ggml : remove redundant src in ggml_cast

* llama : add token support to llama-grammar

* fix inverse token comment

* refactor trigger_patterns to replay tokens instead of the entire string

* add token documentation

* fix test-llama-grammar

* improve test cases for tokens

.github/actions/windows-setup-cuda/action.yml

@@ -65,3 +65,34 @@ runs:
           echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
           echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+    - name: Install Cuda Toolkit 13.1
+      if: ${{ inputs.cuda_version == '13.1' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1"
+          choco install unzip -y
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.1.80-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-13.1.80-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.1.80-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-13.1.80-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-13.2.0.9-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.1.80-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-13.1.68-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-13.1.80-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-13.1.68-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-13.1.78-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_crt-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_cudart-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_nvcc-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_nvrtc-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\libcublas-windows-x86_64-13.2.0.9-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\libnvvm-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_nvtx-windows-x86_64-13.1.68-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_profiler_api-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\visual_studio_integration-windows-x86_64-13.1.68-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_cccl-windows-x86_64-13.1.78-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_1=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

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

@@ -291,6 +291,7 @@ jobs:
                          -DGGML_RVV=ON \
                          -DGGML_RV_ZFH=ON \
                          -DGGML_RV_ZICBOP=ON \
+                         -DGGML_RV_ZIHINTPAUSE=ON \
                          -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
                          -DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake
 

.github/workflows/build-riscv-native.yml

@@ -1,120 +0,0 @@
-name: Build on RISCV Linux Machine by Cloud-V
-on:
-  pull_request:
-  workflow_dispatch:
-  workflow_call:
-
-jobs:
-  debian-13-riscv64-native: # Bianbu 2.2
-    runs-on: [self-hosted, RISCV64]
-
-    steps:
-      - name: Install prerequisites
-        run: |
-          sudo apt-get update || true
-          sudo apt-get install -y libatomic1
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo apt-get update || true
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu \
-                  ccache \
-                  cmake
-
-      - name: Setup ccache
-        run: |
-          mkdir -p $HOME/.ccache
-          ccache -M 5G -d $HOME/.ccache
-          export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
-          export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
-          echo "$GITHUB_WORKSPACE"
-          echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
-          echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
-          echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
-          echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
-
-      - name: Build
-        run: |
-          cmake -B build \
-            -DLLAMA_CURL=OFF \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_OPENMP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=ON \
-            -DLLAMA_BUILD_TOOLS=ON \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DCMAKE_SYSTEM_NAME=Linux \
-            -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
-            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-            -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-            -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-            -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-            -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
-
-          cmake --build build --config Release -j $(nproc)
-
-  # debian-13-riscv64-spacemit-ime-native: # Bianbu 2.2
-  #   runs-on: [self-hosted, RISCV64]
-
-  #   steps:
-  #     - name: Install prerequisites
-  #       run: |
-  #         sudo apt-get update || true
-  #         sudo apt-get install -y libatomic1
-  #     - uses: actions/checkout@v4
-  #     - name: Setup Riscv
-  #       run: |
-  #         sudo apt-get update || true
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 gcc-14-riscv64-linux-gnu \
-  #                 g++-14-riscv64-linux-gnu \
-  #                 ccache \
-  #                 cmake
-  #         sudo apt-get upgrade binutils -y
-
-  #     - name: Setup ccache
-  #       run: |
-  #         mkdir -p $HOME/.ccache
-  #         ccache -M 5G -d $HOME/.ccache
-  #         export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
-  #         export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
-  #         echo "$GITHUB_WORKSPACE"
-  #         echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
-  #         echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
-  #         echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
-  #         echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
-
-  #     - name: Build
-  #       run: |
-  #         cmake -B build \
-  #           -DLLAMA_CURL=OFF \
-  #           -DCMAKE_BUILD_TYPE=Release \
-  #           -DGGML_OPENMP=OFF \
-  #           -DLLAMA_BUILD_EXAMPLES=ON \
-  #           -DLLAMA_BUILD_TOOLS=ON \
-  #           -DLLAMA_BUILD_TESTS=OFF \
-  #           -DCMAKE_SYSTEM_NAME=Linux \
-  #           -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-  #           -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-  #           -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-  #           -DCMAKE_C_COMPILER_LAUNCHER=ccache \
-  #           -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-  #           -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-  #           -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-  #           -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-  #           -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-  #           -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH \
-  #           -DGGML_RVV=ON \
-  #           -DGGML_RV_ZFH=ON \
-  #           -DGGML_RV_ZICBOP=ON \
-  #           -DGGML_CPU_RISCV64_SPACEMIT=ON \
-  #           -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1
-
-  #         cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -547,6 +547,46 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
+  ubuntu-24-wasm-webgpu:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-latest-wasm-webgpu
+          evict-old-files: 1d
+
+      - name: Install Emscripten
+        run: |
+          git clone https://github.com/emscripten-core/emsdk.git
+          cd emsdk
+          ./emsdk install latest
+          ./emsdk activate latest
+
+      - name: Fetch emdawnwebgpu
+        run: |
+          DAWN_TAG="v20251027.212519"
+          EMDAWN_PKG="emdawnwebgpu_pkg-${DAWN_TAG}.zip"
+          echo "Downloading ${EMDAWN_PKG}"
+          curl -L -o emdawn.zip \
+            "https://github.com/google/dawn/releases/download/${DAWN_TAG}/${EMDAWN_PKG}"
+          unzip emdawn.zip
+
+      - name: Build WASM WebGPU
+        run: |
+          source emsdk/emsdk_env.sh
+          emcmake cmake -B build-wasm \
+            -DGGML_WEBGPU=ON \
+            -DLLAMA_CURL=OFF \
+            -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
+
+          cmake --build build-wasm --target test-backend-ops -j $(nproc)
+
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.1.2
@@ -1562,33 +1602,33 @@ jobs:
         run: |
           bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
-  ggml-ci-x64-amd-vulkan:
-    runs-on: [self-hosted, Linux, X64, AMD]
+  # ggml-ci-x64-amd-vulkan:
+  #   runs-on: [self-hosted, Linux, X64, AMD]
 
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v4
 
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+  #     - name: Test
+  #       id: ggml-ci
+  #       run: |
+  #         vulkaninfo --summary
+  #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
-  ggml-ci-x64-amd-rocm:
-    runs-on: [self-hosted, Linux, X64, AMD]
+  # ggml-ci-x64-amd-rocm:
+  #   runs-on: [self-hosted, Linux, X64, AMD]
 
