);
@@ -497,295 +474,6 @@ 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;
@@ -2201,412 +1889,3 @@ 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

@@ -1,38 +0,0 @@
-#include "ggml-backend-impl.h"
-
-#if defined(__riscv) && __riscv_xlen == 64
-#include <asm/hwprobe.h>
-#include <asm/unistd.h>
-#include <unistd.h>
-
-struct riscv64_features {
-    bool has_rvv = false;
-
-    riscv64_features() {
-        struct riscv_hwprobe probe;
-        probe.key = RISCV_HWPROBE_KEY_IMA_EXT_0;
-        probe.value = 0;
-
-        int ret = syscall(__NR_riscv_hwprobe, &probe, 1, 0, NULL, 0);
-
-        if (0 == ret) {
-            has_rvv = !!(probe.value & RISCV_HWPROBE_IMA_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

@@ -1927,10 +1927,6 @@ 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);
@@ -2315,7 +2311,6 @@ 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:
@@ -2839,10 +2834,6 @@ 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/kleidiai/kernels.cpp

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

ggml/src/ggml-cpu/ops.cpp

@@ -7794,7 +7794,7 @@ void ggml_compute_forward_timestep_embedding(
 // ggml_compute_forward_argsort
 
 template<enum ggml_sort_order order>
-struct cmp_argsort {
+struct argsort_cmp {
     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, cmp_argsort<GGML_SORT_ORDER_ASC>{src_data});
+                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_ASC>{src_data});
                 break;
 
             case GGML_SORT_ORDER_DESC:
-                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_DESC>{src_data});
+                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_DESC>{src_data});
                 break;
 
             default:
@@ -7864,72 +7864,6 @@ 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(
@@ -9762,13 +9696,13 @@ static void ggml_compute_forward_solve_tri_f32(const struct ggml_compute_params
         for (int64_t i00 = 0; i00 < n; ++i00) {
             float sum = 0.0f;
             for (int64_t t = 0; t < i00; ++t) {
-                sum += A_batch[i00 * n + t] * X_batch[t * k + i01];
+                sum += A_batch[i00 * n + t] * X_batch[i01 * n + t];
             }
 
             const float diag = A_batch[i00 * n + i00];
-            assert(diag != 0.0f && "Zero diagonal in triangular matrix");
+            GGML_ASSERT(diag != 0.0f && "Zero diagonal in triangular matrix");
 
-            X_batch[i00 * k + i01] = (B_batch[i00 * k + i01] - sum) / diag;
+            X_batch[i01 * n + i00] = (B_batch[i00 * k + i01] - sum) / diag;
         }
     }
 }

ggml/src/ggml-cpu/ops.h

@@ -81,7 +81,6 @@ 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

@@ -124,58 +124,6 @@ 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);
@@ -244,12 +192,6 @@ 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);
@@ -391,77 +333,6 @@ 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;
@@ -856,89 +727,6 @@ 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;
@@ -1440,10 +1228,9 @@ 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 || interleave_block == 4);
+    GGML_ASSERT(interleave_block == 8);
     constexpr int nrows_interleaved = 8;
 
     block_q4_Kx8 * dst = (block_q4_Kx8*)t->data;
@@ -1681,10 +1468,6 @@ 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);
 }
@@ -1718,10 +1501,6 @@ 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);
 }
@@ -1750,10 +1529,6 @@ 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);
 }
@@ -1956,13 +1731,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             nchunk0 = (nr0 + min_chunk_size - 1) / min_chunk_size;
         }
 
-        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)) {
+        if (nth == 1 || nchunk0 < nth || disable_chunking) {
             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;
@@ -2156,9 +1930,6 @@ 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
@@ -2190,16 +1961,6 @@ 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

@@ -80,12 +80,10 @@ 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);
@@ -93,7 +91,6 @@ 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);
@@ -102,12 +99,10 @@ 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);
@@ -115,7 +110,6 @@ 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/simd-mappings.h

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

ggml/src/ggml-cpu/vec.h

@@ -397,118 +397,119 @@ 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) && 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)
+    #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;
 
-    GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+        GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
 
-    int np = (n & ~(ggml_f16_step - 1));
+        const 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);
-    }
-    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);
+            GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
         }
-    }
-#else
-    const int np = 0;
-#endif
+        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);
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
+            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) {
         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
@@ -697,61 +698,60 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 }
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
-#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
-    const int sve_register_length = svcntb() * 8;
-    const int ggml_f16_epr = sve_register_length / 16;
-    const int ggml_f16_step = 2 * ggml_f16_epr;
+#if defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = svcntb() * 8;
+        const int ggml_f16_epr = sve_register_length / 16;
+        const int ggml_f16_step = 2 * ggml_f16_epr;
 
-    GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
-    const int np = (n & ~(ggml_f16_step - 1));
-    svfloat16_t ay1, ay2;
+        GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
+        const int np = (n & ~(ggml_f16_step - 1));
+        svfloat16_t ay1, ay2;
 
-    for (int i = 0; i < np; i += ggml_f16_step) {
-        ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
-        ay1 = GGML_F16x_VEC_MUL(ay1, vx);
-        GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
+        for (int i = 0; i < np; i += ggml_f16_step) {
+            ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
+            ay1 = GGML_F16x_VEC_MUL(ay1, vx);
+            GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
 
-        ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
-        ay2 = GGML_F16x_VEC_MUL(ay2, vx);
-        GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
-    }
-    // leftovers
-    // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
-    if (np < n) {
-        svbool_t pg = svwhilelt_b16(np, n);
-        svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
-        svfloat16_t out = svmul_f16_m(pg, hy, vx);
-        svst1_f16(pg, (__fp16 *)(y + np), out);
-    }
-#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
-    for (int i = 0, vl; i < n; i += vl) {
-        vl = __riscv_vsetvl_e16m2(n - i);
-        vfloat16m2_t vy = __riscv_vle16_v_f16m2((_Float16 *)&y[i], vl);
-        vfloat32m4_t vy32 = __riscv_vfwcvt_f_f_v_f32m4(vy, vl);
-        vy32 = __riscv_vfmul_vf_f32m4(vy32, v, vl);
-        vy = __riscv_vfncvt_f_f_w_f16m2(vy32, vl);
-        __riscv_vse16_v_f16m2((_Float16 *)&y[i], vy, vl);
-    }
-#elif defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F16_STEP - 1));
-
-    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-    GGML_F16_VEC ay[GGML_F16_ARR];
-
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-            ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
-
-            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+            ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
+            ay2 = GGML_F16x_VEC_MUL(ay2, vx);
+            GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
         }
-    }
+        // leftovers
+        // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
+        if (np < n) {
+            svbool_t pg = svwhilelt_b16(np, n);
+            svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
+            svfloat16_t out = svmul_f16_m(pg, hy, vx);
+            svst1_f16(pg, (__fp16 *)(y + np), out);
+        }
+    #elif defined(__riscv_v_intrinsic)
+        // todo: RVV impl
+        // scalar
+        for (int i = 0; i < n; ++i) {
+            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+        }
+    #else
+        const int np = (n & ~(GGML_F16_STEP - 1));
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-    }
+        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+        GGML_F16_VEC ay[GGML_F16_ARR];
+
+        for (int i = 0; i < np; i += GGML_F16_STEP) {
+            for (int j = 0; j < GGML_F16_ARR; j++) {
+                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+                ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {

ggml/src/ggml-cuda/argsort.cu

@@ -44,7 +44,7 @@ static void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
     const dim3 offset_grid((nrows + block_size - 1) / block_size);
     init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
 
-    CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));
+    cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream);
 
     size_t temp_storage_bytes = 0;
 

ggml/src/ggml-cuda/common.cuh

@@ -21,12 +21,10 @@
 #include "ggml-common.h"
 
 #include <array>
-#include <algorithm>
 #include <cassert>
 #include <cfloat>
 #include <cstdio>
 #include <string>
-#include <unordered_map>
 #include <vector>
 
 #if defined(GGML_USE_HIP)
@@ -86,12 +84,12 @@
 
 #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
 #define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_PH1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // MTT S5000
+#define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
-#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_PH1)
-#define GGML_CUDA_CC_IS_PH1(cc)      (cc >= GGML_CUDA_CC_PH1)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
+#define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
 #    define GGML_CUDA_USE_CUB
@@ -214,9 +212,9 @@ static const char * cu_get_error_str(CUresult err) {
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 
-#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
@@ -226,10 +224,6 @@ 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
@@ -252,14 +246,12 @@ static const char * cu_get_error_str(CUresult err) {
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
 static bool fp16_available(const int cc) {
-    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
-        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
+    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
 }
 
 static bool fast_fp16_available(const int cc) {
     return GGML_CUDA_CC_IS_AMD(cc) ||
-        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) ||
-        (GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc));
+        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -276,9 +268,7 @@ static bool fp16_mma_hardware_available(const int cc) {
 }
 
 static bool bf16_mma_hardware_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3 ||
-        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
 }
 
 static bool fp32_mma_hardware_available(const int cc) {
@@ -293,10 +283,6 @@ 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;
 }
@@ -564,12 +550,8 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
     acc += v.y*u.y;
 }
 