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v4
 
-      - name: Test
-        id: ggml-ci
-        run: |
-          amd-smi static
-          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+  #     - name: Test
+  #       id: ggml-ci
+  #       run: |
+  #         amd-smi static
+  #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
   ggml-ci-mac-metal:
     runs-on: [self-hosted, macOS, ARM64]
@@ -1642,6 +1682,337 @@ jobs:
          run: |
            GG_BUILD_KLEIDIAI=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
+  ubuntu-cpu-cmake-riscv64-native:
+    runs-on: RISCV64
+
+    steps:
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+
+          # Install necessary packages
+          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 rustup cmake build-essential libssl-dev wget ccache
+
+          # Set gcc-14 and g++-14 as the default compilers
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
+          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
+          sudo ln -sf /usr/bin/gcc-14 /usr/bin/gcc
+          sudo ln -sf /usr/bin/g++-14 /usr/bin/g++
+
+          # Install Rust stable version
+          rustup install stable
+          rustup default stable
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Check environment
+        run: |
+          uname -a
+          gcc --version
+          g++ --version
+          ldd --version
+          cmake --version
+          rustc --version
+
+      - name: Setup ccache
+        run: |
+          # Set unique cache directory for this job
+          export CCACHE_DIR="$HOME/.ccache/cpu-cmake-rv64-native"
+          mkdir -p "$CCACHE_DIR"
+
+          # Configure ccache for optimal performance
+          ccache --set-config=max_size=5G
+          ccache --set-config=compression=true
+          ccache --set-config=compression_level=6
+          ccache --set-config=cache_dir="$CCACHE_DIR"
+
+          # Enable more aggressive caching
+          ccache --set-config=sloppiness=file_macro,time_macros,include_file_mtime,include_file_ctime
+          ccache --set-config=hash_dir=false
+
+          # Export for subsequent steps
+          echo "CCACHE_DIR=$CCACHE_DIR" >> $GITHUB_ENV
+          echo "PATH=/usr/lib/ccache:$PATH" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=ON \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DGGML_RPC=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L 'main|curl' --verbose --timeout 900
+
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
+  ubuntu-cmake-sanitizer-riscv64-native:
+    runs-on: RISCV64
+
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug]
+
+    steps:
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+
+          # Install necessary packages
+          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 rustup cmake build-essential wget ccache
+
+          # Set gcc-14 and g++-14 as the default compilers
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
+          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
+          sudo ln -sf /usr/bin/gcc-14 /usr/bin/gcc
+          sudo ln -sf /usr/bin/g++-14 /usr/bin/g++
+
+          # Install Rust stable version
+          rustup install stable
+          rustup default stable
+
+      - name: GCC version check
+        run: |
+          gcc --version
+          g++ --version
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Setup ccache
+        run: |
+          # Unique cache directory per matrix combination
+          export CCACHE_DIR="$HOME/.ccache/sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}"
+          mkdir -p "$CCACHE_DIR"
+
+          # Configure ccache
+          ccache --set-config=max_size=5G
+          ccache --set-config=compression=true
+          ccache --set-config=compression_level=6
+          ccache --set-config=cache_dir="$CCACHE_DIR"
+          ccache --set-config=sloppiness=file_macro,time_macros,include_file_mtime,include_file_ctime
+          ccache --set-config=hash_dir=false
+
+          # Export for subsequent steps
+          echo "CCACHE_DIR=$CCACHE_DIR" >> $GITHUB_ENV
+          echo "PATH=/usr/lib/ccache:$PATH" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        if: ${{ matrix.sanitizer != 'THREAD' }}
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+            -DGGML_OPENMP=ON \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+
+  ubuntu-llguidance-riscv64-native:
+    runs-on: RISCV64
+    steps:
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+
+          # Install necessary packages
+          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 rustup cmake build-essential wget ccache
+
+          # Set gcc-14 and g++-14 as the default compilers
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
+          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
+          sudo ln -sf /usr/bin/gcc-14 /usr/bin/gcc
+          sudo ln -sf /usr/bin/g++-14 /usr/bin/g++
+
+          # Install Rust stable version
+          rustup install stable
+          rustup default stable
+
+      - name: GCC version check
+        run: |
+          gcc --version
+          g++ --version
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Setup ccache
+        run: |
+          export CCACHE_DIR="$HOME/.ccache/llguidance-riscv64"
+          mkdir -p "$CCACHE_DIR"
+
+          ccache --set-config=max_size=5G
+          ccache --set-config=compression=true
+          ccache --set-config=compression_level=6
+          ccache --set-config=cache_dir="$CCACHE_DIR"
+          ccache --set-config=sloppiness=file_macro,time_macros,include_file_mtime,include_file_ctime
+          ccache --set-config=hash_dir=false
+
+          echo "CCACHE_DIR=$CCACHE_DIR" >> $GITHUB_ENV
+          echo "PATH=/usr/lib/ccache:$PATH" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DLLAMA_LLGUIDANCE=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+
+  ubuntu-cmake-rpc-riscv64-native:
+    runs-on: RISCV64
+
+    continue-on-error: true
+
+    steps:
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+
+          # Install necessary packages
+          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 rustup cmake build-essential libssl-dev wget ccache
+
+          # Set gcc-14 and g++-14 as the default compilers
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
+          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
+          sudo ln -sf /usr/bin/gcc-14 /usr/bin/gcc
+          sudo ln -sf /usr/bin/g++-14 /usr/bin/g++
+
+          # Install Rust stable version
+          rustup install stable
+          rustup default stable
+
+      - name: GCC version check
+        run: |
+          gcc --version
+          g++ --version
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Setup ccache
+        run: |
+          export CCACHE_DIR="$HOME/.ccache/rpc-riscv64"
+          mkdir -p "$CCACHE_DIR"
+
+          ccache --set-config=max_size=5G
+          ccache --set-config=compression=true
+          ccache --set-config=compression_level=6
+          ccache --set-config=cache_dir="$CCACHE_DIR"
+          ccache --set-config=sloppiness=file_macro,time_macros,include_file_mtime,include_file_ctime
+          ccache --set-config=hash_dir=false
+
+          echo "CCACHE_DIR=$CCACHE_DIR" >> $GITHUB_ENV
+          echo "PATH=/usr/lib/ccache:$PATH" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=ON \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+            -DGGML_RPC=ON
+
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose
+
   ggml-ci-arm64-graviton4-kleidiai:
      runs-on: ah-ubuntu_22_04-c8g_8x
 

.github/workflows/release.yml

@@ -66,14 +66,21 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+          zip -y -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
-      - name: Upload artifacts
+      - name: Upload artifacts (zip)
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
           name: llama-bin-macos-arm64.zip
 
+      - name: Upload artifacts (tar)
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz
+          name: llama-bin-macos-arm64.tar.gz
+
   macOS-x64:
     runs-on: macos-15-intel
 
@@ -120,14 +127,21 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+          zip -y -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
-      - name: Upload artifacts
+      - name: Upload artifacts (zip)
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
           name: llama-bin-macos-x64.zip
 
+      - name: Upload artifacts (tar)
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz
+          name: llama-bin-macos-x64.tar.gz
+
   ubuntu-22-cpu:
     strategy:
       matrix:
@@ -182,14 +196,21 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
+          zip -y -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
-      - name: Upload artifacts
+      - name: Upload artifacts (zip)
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
           name: llama-bin-ubuntu-${{ matrix.build }}.zip
 
+      - name: Upload artifacts (tar)
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
+          name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
+
   ubuntu-22-vulkan:
     runs-on: ubuntu-22.04
 
@@ -235,14 +256,21 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
+          zip -y -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
-      - name: Upload artifacts
+      - name: Upload artifacts (zip)
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
           name: llama-bin-ubuntu-vulkan-x64.zip
 
+      - name: Upload artifacts (tar)
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz
+          name: llama-bin-ubuntu-vulkan-x64.tar.gz
+
   windows-cpu:
     runs-on: windows-2025
 
@@ -298,7 +326,7 @@ jobs:
         run: |
           Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
           Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
-          7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
+          7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
@@ -380,7 +408,7 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
+          7z a -snl llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
@@ -393,7 +421,7 @@ jobs:
 
     strategy:
       matrix:
-        cuda: ['12.4']
+        cuda: ['12.4', '13.1']
 
     steps:
       - name: Clone
@@ -434,7 +462,7 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
+          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
@@ -448,6 +476,7 @@ jobs:
           $dst='.\build\bin\cudart\'
           robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          robocopy "${{env.CUDA_PATH}}\bin\x64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
@@ -517,6 +546,8 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-fallback-bfloat16.spv" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-native-bfloat16.spv" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
@@ -526,7 +557,7 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/umf/latest/bin/umf.dll" ./build/bin
 
           echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a llama-bin-win-sycl-x64.zip ./build/bin/*
+          7z a -snl llama-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
         uses: actions/upload-artifact@v4
@@ -632,7 +663,7 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
+          7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
@@ -685,58 +716,20 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+          zip -y -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+          tar -czvf llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz -C build-apple llama.xcframework
 
-      - name: Upload artifacts
+      - name: Upload artifacts (zip)
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
-          name: llama-${{ steps.tag.outputs.name }}-xcframework
+          name: llama-${{ steps.tag.outputs.name }}-xcframework.zip
 
-  openEuler-cann:
-    strategy:
-      matrix:
-        arch: [x86, aarch64]
-        chip_type: ['910b', '310p']
-        build: ['Release']
-    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.2.rc1-310p-openeuler22.03-py3.11' }}
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Dependencies
-        run: |
-          yum update -y
-          yum install -y git gcc gcc-c++ make cmake libcurl-devel
-          git config --global --add safe.directory "$GITHUB_WORKSPACE"
-
-      - name: Build
-        run: |
-          export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
-
-          cmake -S . -B build \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build }} \
-              -DGGML_CANN=on \
-              -DSOC_TYPE=ascend${{ matrix.chip_type }}
-          cmake --build build -j $(nproc)
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.zip ./build/bin/*
-
-      - name: Upload artifacts
+      - name: Upload artifacts (tar)
         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-${{ matrix.arch }}.zip
+          path: llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz
+          name: llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -759,7 +752,6 @@ jobs:
       - macOS-arm64
       - macOS-x64
       - ios-xcode-build
-      - openEuler-cann
 
     steps:
       - name: Clone
@@ -814,6 +806,7 @@ jobs:
 
           echo "Moving other artifacts..."
           mv -v artifact/*.zip release
+          mv -v artifact/*.tar.gz release
 
       - name: Create release
         id: create_release
@@ -822,6 +815,34 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
           tag_name: ${{ steps.tag.outputs.name }}
+          body: |
+            > [!WARNING]
+            > **Release Format Update**: Linux releases will soon use .tar.gz archives instead of .zip. Please make the necessary changes to your deployment scripts.
+
+            <details open>
+
+            ${{ github.event.head_commit.message }}
+
+            </details>
+
+            **macOS/iOS:**
+            - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
+            - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz)
+            - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz)
+
+            **Linux:**
+            - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.tar.gz)
+            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
+            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
+
+            **Windows:**
+            - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)
+            - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-arm64.zip)
+            - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip)
+            - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.1-x64.zip)
+            - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
+            - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
+            - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
 
       - name: Upload release
         id: upload_release
@@ -833,7 +854,7 @@ jobs:
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
             for (let file of await fs.readdirSync('./release')) {
-              if (path.extname(file) === '.zip') {
+              if (path.extname(file) === '.zip' || file.endsWith('.tar.gz')) {
                 console.log('uploadReleaseAsset', file);
                 await github.repos.uploadReleaseAsset({
                   owner: context.repo.owner,