-#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
-#define V_DOT2_F32_F16_AVAILABLE
-#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
-
 static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
     asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
 #else
 #ifdef FAST_FP16_AVAILABLE
@@ -581,7 +563,7 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
     acc += tmpv.x * tmpu.x;
     acc += tmpv.y * tmpu.y;
 #endif // FAST_FP16_AVAILABLE
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
 }
 
 static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {
@@ -982,154 +964,6 @@ struct ggml_cuda_graph {
 #endif
 };
 
-struct ggml_cuda_concurrent_event {
-    std::vector<cudaEvent_t> join_events;
-    cudaEvent_t              fork_event = nullptr;
-
-    int                                          n_streams = 0;
-    std::unordered_map<const ggml_tensor *, int> stream_mapping;
-
-    const ggml_tensor * join_node;
-
-    ggml_cuda_concurrent_event() = default;
-
-    ggml_cuda_concurrent_event(const ggml_cuda_concurrent_event &) = delete;
-    ggml_cuda_concurrent_event & operator=(const ggml_cuda_concurrent_event &) = delete;
-
-    explicit ggml_cuda_concurrent_event(int n_streams) : n_streams(n_streams) {
-        join_events.resize(n_streams);
-
-        for (size_t i = 0; i < join_events.size(); ++i) {
-            CUDA_CHECK(cudaEventCreateWithFlags(&join_events[i], cudaEventDisableTiming));
-        }
-
-        CUDA_CHECK(cudaEventCreateWithFlags(&fork_event, cudaEventDisableTiming));
-    }
-
-    ggml_cuda_concurrent_event(ggml_cuda_concurrent_event && other) noexcept
-    : join_events(std::move(other.join_events))
-    , fork_event(other.fork_event)
-    , n_streams(other.n_streams)
-    , stream_mapping(std::move(other.stream_mapping))
-    , join_node(other.join_node) {
-        other.fork_event = nullptr;
-    }
-
-    // 1. check if any branches write to overlapping memory ranges (except the join node)
-    // 2. check whether all srcs are either within the branch or outside the nodes covered by ggml_cuda_concurrent_event
-    // we assume all nodes have the same buffer
-    bool is_valid() const {
-        std::vector<std::vector<std::pair<int64_t, int64_t>>> write_ranges;
-        write_ranges.resize(n_streams);
-
-        // get join_node's memory range to exclude from overlap checking.
-        // multiple nodes can use join_node's buffer; we synchronize on the join node.
-        const ggml_tensor * join_t     = join_node->view_src ? join_node->view_src : join_node;
-        const int64_t       join_start = (int64_t) join_t->data;
-        const int64_t       join_end   = join_start + ggml_nbytes(join_t);
-
-        for (const auto & [tensor, stream] : stream_mapping) {
-            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
-            const int64_t       t_start = (int64_t) t->data;
-            const int64_t       t_end   = t_start + ggml_nbytes(t);
-
-            // skip tensors that overlap with join_node's buffer.
-            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
-                continue;
-            }
-
-            // concurrent streams begin from 1
-            write_ranges[stream - 1].emplace_back(t_start, t_end);
-        }
-
-        for (int i = 0; i < n_streams; ++i) {
-            // sorts first by start then by end of write range
-            std::sort(write_ranges[i].begin(), write_ranges[i].end());
-        }
-
-        bool writes_overlap = false;
-        bool dependent_srcs = false;
-        for (const auto & [tensor, stream] : stream_mapping) {
-            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
-            const int64_t       t_start = (int64_t) t->data;
-            const int64_t       t_end   = t_start + ggml_nbytes(t);
-
-            // skip tensors that overlap with join_node's buffer
-            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
-                continue;
-            }
-
-            // check if this buffer's write data overlaps with another stream's
-            std::pair<int64_t, int64_t> data_range = std::make_pair(t_start, t_end);
-            for (int i = 0; i < n_streams; ++i) {
-                if (i == stream - 1) {
-                    continue;
-                }
-                auto it = std::lower_bound(write_ranges[i].begin(), write_ranges[i].end(), data_range);
-
-                if (it != write_ranges[i].end()) {
-                    const std::pair<int64_t, int64_t> & other = *it;
-
-                    // std::lower_bound returns the first element where other >= data_range (lexicographically).
-                    // This guarantees other.first >= data_range.first.
-                    // Therefore, overlap occurs iff other.first < data_range.second
-                    // (i.e., the other range starts before this range ends).
-                    if (other.first < data_range.second) {
-                        GGML_LOG_DEBUG("Writes overlap for %s", tensor->name);
-                        writes_overlap = true;
-                        break;
-                    }
-                }
-            }
-
-            //check if all srcs are either in branch or don't have a branch
-            for (int i = 0; i < GGML_MAX_SRC; ++i) {
-                if (!tensor->src[i]) {
-                    continue;
-                }
-
-                auto it = stream_mapping.find(tensor->src[i]);
-
-                if (it == stream_mapping.end()) {
-                    continue;
-                }
-
-                if (it->second != stream) {
-                    dependent_srcs = true;
-                    break;
-                }
-            }
-
-            if (dependent_srcs || writes_overlap) {
-                break;
-            }
-        }
-
-        return !writes_overlap && !dependent_srcs;
-    }
-
-    ~ggml_cuda_concurrent_event() {
-        if (fork_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(fork_event));
-        }
-        for (cudaEvent_t e : join_events) {
-            if (e != nullptr) {
-                CUDA_CHECK(cudaEventDestroy(e));
-            }
-        }
-    }
-};
-
-struct ggml_cuda_stream_context {
-    std::vector<const ggml_tensor *>                                    original_nodes;
-    std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
-
-    void reset() {
-        original_nodes.clear();
-        concurrent_events.clear();
-    }
-};
-
 struct ggml_backend_cuda_context {
     int device;
     std::string name;
@@ -1140,15 +974,11 @@ struct ggml_backend_cuda_context {
 
     std::unique_ptr<ggml_cuda_graph> cuda_graph;
 
-    int curr_stream_no = 0;
-
     explicit ggml_backend_cuda_context(int device) :
         device(device),
         name(GGML_CUDA_NAME + std::to_string(device)) {
     }
 
-    ggml_cuda_stream_context concurrent_stream_context;
-
     ~ggml_backend_cuda_context();
 
     cudaStream_t stream(int device, int stream) {
@@ -1159,9 +989,9 @@ struct ggml_backend_cuda_context {
         return streams[device][stream];
     }
 
-    cudaStream_t stream() { return stream(device, curr_stream_no); }
-
-    ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }
+    cudaStream_t stream() {
+        return stream(device, 0);
+    }
 
     cublasHandle_t cublas_handle(int device) {
         if (cublas_handles[device] == nullptr) {
@@ -1177,15 +1007,15 @@ struct ggml_backend_cuda_context {
     }
 
     // pool
-    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];
+    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
 
-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, int stream_no);
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
 
     ggml_cuda_pool & pool(int device) {
-        if (pools[device][curr_stream_no] == nullptr) {
-            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no);
+        if (pools[device] == nullptr) {
+            pools[device] = new_pool_for_device(device);
         }
-        return *pools[device][curr_stream_no];
+        return *pools[device];
     }
 
     ggml_cuda_pool & pool() {

ggml/src/ggml-cuda/convert.cuh

@@ -39,15 +39,6 @@ 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_scalar(const char * cxi, char * cdsti) {
+static __device__ void cpy_1_flt(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_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) {
+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) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -40,7 +40,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int ne,
 }
 
 template <typename T>
-static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_flt_transpose(const char * cx, char * cdst, const int ne,
                                const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                const int nb12, const int nb13) {
@@ -86,9 +86,6 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
             }
         }
     }
-
-    GGML_UNUSED_VARS(ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11,
-        nb12, nb13);
 }
 
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
@@ -169,7 +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_scalar_contiguous(const char * cx, char * cdst, const int64_t ne) {
+static __global__ void cpy_flt_contiguous(const char * cx, char * cdst, const int64_t ne) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -183,17 +180,17 @@ static __global__ void cpy_scalar_contiguous(const char * cx, char * cdst, const
 }
 
 template<typename src_t, typename dst_t>
-static void ggml_cpy_scalar_contiguous_cuda(
+static void ggml_cpy_flt_contiguous_cuda(
     const char * cx, char * cdst, const int64_t ne,
 cudaStream_t stream) {
 
     const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_scalar_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+    cpy_flt_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
         (cx, cdst, ne);
 }
 
 template<typename src_t, typename dst_t, bool transposed = false>
-static void ggml_cpy_scalar_cuda(
+static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
@@ -205,7 +202,7 @@ static void ggml_cpy_scalar_cuda(
             ne00n = ne00;
             ne01n = ne01;
             ne02n = ne02;
-        } else {
+        } else if (nb00 > nb02) {
             ne00n = ne00;
             ne01n = ne01*ne02;
             ne02n = 1;
@@ -215,11 +212,11 @@ static void ggml_cpy_scalar_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_scalar_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
+        cpy_flt_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
             (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     } else {
         const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-        cpy_scalar<cpy_1_scalar<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
             (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     }
 }
@@ -387,8 +384,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src1_ddc = (char *) src1->data;
 
     const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
-    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) &&
-        src0->ne[3] == 1 && nb02 == ne00 * ne01 * (int64_t)ggml_element_size(src0);
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) && src0->ne[3] == 1;
 
     if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
@@ -402,132 +398,94 @@ 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_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);
+            ggml_cpy_flt_cuda<float, float, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         } else {
-            ggml_cpy_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);
+            ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<float, nv_bfloat16>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<float, half>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<float, half>        (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<float, half>        (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda
-                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        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_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);
+            ggml_cpy_flt_cuda<half, half, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         } else {
-            ggml_cpy_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);
+            ggml_cpy_flt_cuda<half, half>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<half, nv_bfloat16>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<half, nv_bfloat16>  (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<half, nv_bfloat16>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<half, float>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<half, float>        (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<half, float>          (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
         if (can_be_transposed) {
-            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);
+            ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         } else {
-            ggml_cpy_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);
+            ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<nv_bfloat16, half>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, half>  (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<nv_bfloat16, half>    (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<nv_bfloat16, float>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<nv_bfloat16, float>   (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<float, int32_t>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<float, int32_t>     (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<float, int32_t>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
         if (contiguous_srcs) {
-            ggml_cpy_scalar_contiguous_cuda<int32_t, float>
-                (src0_ddc, src1_ddc, ne, main_stream);
+            ggml_cpy_flt_contiguous_cuda<int32_t, float>     (src0_ddc, src1_ddc, ne, main_stream);
         } else {
-            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);
+            ggml_cpy_flt_cuda<int32_t, float>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,