.github/workflows/winget.yml

@@ -9,6 +9,7 @@ jobs:
   update:
     name: Update Winget Package
     runs-on: ubuntu-latest
+    if: github.repository_owner == 'ggml-org'
 
     steps:
       - name: Install cargo binstall

.gitignore

@@ -134,3 +134,5 @@ poetry.toml
 # IDE
 /*.code-workspace
 /.windsurf/
+# emscripten
+a.out.*

CMakeLists.txt

@@ -33,10 +33,24 @@ endif()
 
 option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF)
 
+option(LLAMA_WASM_MEM64 "llama: use 64-bit memory in WASM builds" ON)
+
 if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
-    option(LLAMA_WASM_SINGLE_FILE "llama: embed WASM inside the generated llama.js" ON)
+    # Use 64-bit memory to support backend_get_memory queries
+    # TODO: analyze performance impact, see https://spidermonkey.dev/blog/2025/01/15/is-memory64-actually-worth-using
+    if (LLAMA_WASM_MEM64)
+      add_compile_options("-sMEMORY64=1")
+      add_link_options("-sMEMORY64=1")
+    endif()
+    add_link_options("-sALLOW_MEMORY_GROWTH=1")
+
+    option(LLAMA_WASM_SINGLE_FILE "llama: embed WASM inside the generated llama.js" OFF)
+    option(LLAMA_BUILD_HTML "llama: build HTML file" ON)
+    if (LLAMA_BUILD_HTML)
+        set(CMAKE_EXECUTABLE_SUFFIX ".html")
+    endif()
 else()
     if (MINGW)
         set(BUILD_SHARED_LIBS_DEFAULT OFF)
@@ -58,6 +72,12 @@ if (MSVC)
     add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
 endif()
 
+if (LLAMA_STANDALONE)
+    # enable parallel builds for msbuild
+    list(APPEND CMAKE_VS_GLOBALS UseMultiToolTask=true)
+    list(APPEND CMAKE_VS_GLOBALS EnforceProcessCountAcrossBuilds=true)
+endif()
+
 if (CMAKE_SYSTEM_NAME STREQUAL "iOS")
     set(LLAMA_TOOLS_INSTALL_DEFAULT OFF)
 else()
@@ -179,11 +199,6 @@ if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
-if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
-endif()
-
 #
 # build the library
 #

CODEOWNERS

@@ -7,16 +7,20 @@
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
 /common/CMakeLists.txt                  @ggerganov
-/common/arg.*                           @ggerganov @ericcurtin
+/common/arg.*                           @ggerganov
 /common/base64.hpp.*                    @ggerganov
 /common/build-info.*                    @ggerganov
+/common/chat.*                          @pwilkin
+/common/chat-peg-parser.*               @aldehir
 /common/common.*                        @ggerganov
 /common/console.*                       @ggerganov
 /common/http.*                          @angt
 /common/llguidance.*                    @ggerganov
 /common/log.*                           @ggerganov
+/common/peg-parser.*                    @aldehir
 /common/sampling.*                      @ggerganov
 /common/speculative.*                   @ggerganov
+/common/unicode.*                       @aldehir
 /convert_*.py                           @CISC
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
@@ -81,14 +85,14 @@
 /src/llama-vocab.*                      @CISC
 /src/models/                            @CISC
 /tests/                                 @ggerganov
+/tests/test-chat-.*                     @pwilkin
 /tools/batched-bench/                   @ggerganov
 /tools/main/                            @ggerganov
 /tools/mtmd/                            @ngxson
 /tools/perplexity/                      @ggerganov
 /tools/quantize/                        @ggerganov
 /tools/rpc/                             @rgerganov
-/tools/run/                             @ericcurtin
-/tools/server/*                         @ngxson @ggerganov @ericcurtin # no subdir
+/tools/server/*                         @ngxson @ggerganov # no subdir
 /tools/server/webui/                    @allozaur
 /tools/tokenize/                        @ggerganov
 /tools/tts/                             @ggerganov

CONTRIBUTING.md

@@ -16,9 +16,10 @@ The project differentiates between 3 levels of contributors:
     - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
 - Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
-- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
+- If your PR becomes stale, rebase it on top of latest `master` to get maintainers attention
 - Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR
 - Consider adding yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs
+- Using AI to generate PRs is permitted. However, you must (1) explicitly disclose how AI was used and (2) conduct a thorough manual review before publishing the PR. Note that trivial tab autocompletions do not require disclosure.
 
 # Pull requests (for maintainers)
 

README.md

@@ -61,7 +61,7 @@ range of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2, AVX512 and AMX support for x86 architectures
-- RVV, ZVFH, ZFH and ZICBOP support for RISC-V architectures
+- RVV, ZVFH, ZFH, ZICBOP and ZIHINTPAUSE support for RISC-V architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
 - Vulkan and SYCL backend support
@@ -276,6 +276,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
+| [ZenDNN](docs/build.md#zendnn) | AMD CPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
@@ -613,3 +614,4 @@ $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 - [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
 - [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
 - [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
+- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain

SECURITY.md

@@ -65,4 +65,6 @@ However, If you have discovered a security vulnerability in this project, please
 
 Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new).
 
+Please note that using AI to identify vulnerabilities and generate reports is permitted. However, you must (1) explicitly disclose how AI was used and (2) conduct a thorough manual review before submitting the report.
+
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.

ci/run.sh

@@ -45,7 +45,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=${LLAMA_FATAL_WARNINGS:-ON} -DLLAMA_CURL=ON -DGGML_SCHED_NO_REALLOC=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 ) 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
+    (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
 
     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 ) 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
+    (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
 
     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 --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 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

cmake/build-info.cmake

@@ -39,26 +39,10 @@ if(Git_FOUND)
     endif()
 endif()
 
-if(MSVC)
-    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    if (CMAKE_VS_PLATFORM_NAME)
-        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
-    else()
-        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
-    endif()
-else()
-    execute_process(
-        COMMAND ${CMAKE_C_COMPILER} --version
-        OUTPUT_VARIABLE OUT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
-    set(BUILD_COMPILER ${OUT})
+set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
 
-    execute_process(
-        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
-        OUTPUT_VARIABLE OUT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    set(BUILD_TARGET ${OUT})
+if(CMAKE_VS_PLATFORM_NAME)
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+else()
+    set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
 endif()

common/CMakeLists.txt

@@ -52,6 +52,8 @@ add_library(${TARGET} STATIC
     chat-parser.h
     chat-parser-xml-toolcall.h
     chat-parser-xml-toolcall.cpp
+    chat-peg-parser.cpp
+    chat-peg-parser.h
     chat.cpp
     chat.h
     common.cpp
@@ -69,12 +71,16 @@ add_library(${TARGET} STATIC
     log.h
     ngram-cache.cpp
     ngram-cache.h
+    peg-parser.cpp
+    peg-parser.h
     regex-partial.cpp
     regex-partial.h
     sampling.cpp
     sampling.h
     speculative.cpp
     speculative.h
+    unicode.cpp
+    unicode.h
     )
 
 if (BUILD_SHARED_LIBS)

common/arg.cpp

@@ -30,6 +30,7 @@
 #include <thread> // for hardware_concurrency
 #include <vector>
 
+#ifndef __EMSCRIPTEN__
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
@@ -41,6 +42,8 @@
 #else
 #include <sys/syslimits.h>
 #endif
+#endif
+
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
 using json = nlohmann::ordered_json;
@@ -212,13 +215,13 @@ struct handle_model_result {
 static handle_model_result common_params_handle_model(
         struct common_params_model & model,
         const std::string & bearer_token,
-        const std::string & model_path_default,
         bool offline) {
     handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
         if (!model.docker_repo.empty()) {  // Handle Docker URLs by resolving them to local paths
             model.path = common_docker_resolve_model(model.docker_repo);
+            model.name = model.docker_repo; // set name for consistency
         } else if (!model.hf_repo.empty()) {
             // short-hand to avoid specifying --hf-file -> default it to --model
             if (model.hf_file.empty()) {
@@ -227,7 +230,8 @@ static handle_model_result common_params_handle_model(
                     if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
                         exit(1); // built without CURL, error message already printed
                     }
-                    model.hf_repo = auto_detected.repo;
+                    model.name    = model.hf_repo;      // repo name with tag
+                    model.hf_repo = auto_detected.repo; // repo name without tag
                     model.hf_file = auto_detected.ggufFile;
                     if (!auto_detected.mmprojFile.empty()) {
                         result.found_mmproj   = true;
@@ -257,8 +261,6 @@ static handle_model_result common_params_handle_model(
                 model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
             }
 
-        } else if (model.path.empty()) {
-            model.path = model_path_default;
         }
     }
 