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

@@ -55,11 +55,11 @@ static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f16(
         ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_h2 + k_KQ_0 + (threadIdx.x % nthreads)*cpy_ne);
 #pragma unroll
         for (int k_KQ_1 = 0; k_KQ_1 < cpy_ne; ++k_KQ_1) {
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
             ggml_cuda_mad(sum,                tmp[k_KQ_1] , ((const half2  *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
 #else
             ggml_cuda_mad(sum, __half22float2(tmp[k_KQ_1]), ((const float2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FP16_AVAILABLE
         }
     }
 

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

@@ -609,7 +609,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
             float KQ_sum_add = 0.0f;
 #pragma unroll
             for (int i0 = 0; i0 < nbatch_fa; i0 += np*warp_size) {
-                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < static_cast<uint32_t>(k_VKQ_sup) ?
+                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < k_VKQ_sup ?
                     expf(KQ_acc[(i0/(np*warp_size))*cpw + jc] - KQ_max[jc]) : 0.0f;
                 KQ_sum_add += val;
                 tmp[i0/(np*warp_size)][jc1] = val;

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

@@ -86,11 +86,11 @@ static __global__ void flash_attn_ext_vec(
 
     constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half,  V_rows_per_thread>();
 #else
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, float, V_rows_per_thread>();
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
@@ -112,13 +112,13 @@ static __global__ void flash_attn_ext_vec(
 
     constexpr int ne_KQ      = ncols*D;
     constexpr int ne_combine = nwarps*V_cols_per_iter*D;
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
     half2            VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
     __shared__ half   KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
 #else
     float2           VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
     __shared__ float  KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
 
     float KQ_max[ncols];
     float KQ_sum[ncols];
@@ -129,11 +129,11 @@ static __global__ void flash_attn_ext_vec(
     }
 
     // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
     half2  Q_reg[ncols][(D/2)/nthreads_KQ]; // Will be initialized completely.
 #else
     float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
     int    Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     float2  Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     if constexpr (Q_q8_1) {
@@ -155,7 +155,7 @@ static __global__ void flash_attn_ext_vec(
                 for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
                     const int i = i0 + threadIdx.x;
 
-                    if (i0 + WARP_SIZE <= int(D/sizeof(int)) || i < int(D/sizeof(int))) {
+                    if (i0 + WARP_SIZE <= D/sizeof(int) || i < D/sizeof(int)) {
                         tmp_q_i32[i] = 0;
                     }
                 }
@@ -191,7 +191,7 @@ static __global__ void flash_attn_ext_vec(
 
         __syncthreads();
     } else {
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
         const half2 scale_h2 = make_half2(scale, scale);
 #pragma unroll
         for (int j = 0; j < ncols; ++j) {
@@ -233,7 +233,7 @@ static __global__ void flash_attn_ext_vec(
                 Q_reg[j][k].y *= scale;
             }
         }
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
     }
 
     const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
@@ -272,7 +272,7 @@ static __global__ void flash_attn_ext_vec(
 
                 KQ_max_new[j] = fmaxf(KQ_max_new[j], sum);
 
-                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == uint32_t(i_KQ_0)) {
+                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == i_KQ_0) {
                     KQ_reg[j] = sum;
                 }
             }
@@ -291,7 +291,7 @@ static __global__ void flash_attn_ext_vec(
             KQ_sum[j] = KQ_sum[j]*KQ_max_scale + KQ_reg[j];
             KQ[j*nthreads + tid] = KQ_reg[j];
 
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
             const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
@@ -303,7 +303,7 @@ static __global__ void flash_attn_ext_vec(
                 VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
                 VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
             }
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
         }
 
 #ifndef GGML_USE_HIP
@@ -314,7 +314,7 @@ static __global__ void flash_attn_ext_vec(
         for (int k0 = 0; k0 < WARP_SIZE; k0 += V_cols_per_iter) {
             const int k = threadIdx.y*WARP_SIZE + k0 + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V);
 
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
             half2 KQ_k[ncols];
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
@@ -353,7 +353,7 @@ static __global__ void flash_attn_ext_vec(
                     }
                 }
             }
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
         }
     }
 
@@ -374,7 +374,7 @@ static __global__ void flash_attn_ext_vec(
 
             KQ_sum[j] = KQ_sum[j]*KQ_max_scale + (threadIdx.x == 0 ? expf(sink - KQ_max[j]) : 0.0f);
 
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
             const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
@@ -386,7 +386,7 @@ static __global__ void flash_attn_ext_vec(
                 VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
                 VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
             }
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
         }
     }
 
@@ -421,7 +421,7 @@ static __global__ void flash_attn_ext_vec(
         const float kqmax_scale = expf(KQ_max[j_VKQ] - kqmax_new);
         KQ_max[j_VKQ] = kqmax_new;
 
-#ifdef V_DOT2_F32_F16_AVAILABLE
+#ifdef FAST_FP16_AVAILABLE
         half2 * VKQ_tmp = (half2 *) KQ + threadIdx.y*(V_cols_per_iter*D/2)
             + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V)*(D/2);
 
@@ -452,7 +452,7 @@ static __global__ void flash_attn_ext_vec(
             ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ,                       &VKQ[j_VKQ][i_VKQ_0/nthreads_V]);
             ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ + V_rows_per_thread/4, &VKQ[j_VKQ][i_VKQ_0/nthreads_V + V_rows_per_thread/4]);
         }
-#endif // V_DOT2_F32_F16_AVAILABLE
+#endif // FAST_FP16_AVAILABLE
 
         KQ_sum[j_VKQ] *= kqmax_scale;
         KQ_sum[j_VKQ] = warp_reduce_sum(KQ_sum[j_VKQ]);

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

@@ -53,7 +53,6 @@
 #include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
-#include "ggml-cuda/solve_tri.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -522,8 +521,7 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
 };
 #endif // defined(GGML_USE_VMM)
 
-std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int                  device,
-                                                                               [[maybe_unused]] int stream_no) {
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
 #if defined(GGML_USE_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
@@ -2719,9 +2717,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_SGD:
             ggml_cuda_opt_step_sgd(ctx, dst);
             break;
-        case GGML_OP_SOLVE_TRI:
-            ggml_cuda_op_solve_tri(ctx, dst);
-            break;
         default:
             return false;
     }
@@ -3006,10 +3001,6 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
                                                 const ggml_tensor * view,
                                                 const ggml_tensor * set_rows) {
-
-    if (rope->op != GGML_OP_ROPE || view->op != GGML_OP_VIEW || set_rows->op != GGML_OP_SET_ROWS) {
-        return false;
-    }
     // ne3 not tested
     if (rope->src[0]->ne[3] != 1) {
         return false;
@@ -3051,12 +3042,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     std::initializer_list<enum ggml_op> topk_moe_ops_delayed_softmax =
         ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
 
-    const auto is_equal = [](const std::initializer_list<enum ggml_op> & list1,
-                             const std::initializer_list<enum ggml_op> & list2) {
-        return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end());
-    };
-
-    if (is_equal(topk_moe_ops_with_norm, ops) &&
+    if (ops.size() == topk_moe_ops_with_norm.size() &&
         ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 9];
@@ -3066,7 +3052,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
+    if (ops.size() == topk_moe_ops.size() &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 4];
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
@@ -3074,7 +3061,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (is_equal(topk_moe_ops_delayed_softmax, ops) &&
+    if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
         ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
         ggml_tensor * weights = cgraph->nodes[node_idx + 5];
@@ -3090,8 +3077,9 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     std::initializer_list<enum ggml_op> mul_mat_id_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_MUL_MAT_ID, GGML_OP_GLU };
     std::initializer_list<enum ggml_op> mul_mat_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_MUL_MAT,    GGML_OP_GLU };
 