@@ -405,7 +407,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
 
     // handle model and download
     {
-        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH, params.offline);
+        auto res = common_params_handle_model(params.model, params.hf_token, params.offline);
         if (params.no_mmproj) {
             params.mmproj = {};
         } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
@@ -415,12 +417,18 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         // only download mmproj if the current example is using it
         for (auto & ex : mmproj_examples) {
             if (ctx_arg.ex == ex) {
-                common_params_handle_model(params.mmproj,    params.hf_token, "", params.offline);
+                common_params_handle_model(params.mmproj,    params.hf_token, params.offline);
                 break;
             }
         }
-        common_params_handle_model(params.speculative.model, params.hf_token, "", params.offline);
-        common_params_handle_model(params.vocoder.model,     params.hf_token, "", params.offline);
+        common_params_handle_model(params.speculative.model, params.hf_token, params.offline);
+        common_params_handle_model(params.vocoder.model,     params.hf_token, params.offline);
+    }
+
+    // model is required (except for server)
+    // TODO @ngxson : maybe show a list of available models in CLI in this case
+    if (params.model.path.empty() && ctx_arg.ex != LLAMA_EXAMPLE_SERVER && !params.usage) {
+        throw std::invalid_argument("error: --model is required\n");
     }
 
     if (params.escape) {
@@ -700,6 +708,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         params.use_jinja = true;
     }
 
+    params.use_color = tty_can_use_colors();
+
     // load dynamic backends
     ggml_backend_load_all();
 
@@ -782,10 +792,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
-        {"-co", "--color"},
-        string_format("colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false"),
-        [](common_params & params) {
-            params.use_color = true;
+        {"-co", "--color"}, "[on|off|auto]",
+        "Colorize output to distinguish prompt and user input from generations ('on', 'off', or 'auto', default: 'auto')\n"
+        "'auto' enables colors when output is to a terminal",
+        [](common_params & params, const std::string & value) {
+            if (is_truthy(value)) {
+                params.use_color = true;
+            } else if (is_falsey(value)) {
+                params.use_color = false;
+            } else if (is_autoy(value)) {
+                params.use_color = tty_can_use_colors();
+            } else {
+                throw std::invalid_argument(
+                    string_format("error: unknown value for --color: '%s'\n", value.c_str()));
+            }
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
     add_opt(common_arg(
@@ -980,7 +1000,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.kv_unified = true;
         }
-    ).set_env("LLAMA_ARG_KV_SPLIT"));
+    ).set_env("LLAMA_ARG_KV_UNIFIED"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1014,7 +1034,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                                params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
                            } else {
                                throw std::runtime_error(
-                                   string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
+                                   string_format("error: unknown value for --flash-attn: '%s'\n", value.c_str()));
                            }
                        }).set_env("LLAMA_ARG_FLASH_ATTN"));
     add_opt(common_arg(
@@ -1221,7 +1241,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -2090,11 +2110,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"-m", "--model"}, "FNAME",
         ex == LLAMA_EXAMPLE_EXPORT_LORA
-            ? std::string("model path from which to load base model")
-            : string_format(
-                "model path (default: `models/$filename` with filename from `--hf-file` "
-                "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH
-            ),
+            ? "model path from which to load base model"
+            : "model path to load",
         [](common_params & params, const std::string & value) {
             params.model.path = value;
         }
@@ -2486,12 +2503,50 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "path to save slot kv cache (default: disabled)",
         [](common_params & params, const std::string & value) {
             params.slot_save_path = value;
+            if (!fs_is_directory(params.slot_save_path)) {
+                throw std::invalid_argument("not a directory: " + value);
+            }
             // if doesn't end with DIRECTORY_SEPARATOR, add it
             if (!params.slot_save_path.empty() && params.slot_save_path[params.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) {
                 params.slot_save_path += DIRECTORY_SEPARATOR;
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--media-path"}, "PATH",
+        "directory for loading local media files; files can be accessed via file:// URLs using relative paths (default: disabled)",
+        [](common_params & params, const std::string & value) {
+            params.media_path = value;
+            if (!fs_is_directory(params.media_path)) {
+                throw std::invalid_argument("not a directory: " + value);
+            }
+            // if doesn't end with DIRECTORY_SEPARATOR, add it
+            if (!params.media_path.empty() && params.media_path[params.media_path.size() - 1] != DIRECTORY_SEPARATOR) {
+                params.media_path += DIRECTORY_SEPARATOR;
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--models-dir"}, "PATH",
+        "directory containing models for the router server (default: disabled)",
+        [](common_params & params, const std::string & value) {
+            params.models_dir = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODELS_DIR"));
+    add_opt(common_arg(
+        {"--models-max"}, "N",
+        string_format("for router server, maximum number of models to load simultaneously (default: %d, 0 = unlimited)", params.models_max),
+        [](common_params & params, int value) {
+            params.models_max = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODELS_MAX"));
+    add_opt(common_arg(
+        {"--no-models-autoload"},
+        "disables automatic loading of models (default: enabled)",
+        [](common_params & params) {
+            params.models_autoload = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_MODELS_AUTOLOAD"));
     add_opt(common_arg(
         {"--jinja"},
         string_format("use jinja template for chat (default: %s)\n", params.use_jinja ? "enabled" : "disabled"),
@@ -2639,7 +2694,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params &, const std::string & value) {
             common_log_set_file(common_log_main(), value.c_str());
         }
-    ));
+    ).set_env("LLAMA_LOG_FILE"));
     add_opt(common_arg(
         {"--log-colors"}, "[on|off|auto]",
         "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
@@ -2653,7 +2708,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
             } else {
                 throw std::invalid_argument(
-                    string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
+                    string_format("error: unknown value for --log-colors: '%s'\n", value.c_str()));
             }
         }
     ).set_env("LLAMA_LOG_COLORS"));
@@ -2674,7 +2729,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_OFFLINE"));
     add_opt(common_arg(
         {"-lv", "--verbosity", "--log-verbosity"}, "N",
-        "Set the verbosity threshold. Messages with a higher verbosity will be ignored.",
+        string_format("Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:\n"
+            " - 0: generic output\n"
+            " - 1: error\n"
+            " - 2: warning\n"
+            " - 3: info\n"
+            " - 4: debug\n"
+            "(default: %d)\n", params.verbosity),
         [](common_params & params, int value) {
             params.verbosity = value;
             common_log_set_verbosity_thold(value);

common/chat-parser-xml-toolcall.cpp

@@ -724,16 +724,10 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
         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 {
+                if (!(form.allow_toolcall_in_think && tc)) {
                     unclosed_reasoning_content += tool_call_start;
+                    continue;
                 }
-                continue;
             } else {
                 reasoning_unclosed = false;
                 std::string reasoning_content;
@@ -781,8 +775,12 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
                 }
             } 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;
+                // This <tool_call> start is in thinking block
+                if (form.allow_toolcall_in_think) {
+                    unclosed_reasoning_content = content.substr(think_start + start_think.size());
+                } else {
+                    unclosed_reasoning_content = content.substr(think_start + start_think.size()) + tool_call_start;
+                }
                 reasoning_unclosed = true;
                 content.resize(think_start);
                 toolcall_in_think = true;
@@ -805,14 +803,35 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
         }
 
         // 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);
+        if (builder.pos() == builder.input().size()) {
+            if (unclosed_reasoning_content.empty()) {
+                rstrip(content);
+                trim_potential_partial_word(content);
+                rstrip(content);
+            } else {
+                rstrip(unclosed_reasoning_content);
+                trim_potential_partial_word(unclosed_reasoning_content);
+                rstrip(unclosed_reasoning_content);
+            }
+        }
+
+        // consume unclosed_reasoning_content if allow_toolcall_in_think is set
+        if (form.allow_toolcall_in_think && !unclosed_reasoning_content.empty()) {
+            if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content) {
+                builder.add_reasoning_content(unclosed_reasoning_content);
+            } else {
+                if (content.empty()) {
+                    content = start_think + unclosed_reasoning_content;
+                } else {
+                    content += "\n\n" + start_think;
+                    content += unclosed_reasoning_content;
+                }
+            }
+            unclosed_reasoning_content.clear();
         }
 
         // Add content
-        if (content.size() != 0) {
+        if (!content.empty()) {
             // 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");
@@ -820,7 +839,7 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
             builder.add_content(content);
         }
 
-        // This <tool_call> start is in thinking block, skip this tool call
+        // This <tool_call> start is in thinking block and toolcall_in_think not set, skip this tool call
         if (toolcall_in_think && !form.allow_toolcall_in_think) {
             continue;
         }
@@ -829,7 +848,7 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
         if (!tc) {
             GGML_ASSERT(builder.pos() == builder.input().size());
             GGML_ASSERT(unclosed_reasoning_content.empty());
-            GGML_ASSERT(!reasoning_unclosed);
+            if (!form.allow_toolcall_in_think) GGML_ASSERT(!reasoning_unclosed);
             break;
         }
 
@@ -854,7 +873,6 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
 
 /**
  * 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

@@ -31,7 +31,7 @@ struct xml_tool_call_format {
     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!!!
+    bool allow_toolcall_in_think = false;
 };
 
 // make a GBNF that accept any strings except those containing any of the forbidden strings.