-    if ((is_equal(mul_mat_bias_glu_ops, ops) || is_equal(mul_mat_id_bias_glu_ops, ops)) &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) {
+    if (ops.size() == 5 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}) ||
+                            ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}))) {
+
         const ggml_tensor * ffn_gate      = cgraph->nodes[node_idx];
         const ggml_tensor * ffn_gate_bias = cgraph->nodes[node_idx + 1];
         const ggml_tensor * ffn_up        = cgraph->nodes[node_idx + 2];
@@ -3103,8 +3091,9 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if ((is_equal(mul_mat_id_glu_ops, ops) || is_equal(mul_mat_glu_ops, ops)) &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+    if (ops.size() == 3 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}) ||
+                            ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}))) {
+
         const ggml_tensor * ffn_gate = cgraph->nodes[node_idx];
         const ggml_tensor * ffn_up   = cgraph->nodes[node_idx + 1];
         const ggml_tensor * glu      = cgraph->nodes[node_idx + 2];
@@ -3114,9 +3103,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    std::initializer_list<enum ggml_op> rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };
-
-    if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+    if (ops.size() == 3 && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
         const ggml_tensor * rope     = cgraph->nodes[node_idx];
         const ggml_tensor * view     = cgraph->nodes[node_idx + 1];
         const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];
@@ -3201,83 +3188,18 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
     // flag used to determine whether it is an integrated_gpu
     const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
-    ggml_cuda_stream_context & stream_ctx = cuda_ctx->stream_context();
-    bool                         is_concurrent_event_active = false;
-    ggml_cuda_concurrent_event * concurrent_event           = nullptr;
-    bool                         should_launch_concurrent_events = false;
-
-    const auto try_launch_concurrent_event = [&](const ggml_tensor * node) {
-        if (stream_ctx.concurrent_events.find(node) != stream_ctx.concurrent_events.end()) {
-            concurrent_event = &stream_ctx.concurrent_events[node];
-
-            is_concurrent_event_active = true;
-
-            GGML_LOG_DEBUG("Launching %d streams at %s\n", concurrent_event->n_streams, node->name);
-
-            cudaStream_t main_stream = cuda_ctx->stream();  // this should be stream 0
-            GGML_ASSERT(cuda_ctx->curr_stream_no == 0);
-            CUDA_CHECK(cudaEventRecord(concurrent_event->fork_event, main_stream));
-
-            for (int i = 1; i <= concurrent_event->n_streams; ++i) {
-                cudaStream_t stream = cuda_ctx->stream(cuda_ctx->device, i);
-                CUDA_CHECK(cudaStreamWaitEvent(stream, concurrent_event->fork_event));
-            }
-        }
-    };
-
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
-            [[maybe_unused]] int prev_i = 0;
 
-            if (stream_ctx.concurrent_events.size() > 0) {
-                should_launch_concurrent_events = true;
-                for (const auto & [tensor, event] : stream_ctx.concurrent_events) {
-                    should_launch_concurrent_events = should_launch_concurrent_events && event.is_valid();
-                }
-            }
-            if (should_launch_concurrent_events) {
-                //Restore the original graph to enable fusion within the streams
-                cgraph->nodes   = const_cast<ggml_tensor **>(stream_ctx.original_nodes.data());
-                cgraph->n_nodes = (int) stream_ctx.original_nodes.size();
-            }
+            [[maybe_unused]] int prev_i = 0;
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
-                if (is_concurrent_event_active) {
-                    GGML_ASSERT(concurrent_event);
-
-                    if (node == concurrent_event->join_node) {
-                        cuda_ctx->curr_stream_no = 0;
-                        for (int i = 1; i <= concurrent_event->n_streams; ++i) {
-                            // Wait on join events of forked streams in the main stream
-                            CUDA_CHECK(cudaEventRecord(concurrent_event->join_events[i - 1],
-                                                       cuda_ctx->stream(cuda_ctx->device, i)));
-                            CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), concurrent_event->join_events[i - 1]));
-                        }
-
-                        is_concurrent_event_active = false;
-                        concurrent_event           = nullptr;
-                    } else {
-                        GGML_ASSERT (concurrent_event->stream_mapping.find(node) != concurrent_event->stream_mapping.end());
-                        cuda_ctx->curr_stream_no = concurrent_event->stream_mapping[node];
-                        GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
-                    }
-                } else if (i - prev_i > 1) {
-                    //the previous node was fused
-                    const ggml_tensor * prev_node = cgraph->nodes[i - 1];
-                    try_launch_concurrent_event(prev_node);
-
-                    if (is_concurrent_event_active) {
-                        cuda_ctx->curr_stream_no = concurrent_event->stream_mapping[node];
-                        GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
-                    }
-                }
-                prev_i = i;
-
 #ifdef GGML_CUDA_DEBUG
                 const int nodes_fused = i - prev_i - 1;
+                prev_i = i;
                 if (nodes_fused > 0) {
                     GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
                 }
@@ -3287,8 +3209,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     continue;
                 }
 
-
-                // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
 
@@ -3581,17 +3501,13 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 }
 #else
                 GGML_UNUSED(integrated);
-#endif  // NDEBUG
+#endif // NDEBUG
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
                     GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                 }
                 GGML_ASSERT(ok);
-
-                if (!is_concurrent_event_active) {
-                    try_launch_concurrent_event(node);
-               }
             }
         }
 
@@ -3731,235 +3647,6 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
     }
 }
 