common/chat-peg-parser.cpp

@@ -0,0 +1,114 @@
+#include "chat-peg-parser.h"
+
+#include <nlohmann/json.hpp>
+
+using json = nlohmann::json;
+
+static std::string_view trim_trailing_space(std::string_view sv) {
+    while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.back()))) {
+        sv.remove_suffix(1);
+    }
+    return sv;
+}
+
+void common_chat_peg_mapper::from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result) {
+    arena.visit(result, [this](const common_peg_ast_node & node) {
+        map(node);
+    });
+}
+
+void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
+    bool is_reasoning = node.tag == common_chat_peg_builder::REASONING;
+    bool is_content = node.tag == common_chat_peg_builder::CONTENT;
+
+    if (is_reasoning) {
+        result.reasoning_content = std::string(trim_trailing_space(node.text));
+    }
+
+    if (is_content) {
+        result.content = std::string(trim_trailing_space(node.text));
+    }
+}
+
+void common_chat_peg_native_mapper::map(const common_peg_ast_node & node) {
+    common_chat_peg_mapper::map(node);
+
+    bool is_tool_open = node.tag == common_chat_peg_native_builder::TOOL_OPEN;
+    bool is_tool_name = node.tag == common_chat_peg_native_builder::TOOL_NAME;
+    bool is_tool_id = node.tag == common_chat_peg_native_builder::TOOL_ID;
+    bool is_tool_args = node.tag == common_chat_peg_native_builder::TOOL_ARGS;
+
+    if (is_tool_open) {
+        result.tool_calls.emplace_back();
+        current_tool = &result.tool_calls.back();
+    }
+
+    if (is_tool_id && current_tool) {
+        current_tool->id = std::string(trim_trailing_space(node.text));
+    }
+
+    if (is_tool_name && current_tool) {
+        current_tool->name = std::string(trim_trailing_space(node.text));
+    }
+
+    if (is_tool_args && current_tool) {
+        current_tool->arguments = std::string(trim_trailing_space(node.text));
+    }
+}
+
+void common_chat_peg_constructed_mapper::map(const common_peg_ast_node & node) {
+    common_chat_peg_mapper::map(node);
+
+    bool is_tool_open = node.tag == common_chat_peg_constructed_builder::TOOL_OPEN;
+    bool is_tool_name = node.tag == common_chat_peg_constructed_builder::TOOL_NAME;
+    bool is_tool_close = node.tag == common_chat_peg_constructed_builder::TOOL_CLOSE;
+    bool is_arg_open = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_OPEN;
+    bool is_arg_close = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_CLOSE;
+    bool is_arg_name = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_NAME;
+    bool is_arg_string = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_STRING_VALUE;
+    bool is_arg_json = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_JSON_VALUE;
+
+    if (is_tool_open) {
+        result.tool_calls.emplace_back();
+        current_tool = &result.tool_calls.back();
+        arg_count = 0;
+    }
+
+    if (is_tool_name) {
+        current_tool->name = std::string(node.text);
+        current_tool->arguments = "{";
+    }
+
+    if (is_arg_open) {
+        needs_closing_quote = false;
+    }
+
+    if (is_arg_name && current_tool) {
+        if (arg_count > 0) {
+            current_tool->arguments += ",";
+        }
+        current_tool->arguments += json(trim_trailing_space(node.text)).dump() + ":";
+        ++arg_count;
+    }
+
+    if (is_arg_string && current_tool) {
+        // Serialize to JSON, but exclude the end quote
+        std::string dumped = json(node.text).dump();
+        current_tool->arguments += dumped.substr(0, dumped.size() - 1);
+        needs_closing_quote = true;
+    }
+
+    if (is_arg_close && current_tool) {
+        if (needs_closing_quote) {
+            current_tool->arguments += "\"";
+        }
+    }
+
+    if (is_arg_json && current_tool) {
+        current_tool->arguments += std::string(trim_trailing_space(node.text));
+    }
+
+    if (is_tool_close && current_tool) {
+        current_tool->arguments += "}";
+    }
+}

common/chat-peg-parser.h

@@ -0,0 +1,105 @@
+#pragma once
+
+#include "chat.h"
+#include "peg-parser.h"
+
+class common_chat_peg_builder : public common_peg_parser_builder {
+  public:
+    static constexpr const char * REASONING_BLOCK = "reasoning-block";
+    static constexpr const char * REASONING = "reasoning";
+    static constexpr const char * CONTENT = "content";
+
+    common_peg_parser reasoning_block(const common_peg_parser & p) { return tag(REASONING_BLOCK, p); }
+    common_peg_parser reasoning(const common_peg_parser & p) { return tag(REASONING, p); }
+    common_peg_parser content(const common_peg_parser & p) { return tag(CONTENT, p); }
+};
+
+inline common_peg_arena build_chat_peg_parser(const std::function<common_peg_parser(common_chat_peg_builder & builder)> & fn) {
+    common_chat_peg_builder builder;
+    builder.set_root(fn(builder));
+    return builder.build();
+}
+
+class common_chat_peg_mapper {
+  public:
+    common_chat_msg & result;
+
+    common_chat_peg_mapper(common_chat_msg & msg) : result(msg) {}
+
+    virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
+    virtual void map(const common_peg_ast_node & node);
+};
+
+class common_chat_peg_native_builder : public common_chat_peg_builder {
+  public:
+    static constexpr const char * TOOL = "tool";
+    static constexpr const char * TOOL_OPEN = "tool-open";
+    static constexpr const char * TOOL_CLOSE = "tool-close";
+    static constexpr const char * TOOL_ID = "tool-id";
+    static constexpr const char * TOOL_NAME = "tool-name";
+    static constexpr const char * TOOL_ARGS = "tool-args";
+
+    common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
+    common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
+    common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
+    common_peg_parser tool_id(const common_peg_parser & p) { return atomic(tag(TOOL_ID, p)); }
+    common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
+    common_peg_parser tool_args(const common_peg_parser & p) { return tag(TOOL_ARGS, p); }
+};
+
+class common_chat_peg_native_mapper : public common_chat_peg_mapper {
+    common_chat_tool_call * current_tool;
+
+  public:
+    common_chat_peg_native_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
+
+    void map(const common_peg_ast_node & node) override;
+};
+
+inline common_peg_arena build_chat_peg_native_parser(const std::function<common_peg_parser(common_chat_peg_native_builder & builder)> & fn) {
+    common_chat_peg_native_builder builder;
+    builder.set_root(fn(builder));
+    return builder.build();
+}
+
+class common_chat_peg_constructed_builder : public common_chat_peg_builder {
+  public:
+    static constexpr const char * TOOL = "tool";
+    static constexpr const char * TOOL_OPEN = "tool-open";
+    static constexpr const char * TOOL_CLOSE = "tool-close";
+    static constexpr const char * TOOL_NAME = "tool-name";
+    static constexpr const char * TOOL_ARG = "tool-arg";
+    static constexpr const char * TOOL_ARG_OPEN = "tool-arg-open";
+    static constexpr const char * TOOL_ARG_CLOSE = "tool-arg-close";
+    static constexpr const char * TOOL_ARG_NAME = "tool-arg-name";
+    static constexpr const char * TOOL_ARG_STRING_VALUE = "tool-arg-string-value";
+    static constexpr const char * TOOL_ARG_JSON_VALUE = "tool-arg-json-value";
+
+    common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
+    common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
+    common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
+    common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
+    common_peg_parser tool_arg(const common_peg_parser & p) { return tag(TOOL_ARG, p); }
+    common_peg_parser tool_arg_open(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_OPEN, p)); }
+    common_peg_parser tool_arg_close(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_CLOSE, p)); }
+    common_peg_parser tool_arg_name(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_NAME, p)); }
+    common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
+    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_JSON_VALUE, p); }
+};
+
+class common_chat_peg_constructed_mapper : public common_chat_peg_mapper {
+    common_chat_tool_call * current_tool;
+    int arg_count = 0;
+    bool needs_closing_quote = false;
+
+  public:
+    common_chat_peg_constructed_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
+
+    void map(const common_peg_ast_node & node) override;
+};
+
+inline common_peg_arena build_chat_peg_constructed_parser(const std::function<common_peg_parser(common_chat_peg_constructed_builder & builder)> & fn) {
+    common_chat_peg_constructed_builder builder;
+    builder.set_root(fn(builder));
+    return builder.build();
+}

common/chat.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include "common.h"
+#include "peg-parser.h"
 #include <functional>
 #include <chrono>
 #include <string>
@@ -76,7 +77,7 @@ struct common_chat_msg_diff {
     size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;
 
-    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg);
+    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & msg_prv, const common_chat_msg & msg_new);
 
     bool operator==(const common_chat_msg_diff & other) const {
         return content_delta == other.content_delta
@@ -124,6 +125,11 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_APRIEL_1_5,
     COMMON_CHAT_FORMAT_XIAOMI_MIMO,
 
+    // These are intended to be parsed by the PEG parser
+    COMMON_CHAT_FORMAT_PEG_SIMPLE,
+    COMMON_CHAT_FORMAT_PEG_NATIVE,
+    COMMON_CHAT_FORMAT_PEG_CONSTRUCTED,
+
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };
 
@@ -154,6 +160,7 @@ struct common_chat_params {
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
+    std::string                         parser;
 };
 
 struct common_chat_syntax {
@@ -163,6 +170,7 @@ struct common_chat_syntax {
     bool                     reasoning_in_content  = false;
     bool                     thinking_forced_open  = false;
     bool                     parse_tool_calls      = true;
+    common_peg_arena         parser                = {};
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
@@ -206,6 +214,7 @@ const char*               common_chat_format_name(common_chat_format format);
 const char*               common_reasoning_format_name(common_reasoning_format format);
 common_reasoning_format   common_reasoning_format_from_name(const std::string & format);
 common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+common_chat_msg           common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 

common/common.cpp

@@ -694,7 +694,7 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
 
 // Validate if a filename is safe to use
 // To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
-bool fs_validate_filename(const std::string & filename) {
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
     if (!filename.length()) {
         // Empty filename invalid
         return false;
@@ -754,10 +754,14 @@ bool fs_validate_filename(const std::string & filename) {
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
-            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
+            || c == ':' || c == '*' // Illegal characters
             || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
             return false;
         }
+        if (!allow_subdirs && (c == '/' || c == '\\')) {
+            // Subdirectories not allowed, reject path separators
+            return false;
+        }
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -782,11 +786,29 @@ bool fs_validate_filename(const std::string & filename) {
 #include <iostream>
 
 
+#ifdef _WIN32
+static std::wstring utf8_to_wstring(const std::string & str) {
+    if (str.empty()) {
+        return std::wstring();
+    }
+
+    int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
+
+    if (size <= 0) {
+        return std::wstring();
+    }
+
+    std::wstring wstr(size, 0);
+    MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
+
+    return wstr;
+}
+#endif
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
-    std::wstring wpath = converter.from_bytes(path);
+    std::wstring wpath = utf8_to_wstring(path);
 
     // if the path already exists, check whether it's a directory
     const DWORD attributes = GetFileAttributesW(wpath.c_str());
@@ -859,6 +881,11 @@ bool fs_create_directory_with_parents(const std::string & path) {
 #endif // _WIN32
 }
 
+bool fs_is_directory(const std::string & path) {
+    std::filesystem::path dir(path);
+    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -893,6 +920,8 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
+#elif defined(__EMSCRIPTEN__)
+        GGML_ABORT("not implemented on this platform");
 #else
 #  error Unknown architecture
 #endif
@@ -912,7 +941,7 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
-std::vector<common_file_info> fs_list_files(const std::string & path) {
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
     std::vector<common_file_info> files;
     if (path.empty()) return files;
 
@@ -927,14 +956,22 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
             const auto & p = entry.path();
             if (std::filesystem::is_regular_file(p)) {
                 common_file_info info;
-                info.path = p.string();
-                info.name = p.filename().string();
+                info.path   = p.string();
+                info.name   = p.filename().string();
+                info.is_dir = false;
                 try {
                     info.size = static_cast<size_t>(std::filesystem::file_size(p));
                 } catch (const std::filesystem::filesystem_error &) {
                     info.size = 0;
                 }
                 files.push_back(std::move(info));
+            } else if (include_directories && std::filesystem::is_directory(p)) {
+                common_file_info info;
+                info.path   = p.string();
+                info.name   = p.filename().string();
+                info.size   = 0; // Directories have no size
+                info.is_dir = true;
+                files.push_back(std::move(info));
             }
         } catch (const std::filesystem::filesystem_error &) {
             // skip entries we cannot inspect
@@ -945,6 +982,32 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
     return files;
 }
 
+//
+// TTY utils
+//
+
+bool tty_can_use_colors() {
+    // Check NO_COLOR environment variable (https://no-color.org/)
+    if (const char * no_color = std::getenv("NO_COLOR")) {
+        if (no_color[0] != '\0') {
+            return false;
+        }
+    }
+
+    // Check TERM environment variable
+    if (const char * term = std::getenv("TERM")) {
+        if (std::strcmp(term, "dumb") == 0) {
+            return false;
+        }
+    }
+
+    // Check if stdout and stderr are connected to a terminal
+    // We check both because log messages can go to either
+    bool stdout_is_tty = isatty(fileno(stdout));
+    bool stderr_is_tty = isatty(fileno(stderr));
+
+    return stdout_is_tty || stderr_is_tty;
+}
 
 //
 // Model utils

common/common.h

@@ -12,6 +12,10 @@
 #include <vector>
 #include <map>
 
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#define _WIN32_WINNT 0x0A00
+#endif
+
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
 #else
@@ -26,8 +30,6 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
-#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();
@@ -223,6 +225,7 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
+    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
 struct common_params_speculative {
@@ -369,7 +372,7 @@ struct common_params {
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 0;
+    int32_t verbosity                  = 3;  // LOG_LEVEL_INFO
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
@@ -478,9 +481,15 @@ struct common_params {
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
+    // router server configs
+    std::string models_dir = ""; // directory containing models for the router server
+    int models_max = 4;          // maximum number of models to load simultaneously
+    bool models_autoload = true; // automatically load models when requested via the router server
+
     bool log_json = false;
 
     std::string slot_save_path;
+    std::string media_path; // path to directory for loading media files
 
     float slot_prompt_similarity = 0.1f;
 
@@ -631,8 +640,9 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 // Filesystem utils
 //
 
-bool fs_validate_filename(const std::string & filename);
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
 bool fs_create_directory_with_parents(const std::string & path);
+bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
@@ -641,8 +651,16 @@ struct common_file_info {
     std::string path;
     std::string name;
     size_t      size = 0; // in bytes
+    bool        is_dir = false;
 };
-std::vector<common_file_info> fs_list_files(const std::string & path);
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
+
+//
+// TTY utils
+//
+
+// Auto-detect if colors can be enabled based on terminal and environment
+bool tty_can_use_colors();
 
 //
 // Model utils

common/download.cpp

@@ -24,6 +24,7 @@
 #include "http.h"
 #endif
 
+#ifndef __EMSCRIPTEN__
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
@@ -35,6 +36,8 @@
 #else
 #include <sys/syslimits.h>
 #endif
+#endif
+
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
 // isatty
@@ -430,7 +433,7 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-    curl_easy_setopt(curl.get(), CURLOPT_VERBOSE, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_VERBOSE, 0L);
     typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
     auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
         auto data_vec = static_cast<std::vector<char> *>(data);
@@ -517,16 +520,18 @@ static bool common_pull_file(httplib::Client & cli,
         headers.emplace("Range", "bytes=" + std::to_string(existing_size) + "-");
     }
 
-    std::atomic<size_t> downloaded{existing_size};
+    const char * func = __func__; // avoid __func__ inside a lambda
+    size_t downloaded = existing_size;
+    size_t progress_step = 0;
 
     auto res = cli.Get(resolve_path, headers,
         [&](const httplib::Response &response) {
             if (existing_size > 0 && response.status != 206) {
-                LOG_WRN("%s: server did not respond with 206 Partial Content for a resume request. Status: %d\n", __func__, response.status);
+                LOG_WRN("%s: server did not respond with 206 Partial Content for a resume request. Status: %d\n", func, response.status);
                 return false;
             }
             if (existing_size == 0 && response.status != 200) {
-                LOG_WRN("%s: download received non-successful status code: %d\n", __func__, response.status);
+                LOG_WRN("%s: download received non-successful status code: %d\n", func, response.status);
                 return false;
             }
             if (total_size == 0 && response.has_header("Content-Length")) {
@@ -534,7 +539,7 @@ static bool common_pull_file(httplib::Client & cli,
                     size_t content_length = std::stoull(response.get_header_value("Content-Length"));
                     total_size = existing_size + content_length;
                 } catch (const std::exception &e) {
-                    LOG_WRN("%s: invalid Content-Length header: %s\n", __func__, e.what());
+                    LOG_WRN("%s: invalid Content-Length header: %s\n", func, e.what());
                 }
             }
             return true;
@@ -542,11 +547,16 @@ static bool common_pull_file(httplib::Client & cli,
         [&](const char *data, size_t len) {
             ofs.write(data, len);
             if (!ofs) {
-                LOG_ERR("%s: error writing to file: %s\n", __func__, path_tmp.c_str());
+                LOG_ERR("%s: error writing to file: %s\n", func, path_tmp.c_str());
                 return false;
             }
             downloaded += len;
-            print_progress(downloaded, total_size);
+            progress_step += len;
+
+            if (progress_step >= total_size / 1000 || downloaded == total_size) {
+                print_progress(downloaded, total_size);
+                progress_step = 0;
+            }
             return true;
         },
         nullptr
@@ -1047,7 +1057,7 @@ std::string common_docker_resolve_model(const std::string &) {
 std::vector<common_cached_model_info> common_list_cached_models() {
     std::vector<common_cached_model_info> models;
     const std::string cache_dir = fs_get_cache_directory();
-    const std::vector<common_file_info> files = fs_list_files(cache_dir);
+    const std::vector<common_file_info> files = fs_list(cache_dir, false);
     for (const auto & file : files) {
         if (string_starts_with(file.name, "manifest=") && string_ends_with(file.name, ".json")) {
             common_cached_model_info model_info;

common/download.h

@@ -14,8 +14,10 @@ struct common_cached_model_info {
     std::string model;
     std::string tag;
     size_t      size = 0; // GGUF size in bytes
+    // return string representation like "user/model:tag"
+    // if tag is "latest", it will be omitted
     std::string to_string() const {
-        return user + "/" + model + ":" + tag;
+        return user + "/" + model + (tag == "latest" ? "" : ":" + tag);
     }
 };
 