-static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
-
-    static bool enable_graph_optimization = [] {
-        const char * env = getenv("GGML_CUDA_GRAPH_OPT");
-        return env != nullptr && atoi(env) == 1;
-    }();
-
-    if (!enable_graph_optimization) {
-        return;
-    }
-
-    GGML_ASSERT(ggml_backend_cuda_get_device_count() == 1 && "compute graph optimization is only supported on single GPU in the CUDA backend");
-    GGML_LOG_DEBUG("Optimizing CUDA graph %p with %d nodes\n", cgraph->nodes, cgraph->n_nodes);
-
-    ggml_cuda_stream_context & stream_context = cuda_ctx->stream_context();
-    stream_context.reset();
-
-    // number of out-degrees for a particular node
-    std::unordered_map<const ggml_tensor *, int> fan_out;
-    // reverse mapping of node to index in the cgraph
-    std::unordered_map<const ggml_tensor *, int> node_indices;
-
-    const auto & is_noop = [](const ggml_tensor * node) -> bool {
-        return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE ||
-               node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
-    };
-
-    const auto & depends_on = [](const ggml_tensor * dst, const ggml_tensor * src) -> bool {
-        for (uint32_t s = 0; s < GGML_MAX_SRC; ++s) {
-            if (dst->src[s] == src) {
-                return true;
-            }
-        }
-        // implicit dependency if they view the same tensor
-        const ggml_tensor * dst2 = dst->view_src ? dst->view_src : dst;
-        const ggml_tensor * src2 = src->view_src ? src->view_src : src;
-        if (dst2 == src2) {
-            return true;
-        }
-        return false;
-    };
-
-    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
-        const ggml_tensor * node = cgraph->nodes[node_idx];
-        node_indices[node]       = node_idx;
-
-        if (is_noop(node)) {
-            continue;
-        }
-        for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
-            const ggml_tensor * src = cgraph->nodes[node_idx]->src[src_idx];
-            //TODO: check why nrows > 1 fails
-            if (node && !is_noop(node) && ggml_nrows(node) <= 1) {
-                fan_out[src] += 1;
-            }
-        }
-    }
-
-    // Target Q, K, V for concurrency
-    // this is a more general way to find nodes which can be candidates for concurrency (although it has not been tested for anything else):
-    // 1. find fan-out (fork) nodes where the same input is used at least N times (in QKV, it would be "attn-norm")
-    // 2. find the join node, where 2 or more of the outputs are required (in QKV, this would "KQ" or "flash-attn")
-    // 3. account for all branches from the fork to the join
-    // 4. To extend lifetimes of the tensors, we interleave the branches (see below for more details)
-    // 5. save the original cgraph and restore it in graph_compute, to enable fusion within streams
-    // See discussion: https://github.com/ggml-org/llama.cpp/pull/16991#issuecomment-3522620030
-
-    const int min_fan_out = 3;
-    const int max_fan_out = 3;
-
-    // store {fork_idx, join_idx}
-    std::vector<std::pair<int, int>> concurrent_node_ranges;
-
-    // save the original nodes
-    std::vector<const ggml_tensor *> original_nodes;
-    original_nodes.reserve(cgraph->n_nodes);
-    for (int i = 0; i < cgraph->n_nodes; ++i) {
-        original_nodes.push_back(cgraph->nodes[i]);
-    }
-    cuda_ctx->stream_context().original_nodes = std::move(original_nodes);
-
-    for (const auto & [root_node, count] : fan_out) {
-        if (count >= min_fan_out && count <= max_fan_out) {
-            const int root_node_idx = node_indices[root_node];
-
-            bool is_part_of_event = false;
-            for (const auto & [start, end] : concurrent_node_ranges) {
-                if (root_node_idx >= start && root_node_idx <= end) {
-                    is_part_of_event = true;
-                }
-            }
-
-            if (is_part_of_event) {
-                continue;
-            }
-
-            std::vector<std::vector<const ggml_tensor *>> nodes_per_branch;
-            for (int i = root_node_idx + 1; i < cgraph->n_nodes; ++i) {
-                const ggml_tensor * node = cgraph->nodes[i];
-                if (!is_noop(node) && depends_on(node, root_node)) {
-                    nodes_per_branch.push_back({ node });
-                }
-            }
-
-            GGML_ASSERT(nodes_per_branch.size() == (size_t) count);
-
-            //find the join point
-            const ggml_tensor * join_node = nullptr;
-
-            const auto & belongs_to_branch = [&](const ggml_tensor *                      node,
-                                                 const std::vector<const ggml_tensor *> & branch) -> bool {
-                for (const ggml_tensor * n : branch) {
-                    if (depends_on(node, n)) {
-                        return true;
-                    }
-                }
-                return false;
-            };
-
-            for (int i = root_node_idx + 1; i < cgraph->n_nodes; ++i) {
-                const ggml_tensor * curr_node = cgraph->nodes[i];
-
-                int num_joins = 0;
-                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
-                    if (belongs_to_branch(curr_node, nodes_per_branch[branch_idx])) {
-                        num_joins++;
-                    }
-                }
-
-                if (num_joins >= 2) {
-                    join_node = curr_node;
-                    break;
-                }
-
-                bool found_branch = false;
-                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
-                    std::vector<const ggml_tensor *> & branch_vec = nodes_per_branch[branch_idx];
-                    if (belongs_to_branch(curr_node, branch_vec)) {
-                        //continue accumulating
-                        if (std::find(branch_vec.begin(), branch_vec.end(), curr_node) == branch_vec.end()) {
-                            branch_vec.push_back(curr_node);
-                        }
-                        found_branch = true;
-                    }
-                }
-
-                if (!found_branch && is_noop(curr_node)) {
-                    // we can put it in any branch because it will be ignored
-                    nodes_per_branch[0].push_back({ curr_node });
-                }
-            }
-
-            if (join_node) {
-                //Create ggml_cuda_concurrent_event
-                ggml_cuda_concurrent_event concurrent_event(nodes_per_branch.size());
-                concurrent_event.join_node = join_node;
-
-                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
-                    for (const ggml_tensor * n : nodes_per_branch[branch_idx]) {
-                        concurrent_event.stream_mapping[n] = branch_idx + 1;
-                    }
-                }
-
-                int fork_node_idx = node_indices[root_node];
-                int join_node_idx = node_indices[join_node];
-
-                int       current_branch_idx = 0;
-                int       current_node_idx   = fork_node_idx + 1;
-                const int n_branches         = nodes_per_branch.size();
-
-                int total_branch_nodes = 0;
-                for (std::vector<const ggml_tensor *> branch_nodes : nodes_per_branch) {
-                    total_branch_nodes += branch_nodes.size();
-                }
-
-                // there are other nodes in the middle which are unaccounted for
-                // usually (cpy) nodes, then ignore this fork
-                if (join_node_idx - fork_node_idx - 1 != total_branch_nodes) {
-                    GGML_LOG_DEBUG(
-                        "Skipping %s because the number of nodes in the middle is not equal to the total number of "
-                        "branch nodes %d != %d\n",
-                        root_node->name, join_node_idx - fork_node_idx - 1, total_branch_nodes);
-                    continue;
-                }
-
-                std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> & concurrent_events = cuda_ctx->stream_context().concurrent_events;
-                GGML_ASSERT(concurrent_events.find(root_node) == concurrent_events.end());
-                concurrent_events.emplace(root_node, std::move(concurrent_event));
-                GGML_LOG_DEBUG("Adding stream at node %s %p\n", root_node->name, root_node);
-                concurrent_node_ranges.emplace_back(fork_node_idx, join_node_idx);
-
-                // interleave tensors to extend lifetimes so that ggml graph doesn't recycle them
-                // example transformation:
-                // [attn-norm, QMul, QNorm, QRope, KMul, KNorm, KRope, VMul, attn] ->
-                // [attn-norm, QMul, KMul, VMul, QNorm, VNorm, QRope, KRope, attn]
-                while (current_node_idx < join_node_idx) {
-                    std::vector<const ggml_tensor *> & branch_nodes = nodes_per_branch[current_branch_idx];
-
-                    bool has_node = false;
-                    for (std::vector<const ggml_tensor *> branch_node : nodes_per_branch) {
-                        has_node |= branch_node.size() > 0;
-                    }
-
-                    GGML_ASSERT(has_node);
-
-                    if (branch_nodes.empty()) {
-                        current_branch_idx = (current_branch_idx + 1) % n_branches;
-                        continue;
-                    }
-
-                    cgraph->nodes[current_node_idx] = const_cast<ggml_tensor *>(branch_nodes.front());
-                    current_node_idx++;
-                    branch_nodes.erase(branch_nodes.begin());
-
-                    // append all empty nodes
-                    while (!branch_nodes.empty() && is_noop(branch_nodes.front())) {
-                        cgraph->nodes[current_node_idx] = const_cast<ggml_tensor *>(branch_nodes.front());
-                        current_node_idx++;
-                        branch_nodes.erase(branch_nodes.begin());
-                    }
-
-                    current_branch_idx = (current_branch_idx + 1) % n_branches;
-                }
-            }
-        }
-    }
-}
-
 static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .get_name                = */ ggml_backend_cuda_get_name,
     /* .free                    = */ ggml_backend_cuda_free,
@@ -3974,7 +3661,7 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .event_record            = */ ggml_backend_cuda_event_record,
     /* .event_wait              = */ ggml_backend_cuda_event_wait,
-    /* .graph_optimize          = */ ggml_backend_cuda_graph_optimize,
+    /* .graph_optimize          = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_cuda_guid() {
@@ -4057,110 +3744,10 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
-#if defined(__linux__)
-// Helper function to get available memory from /proc/meminfo for UMA systems
-static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_kb, long * free_swap_kb) {
-    FILE * meminfo_file = nullptr;
-    // 2KB buffer for reading /proc/meminfo since it does not report size info, should be enough
-    const size_t BUFFER_SIZE = 2048;
-    auto file_buffer = std::make_unique<char[]>(BUFFER_SIZE);
-    size_t bytes_read = 0;
-    long huge_tlb_total_pages = -1;
-    long huge_tlb_free_pages = -1;
-    long huge_tlb_page_size = -1;
-
-    if (available_memory_kb == nullptr || free_swap_kb == nullptr) {
-        return false;
-    }
-
-    meminfo_file = fopen("/proc/meminfo", "r");
-    if (meminfo_file == nullptr) {
-        GGML_LOG_ERROR("%s: failed to open /proc/meminfo\n", __func__);
-        return false;
-    }
-
-    // Read file into buffer
-    bytes_read = fread(file_buffer.get(), 1, BUFFER_SIZE - 1, meminfo_file);
-    fclose(meminfo_file);
-
-    if (bytes_read == 0) {
-        GGML_LOG_ERROR("%s: failed to read from /proc/meminfo\n", __func__);
-        return false;
-    }
-    file_buffer[bytes_read] = '\0';
-
-    *available_memory_kb = -1;
-    *free_swap_kb = -1;
-
-    // Parse the file buffer line by line
-    char * line = file_buffer.get();
-    char * line_next;
-    while (line < file_buffer.get() + bytes_read) {
-        // Find the end of the current line
-        line_next = strchr(line, '\n');
-        if (line_next != nullptr) {
-            *line_next = '\0';
-            line_next++;
-        } else {
-            line_next = file_buffer.get() + bytes_read;
-        }
-
-        long value;
-        if (sscanf(line, "MemAvailable: %ld kB", &value) == 1) {
-            *available_memory_kb = value;
-        } else if (sscanf(line, "SwapFree: %ld kB", &value) == 1) {
-            *free_swap_kb = value;
-        } else if (sscanf(line, "HugePages_Total: %ld", &value) == 1) {
-            huge_tlb_total_pages = value;
-        } else if (sscanf(line, "HugePages_Free: %ld", &value) == 1) {
-            huge_tlb_free_pages = value;
-        } else if (sscanf(line, "Hugepagesize: %ld kB", &value) == 1) {
-            huge_tlb_page_size = value;
-        }
-
-        line = line_next;
-    }
-
-    if (huge_tlb_total_pages != 0 && huge_tlb_total_pages != -1) {
-        *available_memory_kb = huge_tlb_free_pages * huge_tlb_page_size;
-
-        // Hugetlbfs pages are not swappable.
-        *free_swap_kb = 0;
-    }
-
-    GGML_LOG_DEBUG("%s: final available_memory_kb: %ld\n", __func__, *available_memory_kb);
-    return true;
-}
-#endif // defined(__linux__)
-
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     ggml_cuda_set_device(ctx->device);
     CUDA_CHECK(cudaMemGetInfo(free, total));
-
-// ref: https://github.com/ggml-org/llama.cpp/pull/17368
-#if defined(__linux__)
-    // Check if this is a UMA (Unified Memory Architecture) system
-    cudaDeviceProp prop;
-    CUDA_CHECK(cudaGetDeviceProperties(&prop, ctx->device));
-
-    // Check if UMA is explicitly enabled via environment variable
-    bool uma_env = getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr;
-    bool is_uma = prop.integrated > 0 || uma_env;
-
-    if (is_uma) {
-        // For UMA systems (like DGX Spark), use system memory info
-        long available_memory_kb = 0;
-        long free_swap_kb = 0;
-
-        if (ggml_backend_cuda_get_available_uma_memory(&available_memory_kb, &free_swap_kb) && available_memory_kb > 0) {
-            *free = (size_t)available_memory_kb * 1024;
-        } else {
-            GGML_LOG_ERROR("%s: /proc/meminfo reading failed, using cudaMemGetInfo\n", __func__);
-        }
-    }
-#endif // defined(__linux__)
-
 }
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
@@ -4424,9 +4011,6 @@ 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;
                 }
@@ -4564,8 +4148,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return true;
-        case GGML_OP_SOLVE_TRI:
-            return op->src[0]->ne[0] <= 64 && op->src[1]->ne[0] <= 32;
         default:
             return false;
     }

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(AMD_MFMA_AVAILABLE)
+#if defined(GGML_USE_HIP)
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
@@ -149,34 +149,6 @@ 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};
@@ -264,32 +236,6 @@ 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}};
@@ -339,34 +285,6 @@ 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}};
 
@@ -402,7 +320,6 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif  // defined(AMD_WMMA_AVAILABLE)
     };
 
     template <int I, int J>
@@ -436,30 +353,6 @@ 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) {
@@ -746,34 +639,12 @@ 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)
@@ -794,36 +665,6 @@ 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;
@@ -850,7 +691,6 @@ namespace ggml_cuda_mma {
                                                      acc[0],
                                                      0, 0, 0);
 #endif // defined(CDNA3)
-
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -889,37 +729,10 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
             : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
 #else
-        tile      <16, 8, float> * D16 = reinterpret_cast<tile      <16, 8, float> *>(&D);
-        const tile<16, 8, half2> * A16 = reinterpret_cast<const tile<16, 8, half2> *>(&A);
+        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
+        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
         mma(D16[0], A16[0], B);
         mma(D16[1], A16[1], B);
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     }
-
-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

@@ -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) || amd_wmma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc);
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc) || amd_wmma_available(cc);
+            return ampere_mma_available(cc);
         default:
             return false;
     }

ggml/src/ggml-cuda/mmf.cuh

@@ -2,7 +2,6 @@
 
 #include "mma.cuh"
 #include "common.cuh"
-#include "convert.cuh"
 
 using namespace ggml_cuda_mma;
 
@@ -28,35 +27,20 @@ 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) {
-// 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
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
     constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
-    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 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;
@@ -177,11 +161,11 @@ 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] = ggml_cuda_cast<T>(tmp);
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
                     } 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] = ggml_cuda_cast<T>(tmp);
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
                     }
                 }
             } else {
@@ -255,7 +239,7 @@ 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)) || defined(AMD_WMMA_AVAILABLE)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
 //This kernel is for larger batch sizes of mul_mat_id
@@ -269,35 +253,20 @@ 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) {
-// 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
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
     constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
-    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.
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work butr 16 is ~1% faster.
 
     typedef tile<I_preferred, 8, T>     tile_A;
     typedef tile<8,           8, T>     tile_B;
     typedef tile<I_preferred, 8, float> tile_C;
-#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;
@@ -439,7 +408,7 @@ 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] = ggml_cuda_cast<T>(tmp);
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
                 }
 
                 if (itB + 1 < ntB) {
@@ -523,7 +492,7 @@ 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)) || defined(AMD_WMMA_AVAILABLE)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
 template<typename T, int cols_per_block, int nwarps>
@@ -585,8 +554,7 @@ 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<16, 8, T>     tile_B_16;
-    typedef tile< 8, 8, T>     tile_B_8;
+    typedef tile< 8, 8, T>     tile_B;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
     GGML_ASSERT(stride_row   % 2 == 0);
@@ -613,8 +581,7 @@ 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_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_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
     const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;

ggml/src/ggml-cuda/mmq.cu

@@ -306,11 +306,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    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;
+    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;
 }

ggml/src/ggml-cuda/solve_tri.cu

@@ -1,203 +0,0 @@
-#include "common.cuh"
-#include "ggml.h"
-#include "solve_tri.cuh"
-
-#define MAX_N_FAST 64
-#define MAX_K_FAST 32
-
-// ======================
-// Fast Kernel (n <= 64, k <= 32) - Warp-based parallel reduction
-// ======================
-// When ncols_template == 0 the bounds for the loops in this function are not
-// known and can't be unrolled. As we want to keep pragma unroll for all other
-// cases we supress the clang transformation warning here.
-#ifdef __clang__
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wpass-failed"
-#endif  // __clang__
-template <int n_template, int k_template>
-static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
-                                          const float * __restrict__ B,
-                                          float * __restrict__ X,
-                                          const uint3  ne02,
-                                          const size_t nb02,
-                                          const size_t nb03,
-                                          const size_t nb12,
-                                          const size_t nb13,
-                                          const size_t nb2,
-                                          const size_t nb3,
-                                          const int    n_arg,
-                                          const int    k_arg) {
-    const int n = n_template == 0 ? n_arg : n_template;
-    const int k = k_template == 0 ? k_arg : k_template;
-
-    const int batch_idx = blockIdx.x;
-    const int lane      = threadIdx.x;
-    const int col_idx   = threadIdx.y;
-
-    if (col_idx >= k) {
-        return;
-    }
-
-    const uint2   i02_i03 = fast_div_modulo(batch_idx, ne02);
-    const int64_t i02     = i02_i03.y;
-    const int64_t i03     = i02_i03.x;
-
-    const float * const A_batch = (const float *) (A + i02 * nb02 + i03 * nb03);
-    const float * const B_batch = (const float *) (B + i02 * nb12 + i03 * nb13);
-    float *             X_batch = (float *) (X + i02 * nb2 + i03 * nb3);
-
-    __shared__ float sA[MAX_N_FAST * MAX_N_FAST];
-    __shared__ float sXt[MAX_N_FAST * (MAX_K_FAST + 1)];
-
-    const int offset = threadIdx.x + threadIdx.y * blockDim.x;
-
-#pragma unroll
-    for (int i = 0; i < n * n; i += k * WARP_SIZE) {
-        int i0 = i + offset;
-        if (i0 < n * n) {
-            sA[i0] = A_batch[i0];
-        }
-    }
-
-    const int rows_per_warp = (n + WARP_SIZE - 1) / WARP_SIZE;
-
-#pragma unroll
-    for (int i = 0; i < rows_per_warp; i++) {
-        const int i0 = lane + i * WARP_SIZE;
-        if (i0 < n) {
-            sXt[col_idx * n + i0] = B_batch[i0 * k + col_idx];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int row = 0; row < n; ++row) {
-        float sum = 0.0f;
-
-        {
-            int j = lane;
-            if (j < row) {
-                sum += sA[row * n + j] * sXt[col_idx * n + j];
-            }
-        }
-        if (row >= WARP_SIZE) {
-            int j = WARP_SIZE + lane;
-            if (j < row) {
-                sum += sA[row * n + j] * sXt[col_idx * n + j];
-            }
-        }
-
-        sum = warp_reduce_sum(sum);
-
-        if (lane == 0) {
-            const float b_val      = sXt[col_idx * n + row];
-            const float a_diag     = sA[row * n + row];
-            // no safeguards for division by zero because that indicates corrupt
-            // data anyway
-            sXt[col_idx * n + row] = (b_val - sum) / a_diag;
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int i = 0; i < rows_per_warp; i++) {
-        const int i0 = lane + i * WARP_SIZE;
-        if (i0 < n) {
-            X_batch[i0 * k + col_idx] = sXt[col_idx * n + i0];
-        }
-    }
-}
-#ifdef __clang__
-#    pragma clang diagnostic pop
-#endif  // __clang__
-
-static void solve_tri_f32_cuda(const float * A,
-                               const float * B,
-                               float *       X,
-                               int           n,
-                               int           k,
-                               int64_t       ne02,
-                               int64_t       ne03,
-                               size_t        nb02,
-                               size_t        nb03,
-                               size_t        nb12,
-                               size_t        nb13,
-                               size_t        nb2,
-                               size_t        nb3,
-                               cudaStream_t  stream) {
-    const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
-    dim3        threads(WARP_SIZE, k);
-    dim3        grid(ne02 * ne03);
-    if (n == 64) {
-        switch (k) {
-            case 32:
-                solve_tri_f32_fast<64, 32>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 16:
-                solve_tri_f32_fast<64, 16>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 14:
-                solve_tri_f32_fast<64, 14>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 12:
-                solve_tri_f32_fast<64, 12>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 10:
-                solve_tri_f32_fast<64, 10>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 8:
-                solve_tri_f32_fast<64, 8>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 6:
-                solve_tri_f32_fast<64, 6>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 4:
-                solve_tri_f32_fast<64, 4>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 2:
-                solve_tri_f32_fast<64, 2>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            case 1:
-                solve_tri_f32_fast<64, 1>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
-                break;
-            default:
-                solve_tri_f32_fast<0, 0>
-                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
-        }
-    } else {  // run general case
-        solve_tri_f32_fast<0, 0>
-            <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
-    }
-}
-
-void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];  // A (triangular n x x matrix)
-    const ggml_tensor * src1 = dst->src[1];  // B (right hand side of n x k equation columns)
-
-    ggml_is_contiguous(src0);
-    ggml_is_contiguous(src1);
-
-    const int64_t n = src0->ne[0];
-    const int64_t k = src1->ne[0];
-
-    GGML_ASSERT(n <= 64);
-    GGML_ASSERT(k <= 32);
-
-    solve_tri_f32_cuda((const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k, src0->ne[2],
-                       src0->ne[3], src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
-                       src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
-                       dst->nb[3] / sizeof(float), ctx.stream());
-}

ggml/src/ggml-cuda/solve_tri.cuh

@@ -1,3 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/vendors/hip.h

@@ -105,7 +105,7 @@
 #define cudaStreamNonBlocking hipStreamNonBlocking
 #define cudaStreamPerThread hipStreamPerThread
 #define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent hipStreamWaitEvent
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
 #define cudaGraphExec_t hipGraphExec_t
 #define cudaGraphNode_t hipGraphNode_t
 #define cudaKernelNodeParams hipKernelNodeParams

ggml/src/ggml-hexagon/CMakeLists.txt

@@ -43,14 +43,6 @@ 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")
@@ -69,8 +61,6 @@ 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