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 = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
@@ -974,7 +974,7 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
+            throw std::invalid_argument("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
         }
         if (!_warnings.empty()) {
             fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());

common/log.cpp

@@ -1,3 +1,4 @@
+#include "common.h"
 #include "log.h"
 
 #include <chrono>
@@ -26,30 +27,6 @@ void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
-// Auto-detect if colors should be enabled based on terminal and environment
-static bool common_log_should_use_colors_auto() {
-    // Check NO_COLOR environment variable (https://no-color.org/)
-    if (const char * no_color = std::getenv("NO_COLOR")) {
-        if (no_color[0] != '\0') {
-            return false;
-        }
-    }
-
-    // Check TERM environment variable
-    if (const char * term = std::getenv("TERM")) {
-        if (std::strcmp(term, "dumb") == 0) {
-            return false;
-        }
-    }
-
-    // Check if stdout and stderr are connected to a terminal
-    // We check both because log messages can go to either
-    bool stdout_is_tty = isatty(fileno(stdout));
-    bool stderr_is_tty = isatty(fileno(stderr));
-
-    return stdout_is_tty || stderr_is_tty;
-}
-
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -391,7 +368,7 @@ struct common_log * common_log_main() {
     static std::once_flag    init_flag;
     std::call_once(init_flag, [&]() {
         // Set default to auto-detect colors
-        log.set_colors(common_log_should_use_colors_auto());
+        log.set_colors(tty_can_use_colors());
     });
 
     return &log;
@@ -422,7 +399,7 @@ void common_log_set_file(struct common_log * log, const char * file) {
 
 void common_log_set_colors(struct common_log * log, log_colors colors) {
     if (colors == LOG_COLORS_AUTO) {
-        log->set_colors(common_log_should_use_colors_auto());
+        log->set_colors(tty_can_use_colors());
         return;
     }
 
@@ -443,8 +420,22 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
+static int common_get_verbosity(enum ggml_log_level level) {
+    switch (level) {
+        case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;
+        case GGML_LOG_LEVEL_INFO:  return LOG_LEVEL_INFO;
+        case GGML_LOG_LEVEL_WARN:  return LOG_LEVEL_WARN;
+        case GGML_LOG_LEVEL_ERROR: return LOG_LEVEL_ERROR;
+        case GGML_LOG_LEVEL_CONT:  return LOG_LEVEL_INFO; // same as INFO
+        case GGML_LOG_LEVEL_NONE:
+        default:
+            return LOG_LEVEL_OUTPUT;
+    }
+}
+
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
-    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+    auto verbosity = common_get_verbosity(level);
+    if (verbosity <= common_log_verbosity_thold) {
         common_log_add(common_log_main(), level, "%s", text);
     }
 }

common/log.h

@@ -21,8 +21,14 @@
 #    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#define LOG_DEFAULT_DEBUG 1
-#define LOG_DEFAULT_LLAMA 0
+#define LOG_LEVEL_DEBUG  4
+#define LOG_LEVEL_INFO   3
+#define LOG_LEVEL_WARN   2
+#define LOG_LEVEL_ERROR  1
+#define LOG_LEVEL_OUTPUT 0 // output data from tools
+
+#define LOG_DEFAULT_DEBUG LOG_LEVEL_DEBUG
+#define LOG_DEFAULT_LLAMA LOG_LEVEL_INFO
 
 enum log_colors {
     LOG_COLORS_AUTO     = -1,
@@ -67,10 +73,11 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 //   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
 //   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
 //
-// I - info    (stdout, V = 0)
-// W - warning (stderr, V = 0)
-// E - error   (stderr, V = 0)
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
+// I - info    (stdout, V = LOG_DEFAULT_INFO)
+// W - warning (stderr, V = LOG_DEFAULT_WARN)
+// E - error   (stderr, V = LOG_DEFAULT_ERROR)
+// O - output  (stdout, V = LOG_DEFAULT_OUTPUT)
 //
 
 void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
@@ -95,14 +102,14 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps);   // w
         } \
     } while (0)
 
-#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
-#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
+#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, LOG_LEVEL_OUTPUT, __VA_ARGS__)
+#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity,        __VA_ARGS__)
 
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_LEVEL_DEBUG,  __VA_ARGS__)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  LOG_LEVEL_INFO,   __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  LOG_LEVEL_WARN,   __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, LOG_LEVEL_ERROR,  __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  LOG_LEVEL_INFO,   __VA_ARGS__) // same as INFO
 
 #define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)