@@ -9,7 +9,6 @@
 #include <chrono>
 #include <mutex>
 #include <string>
-#include <stdexcept>
 
 #ifdef _WIN32
 #    include <sal.h>
@@ -241,23 +240,6 @@ 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
@@ -1930,15 +1912,6 @@ 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];
@@ -1986,7 +1959,16 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
     }
 
     // src0 & src1 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, src1, dst)) {
+    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)) {
         return false;
     }
 
@@ -2034,7 +2016,20 @@ 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 (!hex_supported_buffer(sess, src0, src1, src2, dst)) {
+    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)) {
         return false;
     }
 
@@ -2068,7 +2063,16 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, src1, dst)) {
+    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)) {
         return false;
     }
 
@@ -2100,7 +2104,20 @@ static bool ggml_hexagon_supported_add_id(const struct ggml_hexagon_session * se
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, src1, src2, dst)) {
+    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)) {
         return false;
     }
 
@@ -2127,7 +2144,12 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
     }
 
     // src0 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, dst)) {
+    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)) {
         return false;
     }
 
@@ -2164,7 +2186,16 @@ static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, src1, dst)) {
+    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)) {
         return false;
     }
 
@@ -2217,7 +2248,16 @@ static bool ggml_hexagon_supported_softmax(const struct ggml_hexagon_session * s
     }
 
     // src0, src1 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, src1, dst)) {
+    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)) {
         return false;
     }
 
@@ -2229,7 +2269,7 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
 
     int mode = op_params[2];
 
-    if ((mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
+    if ((mode & GGML_ROPE_TYPE_NEOX) || (mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
         return false;
     }
     if (mode & 1) {
@@ -2272,7 +2312,20 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
     }
 
     // src0, src1, src2 & dst must be mapped to the same session
-    if (!hex_supported_buffer(sess, src0, src1, src2, dst)) {
+    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)) {
         return false;
     }
 
@@ -2293,26 +2346,6 @@ 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;
@@ -2327,6 +2360,10 @@ 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();
 
@@ -2348,27 +2385,55 @@ 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
-    dspqueue_buffers_init(bufs, src0, false, false);
+    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;
 
     // 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.
-    dspqueue_buffers_init(&bufs[1], src1, true, true);
+    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
 
     // 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.
-    dspqueue_buffers_init(&bufs[2], dst, true, false);
+    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);
 
-    auto * sess = get_session_from_tensor(src0);
+    // Primary DSP session from the src0 (normally weight) tensor
+    auto sess = src0_buf->sess;
 
     if (opt_verbose) {
-        hex_print_op_info(op, sess, 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);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2398,6 +2463,11 @@ 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();
 
@@ -2420,32 +2490,66 @@ 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
-    dspqueue_buffers_init(bufs, src0, false, false);
+    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;
 
     // 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.
-    dspqueue_buffers_init(&bufs[1], src1, true, true);
+    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
 
     // 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.
-    dspqueue_buffers_init(&bufs[2], src2, true, true);
+    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
 
     // 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.
-    dspqueue_buffers_init(&bufs[3], dst, true, false);
+    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);
 
-    auto * sess = get_session_from_tensor(src0);
+    // Primary DSP session from the src0 (normally weight) tensor
+    auto sess = src0_buf->sess;
 
     if (opt_verbose) {
-        hex_print_op_info(op, sess, 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_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);
+
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2477,6 +2581,10 @@ 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;
 
@@ -2513,30 +2621,60 @@ 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.
-    dspqueue_buffers_init(bufs, src0, true, true);
+    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;
 
     // 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.
-    dspqueue_buffers_init(&bufs[1], src1, true, true);
+    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
 
     // 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.
-    dspqueue_buffers_init(&bufs[2], dst, true, false);
+    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);
 
-    auto * sess = get_session_from_tensor(src0);
+    // Primary DSP session from the src0 tensor
+    ggml_hexagon_session * sess = src0_buf->sess;
 
     if (opt_verbose) {
-        hex_print_op_info(op, sess, req.flags);
+        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);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2567,6 +2705,11 @@ 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;
 
@@ -2598,19 +2741,58 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     init_htp_tensor(&req.dst, dst);
 
     dspqueue_buffer bufs[4];
-    // First buffer = input activations
-    dspqueue_buffers_init(bufs, src0, true, true);
-    // Second buffer = experts bias
-    dspqueue_buffers_init(&bufs[1], src1, true, true);
-    // Third buffer = activated experts
-    dspqueue_buffers_init(&bufs[2], src2, true, true);
-    // Forth buffer = output activations
-    dspqueue_buffers_init(&bufs[3], dst, true, true);
+    memset(bufs, 0, sizeof(bufs));
 
-    auto * sess = get_session_from_tensor(src0);
+    // 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;
+
+    // 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
+
+    // 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
+
+    // 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);
+
+    // Primary DSP session from the src0 tensor
+    ggml_hexagon_session * sess = src0_buf->sess;
 
     if (opt_verbose) {
-        hex_print_op_info(op, sess, req.flags);
+        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);
+
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             hex_dump_dspbuf(src1, &bufs[1]);
@@ -2704,33 +2886,71 @@ 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.
-    size_t n_bufs = dspqueue_buffers_init(bufs, src0, true, true);
+    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;
 
-    // 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);
+    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 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.
-    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], dst, true, false);
+    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;
 
     // Primary DSP session from the src0 tensor
-    auto * sess = get_session_from_tensor(src0);
+    ggml_hexagon_session * sess = src0_buf->sess;
 
     if (opt_verbose) {
-        hex_print_op_info(op, sess, 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);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             if (src1) {
@@ -2803,40 +3023,85 @@ 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.
-    size_t n_bufs = dspqueue_buffers_init(bufs, src0, true, true);
+    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;
 
     // 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.
-    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], src1, true, true);
+    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;
 
-    // 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);
+    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;
+    }
 
     // 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.
-    n_bufs += dspqueue_buffers_init(&bufs[n_bufs], dst, true, false);
+    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;
 
     // Primary DSP session from the src0 tensor
-    auto * sess = get_session_from_tensor(src0);
+    ggml_hexagon_session * sess = src0_buf->sess;
 
     if (opt_verbose) {
-        hex_print_op_info(op, sess, 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);
         if (opt_verbose > 1) {
             hex_dump_dspbuf(src0, &bufs[0]);
             if (src1) {

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

@@ -390,12 +390,6 @@ 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

@@ -106,32 +106,33 @@ 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;
 
-    const bool src1_valid = src1->ne[0];
-    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
+    bool src1_valid = src1->ne[0];
     if (!src1_valid) {
-        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;
+        data_src1     = data_src0;
+        src1_row_size = src0_row_size;
     }
 
     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 bool opt_path = ((1 == is_aligned) && !(nb01 & (VLEN - 1)));
+    const int32_t swapped = op_params[1];
+
+    const int nc = (src1_valid) ? ne0 : ne0 / 2;
+
     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));
@@ -141,7 +142,12 @@ 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 (opt_path) {
+        if (!src1_valid) {
+            src0 += swapped ? nc : 0;
+            src1 += swapped ? 0 : nc;
+        }
+
+        if (1 == 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);
@@ -212,7 +218,7 @@ 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) ? ne00 : ne00 / 2;
+    const int nc = (src1_valid) ? ne0 : ne0 / 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

@@ -66,13 +66,6 @@ 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

@@ -16,14 +16,6 @@
 #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;
@@ -42,12 +34,6 @@ 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;
@@ -56,9 +42,9 @@ 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_guard(neg_vec_in, max_exp, inf);
+                *p_vec_out++          = hvx_vec_exp_fp32(neg_vec_in);
             } else {
-                *p_vec_out++ = hvx_vec_exp_fp32_guard(*p_vec_in1++, max_exp, inf);
+                *p_vec_out++ = hvx_vec_exp_fp32(*p_vec_in1++);
             }
         }
     } else {
@@ -68,9 +54,9 @@ 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_guard(neg_vec_in, max_exp, inf);
+                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32(neg_vec_in);
             } else {
-                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32_guard(in, max_exp, inf);
+                *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_exp_fp32(in);
             }
         }
     }
@@ -84,9 +70,9 @@ 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_guard(neg_vec_in, max_exp, inf);
+            vec_out = hvx_vec_exp_fp32(neg_vec_in);
         } else {
-            vec_out = hvx_vec_exp_fp32_guard(in, max_exp, inf);
+            vec_out = hvx_vec_exp_fp32(in);
         }
 
         hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, vec_out);

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

@@ -16,15 +16,6 @@
 #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;
@@ -41,22 +32,19 @@ 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_guard(*p_vec_in++, nan_inf_mask);
+            *p_vec_out++ = hvx_vec_inverse_fp32(*p_vec_in++);
         }
     } 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_guard(in, nan_inf_mask);
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = hvx_vec_inverse_fp32(in);
         }
     }
 
@@ -65,7 +53,7 @@ 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_guard(in, nan_inf_mask);
+        HVX_Vector out = hvx_vec_inverse_fp32(in);
 
         hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, out);
     }

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

@@ -401,9 +401,7 @@ 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");
     }
 
-    static const float kInf    = INFINITY;
-    const HVX_Vector   inf     = hvx_vec_splat_fp32(kInf);
-    HVX_Vector         val_vec = hvx_vec_splat_fp32(val);
+    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
 
     if (0 == unaligned_loop) {
         HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
@@ -411,24 +409,17 @@ 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           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;
+            HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*vec_in1++, val_vec);
+            *vec_out++   = Q6_Vsf_equals_Vqf32(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);
 
-            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_Vector out = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
 
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = out;
+            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
         }
     }
 
@@ -438,12 +429,8 @@ void hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t *
 
         HVX_Vector in = *(HVX_UVector *) srcf;
 
-        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);
+        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));
     }
 }
 

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

@@ -12,35 +12,6 @@
 #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;
@@ -272,16 +243,19 @@ 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) {
-    HVX_VectorAlias u = { .v = v };
+    union {
+        HVX_Vector v;
+        __fp16 d[64];
+    } 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.fp16 + (16 * i), 16);
+        htp_dump_fp16_line(pref, u.d + (16 * i), 16);
     }
     if (n1) {
-        htp_dump_fp16_line(pref, u.fp16 + (16 * i), n1);
+        htp_dump_fp16_line(pref, u.d + (16 * i), n1);
     }
 }
 
@@ -437,8 +411,8 @@ 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;
@@ -517,7 +491,7 @@ 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_vxor_VV(v, mask);
+    return Q6_V_vor_VV(v, mask);
 }
 
 static inline HVX_Vector hvx_vec_abs_fp32(HVX_Vector v) {
@@ -532,7 +506,7 @@ 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_vxor_VV(v, mask);
+    return Q6_V_vor_VV(v, mask);
 #endif  // __HTP_ARCH__ > 75
 }
 
@@ -960,18 +934,6 @@ 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;
@@ -981,16 +943,9 @@ 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_guard(v_src[i], one, max_exp, min_exp);
+        v_dst[i] = hvx_vec_fast_sigmoid_fp32(v_src[i]);
     }
 }
 

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

@@ -143,25 +143,16 @@ 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);
-        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_acquire_cached(ctx->vtcm_rctx, 1000000);
         HAP_compute_res_release_cached(ctx->vtcm_rctx);
         qurt_thread_set_priority(qurt_thread_get_id(), ctx->thread_prio);
 
-        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_acquire_cached(ctx->vtcm_rctx, 1000000);
         ctx->vtcm_valid = true;
     }
 
@@ -210,7 +201,7 @@ 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, 0, vtcm_size);
+    HAP_compute_res_attr_set_vtcm_param_v2(&attr, vtcm_size, vtcm_size, 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

@@ -24,10 +24,6 @@
 #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]; \
@@ -150,57 +146,6 @@ 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,
@@ -267,9 +212,6 @@ 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;
@@ -305,35 +247,20 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
                 float *       dst_data_loc = dst_data;
 
                 if (1 == opt_path) {
-                    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);
-                    }
+                    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];
 
-                        if (is_neox) {
-                            const float x0 = src_loc[0];
-                            const float x1 = src_loc[rope_ctx->n_dims/2];
+                        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[rope_ctx->n_dims/2] = x0 * sin_theta + x1 * cos_theta;
+                        dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
+                        dst_data_loc[1] = x0 * sin_theta + x1 * cos_theta;
 
-                            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;
-                        }
+                        src_loc += 2;
+                        dst_data_loc += 2;
                     }
                 }
 

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

@@ -1009,64 +1009,6 @@ 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

@@ -128,8 +128,6 @@ 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

@@ -905,7 +905,6 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_LEAKY_RELU:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ARGSORT:
-        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

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

@@ -11,7 +11,6 @@
 #include <cassert>
 #include <algorithm>
 #include <limits>
-#include <cmath>
 
 static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
     if (!t) {
@@ -406,10 +405,6 @@ 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);
@@ -3682,19 +3677,14 @@ 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,
-        /*.ne0   =*/ ne0,
-        /*.ne1   =*/ ne1,
-        /*.ne2   =*/ ne2,
-        /*.ne3   =*/ ne3,
-        /*.top_k =*/ nth,
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
     };
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
@@ -3714,20 +3704,15 @@ 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,
-            /*.ne0   =*/ ne0,
-            /*.ne1   =*/ ne1,
-            /*.ne2   =*/ ne2,
-            /*.ne3   =*/ ne3,
-            /*.top_k =*/ ne00,
-            /*.len   =*/ len,
+            .ne00 = ne00,
+            .ne01 = ne01,
+            .ne02 = ne02,
+            .ne03 = ne03,
+            .nb00 = nb00,
+            .nb01 = nb01,
+            .nb02 = nb02,
+            .nb03 = nb03,
+            .len  = len,
         };
 
         // merges per row
@@ -3751,118 +3736,6 @@ 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

@@ -81,7 +81,6 @@ 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

@@ -202,10 +202,6 @@ 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

@@ -4670,12 +4670,11 @@ 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 = ib*ntg.x;
-    const int i01 = tgpig[0] % args.ne01;
-    const int i02 = tgpig[1];
-    const int i03 = tgpig[2];
+    const int i00 = (tgpig[0]/args.ne01)*ntg.x;
+    const int i01 =  tgpig[0]%args.ne01;
+    const int i02 =  tgpig[1];
+    const int i03 =  tgpig[2];
 
     device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
@@ -4711,11 +4710,9 @@ kernel void kernel_argsort_f32_i32(
         }
     }
 
-    const int64_t i0 = ib*args.top_k;
-
     // copy the result to dst without the padding
-    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;
+    if (i00 + col < args.ne00) {
+        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
 
         dst[col] = shmem_i32[col];
     }
@@ -4750,22 +4747,22 @@ kernel void kernel_argsort_merge_f32_i32(
 
     const int start = im * (2 * 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 len0 = MIN(args.len, MAX(0, args.ne00 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne00 - (int)(start + args.len)));
 
     const int total = len0 + len1;
 
     device const int32_t * tmp0 = tmp + start
-        + i01*args.ne0
-        + i02*args.ne0*args.ne01
-        + i03*args.ne0*args.ne01*args.ne02;
+        + i01*args.ne00
+        + i02*args.ne00*args.ne01
+        + i03*args.ne00*args.ne01*args.ne02;
 
     device const int32_t * tmp1 = tmp0 + args.len;
 
     dst += start
-        + i01*args.top_k
-        + i02*args.top_k*args.ne01
-        + i03*args.top_k*args.ne01*args.ne02;
+        + i01*args.ne00
+        + i02*args.ne00*args.ne01
+        + i03*args.ne00*args.ne01*args.ne02;
 
     device const float * src0_row = (device const float *)(src0
         + args.nb01*i01
@@ -4779,11 +4776,7 @@ 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(MIN(k0 + chunk, total), args.top_k);
-
-    if (k0 >= args.top_k) {
-        return;
-    }
+    const int k1 = min(k0 + chunk, total);
 
     if (k0 >= total) {
         return;

ggml/src/ggml-opencl/CMakeLists.txt

@@ -70,7 +70,6 @@ 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
@@ -110,9 +109,6 @@ 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

@@ -449,9 +449,6 @@ 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;
@@ -512,7 +509,6 @@ 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;
@@ -1556,66 +1552,6 @@ 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
@@ -1889,24 +1825,6 @@ 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
@@ -3041,10 +2959,6 @@ 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:
@@ -3093,8 +3007,6 @@ 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:
@@ -3163,7 +3075,6 @@ 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:
             {
@@ -5282,224 +5193,6 @@ 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);
@@ -7202,23 +6895,9 @@ 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) {
-            // For KQ
-            if (ggml_is_permuted(src0) && ggml_is_permuted(src1) &&
-                nb00 <= nb02 &&
-                nb02 <= nb01 &&
-                nb01 <= nb03 &&
-                nb10 <= nb12 &&
-                nb12 <= nb11 &&
-                nb11 <= nb13) {
-                ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
-                return;
-            }
-            // For KQV
-            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-                ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
-                return;
-            }
+        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0){
+            ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
+            return;
         }
     }
 
@@ -9398,24 +9077,6 @@ 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:
@@ -9517,12 +9178,6 @@ 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

@@ -1,39 +0,0 @@
-
-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

@@ -1,53 +0,0 @@
-#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];
-}