common/peg-parser.h

@@ -0,0 +1,459 @@
+#pragma once
+
+#include <nlohmann/json_fwd.hpp>
+
+#include <memory>
+#include <unordered_map>
+#include <string>
+#include <string_view>
+#include <functional>
+#include <vector>
+#include <variant>
+
+struct common_grammar_builder;
+
+class common_peg_parser_builder;
+
+using common_peg_parser_id = size_t;
+constexpr common_peg_parser_id COMMON_PEG_INVALID_PARSER_ID = static_cast<common_peg_parser_id>(-1);
+
+using common_peg_ast_id = size_t;
+constexpr common_peg_ast_id COMMON_PEG_INVALID_AST_ID = static_cast<common_peg_ast_id>(-1);
+
+// Lightweight wrapper around common_peg_parser_id for convenience
+class common_peg_parser {
+    common_peg_parser_id id_;
+    common_peg_parser_builder & builder_;
+
+  public:
+    common_peg_parser(const common_peg_parser & other) : id_(other.id_), builder_(other.builder_) {}
+    common_peg_parser(common_peg_parser_id id, common_peg_parser_builder & builder) : id_(id), builder_(builder) {}
+
+    common_peg_parser & operator=(const common_peg_parser & other);
+    common_peg_parser & operator+=(const common_peg_parser & other);
+    common_peg_parser & operator|=(const common_peg_parser & other);
+
+    operator common_peg_parser_id() const { return id_; }
+    common_peg_parser_id id() const { return id_; }
+
+    common_peg_parser_builder & builder() const { return builder_; }
+
+    // Creates a sequence
+    common_peg_parser operator+(const common_peg_parser & other) const;
+
+    // Creates a sequence separated by spaces.
+    common_peg_parser operator<<(const common_peg_parser & other) const;
+
+    // Creates a choice
+    common_peg_parser operator|(const common_peg_parser & other) const;
+
+    common_peg_parser operator+(const char * str) const;
+    common_peg_parser operator+(const std::string & str) const;
+    common_peg_parser operator<<(const char * str) const;
+    common_peg_parser operator<<(const std::string & str) const;
+    common_peg_parser operator|(const char * str) const;
+    common_peg_parser operator|(const std::string & str) const;
+};
+
+common_peg_parser operator+(const char * str, const common_peg_parser & p);
+common_peg_parser operator+(const std::string & str, const common_peg_parser & p);
+common_peg_parser operator<<(const char * str, const common_peg_parser & p);
+common_peg_parser operator<<(const std::string & str, const common_peg_parser & p);
+common_peg_parser operator|(const char * str, const common_peg_parser & p);
+common_peg_parser operator|(const std::string & str, const common_peg_parser & p);
+
+enum common_peg_parse_result_type {
+    COMMON_PEG_PARSE_RESULT_FAIL            = 0,
+    COMMON_PEG_PARSE_RESULT_SUCCESS         = 1,
+    COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT = 2,
+};
+
+const char * common_peg_parse_result_type_name(common_peg_parse_result_type type);
+
+struct common_peg_ast_node {
+    common_peg_ast_id id;
+    std::string rule;
+    std::string tag;
+    size_t start;
+    size_t end;
+    std::string_view text;
+    std::vector<common_peg_ast_id> children;
+
+    bool is_partial = false;
+};
+
+struct common_peg_parse_result;
+
+using common_peg_ast_visitor = std::function<void(const common_peg_ast_node & node)>;
+
+class common_peg_ast_arena {
+    std::vector<common_peg_ast_node> nodes_;
+  public:
+    common_peg_ast_id add_node(
+        const std::string & rule,
+        const std::string & tag,
+        size_t start,
+        size_t end,
+        std::string_view text,
+        std::vector<common_peg_ast_id> children,
+        bool is_partial = false
+    ) {
+        common_peg_ast_id id = nodes_.size();
+        nodes_.push_back({id, rule, tag, start, end, text, std::move(children), is_partial});
+        return id;
+    }
+
+    const common_peg_ast_node & get(common_peg_ast_id id) const { return nodes_.at(id); }
+
+    size_t size() const { return nodes_.size(); }
+
+    void clear() { nodes_.clear(); }
+
+    void visit(common_peg_ast_id id, const common_peg_ast_visitor & visitor) const;
+    void visit(const common_peg_parse_result & result, const common_peg_ast_visitor & visitor) const;
+};
+
+struct common_peg_parse_result {
+    common_peg_parse_result_type type = COMMON_PEG_PARSE_RESULT_FAIL;
+    size_t start = 0;
+    size_t end = 0;
+
+    std::vector<common_peg_ast_id> nodes;
+
+    common_peg_parse_result() = default;
+
+    common_peg_parse_result(common_peg_parse_result_type type, size_t start)
+        : type(type), start(start), end(start) {}
+
+    common_peg_parse_result(common_peg_parse_result_type type, size_t start, size_t end)
+        : type(type), start(start), end(end) {}
+
+    common_peg_parse_result(common_peg_parse_result_type type, size_t start, size_t end, std::vector<common_peg_ast_id> nodes)
+        : type(type), start(start), end(end), nodes(std::move(nodes)) {}
+
+    bool fail() const { return type == COMMON_PEG_PARSE_RESULT_FAIL; }
+    bool need_more_input() const { return type == COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT; }
+    bool success() const { return type == COMMON_PEG_PARSE_RESULT_SUCCESS; }
+};
+
+struct common_peg_parse_context {
+    std::string input;
+    bool is_partial;
+    common_peg_ast_arena ast;
+
+    int parse_depth;
+
+    common_peg_parse_context()
+        : is_partial(false), parse_depth(0) {}
+
+    common_peg_parse_context(const std::string & input)
+        : input(input), is_partial(false), parse_depth(0) {}
+
+    common_peg_parse_context(const std::string & input, bool is_partial)
+        : input(input), is_partial(is_partial), parse_depth(0) {}
+};
+
+class common_peg_arena;
+
+// Parser variants
+struct common_peg_epsilon_parser {};
+
+struct common_peg_start_parser {};
+
+struct common_peg_end_parser {};
+
+struct common_peg_literal_parser {
+    std::string literal;
+};
+
+struct common_peg_sequence_parser {
+    std::vector<common_peg_parser_id> children;
+};
+
+struct common_peg_choice_parser {
+    std::vector<common_peg_parser_id> children;
+};
+
+struct common_peg_repetition_parser {
+    common_peg_parser_id child;
+    int min_count;
+    int max_count;  // -1 for unbounded
+};
+
+struct common_peg_and_parser {
+    common_peg_parser_id child;
+};
+
+struct common_peg_not_parser {
+    common_peg_parser_id child;
+};
+
+struct common_peg_any_parser {};
+
+struct common_peg_space_parser {};
+
+struct common_peg_chars_parser {
+    struct char_range {
+        uint32_t start;
+        uint32_t end;
+        bool contains(uint32_t codepoint) const { return codepoint >= start && codepoint <= end; }
+    };
+
+    std::string pattern;
+    std::vector<char_range> ranges;
+    bool negated;
+    int min_count;
+    int max_count;  // -1 for unbounded
+};
+
+struct common_peg_json_string_parser {};
+
+struct common_peg_until_parser {
+    std::vector<std::string> delimiters;
+};
+
+struct common_peg_schema_parser {
+    common_peg_parser_id child;
+    std::string name;
+    std::shared_ptr<nlohmann::ordered_json> schema;
+
+    // Indicates if the GBNF should accept a raw string that matches the schema.
+    bool raw;
+};
+
+struct common_peg_rule_parser {
+    std::string name;
+    common_peg_parser_id child;
+    bool trigger;
+};
+
+struct common_peg_ref_parser {
+    std::string name;
+};
+
+struct common_peg_atomic_parser {
+    common_peg_parser_id child;
+};
+
+struct common_peg_tag_parser {
+    common_peg_parser_id child;
+    std::string tag;
+};
+
+// Variant holding all parser types
+using common_peg_parser_variant = std::variant<
+    common_peg_epsilon_parser,
+    common_peg_start_parser,
+    common_peg_end_parser,
+    common_peg_literal_parser,
+    common_peg_sequence_parser,
+    common_peg_choice_parser,
+    common_peg_repetition_parser,
+    common_peg_and_parser,
+    common_peg_not_parser,
+    common_peg_any_parser,
+    common_peg_space_parser,
+    common_peg_chars_parser,
+    common_peg_json_string_parser,
+    common_peg_until_parser,
+    common_peg_schema_parser,
+    common_peg_rule_parser,
+    common_peg_ref_parser,
+    common_peg_atomic_parser,
+    common_peg_tag_parser
+>;
+
+class common_peg_arena {
+    std::vector<common_peg_parser_variant> parsers_;
+    std::unordered_map<std::string, common_peg_parser_id> rules_;
+    common_peg_parser_id root_ = COMMON_PEG_INVALID_PARSER_ID;
+
+  public:
+    const common_peg_parser_variant & get(common_peg_parser_id id) const { return parsers_.at(id); }
+    common_peg_parser_variant & get(common_peg_parser_id id) { return parsers_.at(id); }
+
+    size_t size() const { return parsers_.size(); }
+    bool empty() const { return parsers_.empty(); }
+
+    common_peg_parser_id get_rule(const std::string & name) const;
+    bool has_rule(const std::string & name) const { return rules_.find(name) != rules_.end(); }
+
+    common_peg_parser_id root() const { return root_; }
+    void set_root(common_peg_parser_id id) { root_ = id; }
+
+    common_peg_parse_result parse(common_peg_parse_context & ctx, size_t start = 0) const;
+    common_peg_parse_result parse(common_peg_parser_id id, common_peg_parse_context & ctx, size_t start) const;
+
+    void resolve_refs();
+
+    void build_grammar(const common_grammar_builder & builder, bool lazy = false) const;
+
+    std::string dump(common_peg_parser_id id) const;
+
+    nlohmann::json to_json() const;
+    static common_peg_arena from_json(const nlohmann::json & j);
+
+    std::string save() const;
+    void load(const std::string & data);
+
+    friend class common_peg_parser_builder;
+
+  private:
+    common_peg_parser_id add_parser(common_peg_parser_variant parser);
+    void add_rule(const std::string & name, common_peg_parser_id id);
+
+    common_peg_parser_id resolve_ref(common_peg_parser_id id);
+};
+
+class common_peg_parser_builder {
+    common_peg_arena arena_;
+
+    common_peg_parser wrap(common_peg_parser_id id) { return common_peg_parser(id, *this); }
+    common_peg_parser add(const common_peg_parser_variant & p) { return wrap(arena_.add_parser(p)); }
+
+  public:
+    common_peg_parser_builder();
+
+    // Match nothing, always succeed.
+    //   S -> ε
+    common_peg_parser eps() { return add(common_peg_epsilon_parser{}); }
+
+    // Matches the start of the input.
+    //   S -> ^
+    common_peg_parser start() { return add(common_peg_start_parser{}); }
+
+    // Matches the end of the input.
+    //   S -> $
+    common_peg_parser end() { return add(common_peg_end_parser{}); }
+
+    // Matches an exact literal string.
+    //   S -> "hello"
+    common_peg_parser literal(const std::string & literal) { return add(common_peg_literal_parser{literal}); }
+
+    // Matches a sequence of parsers in order, all must succeed.
+    //   S -> A B C
+    common_peg_parser sequence() { return add(common_peg_sequence_parser{}); }
+    common_peg_parser sequence(const std::vector<common_peg_parser_id> & parsers);
+    common_peg_parser sequence(const std::vector<common_peg_parser> & parsers);
+    common_peg_parser sequence(std::initializer_list<common_peg_parser> parsers);
+
+    // Matches the first parser that succeeds from a list of alternatives.
+    //   S -> A | B | C
+    common_peg_parser choice() { return add(common_peg_choice_parser{}); }
+    common_peg_parser choice(const std::vector<common_peg_parser_id> & parsers);
+    common_peg_parser choice(const std::vector<common_peg_parser> & parsers);
+    common_peg_parser choice(std::initializer_list<common_peg_parser> parsers);
+
+    // Matches one or more repetitions of a parser.
+    //   S -> A+
+    common_peg_parser one_or_more(const common_peg_parser & p) { return repeat(p, 1, -1); }
+
+    // Matches zero or more repetitions of a parser, always succeeds.
+    //   S -> A*
+    common_peg_parser zero_or_more(const common_peg_parser & p) { return repeat(p, 0, -1); }
+
+    // Matches zero or one occurrence of a parser, always succeeds.
+    //   S -> A?
+    common_peg_parser optional(const common_peg_parser & p) { return repeat(p, 0, 1); }
+
+    // Positive lookahead: succeeds if child parser succeeds, consumes no input.
+    //   S -> &A
+    common_peg_parser peek(const common_peg_parser & p) { return add(common_peg_and_parser{p}); }
+
+    // Negative lookahead: succeeds if child parser fails, consumes no input.
+    //   S -> !A
+    common_peg_parser negate(const common_peg_parser & p) { return add(common_peg_not_parser{p}); }
+
+    // Matches any single character.
+    //   S -> .
+    common_peg_parser any() { return add(common_peg_any_parser{}); }
+
+    // Matches between min and max repetitions of characters from a character class.
+    //   S -> [a-z]{m,n}
+    //
+    // Use -1 for max to represent unbounded repetition (equivalent to {m,})
+    common_peg_parser chars(const std::string & classes, int min = 1, int max = -1);
+
+    // Creates a lightweight reference to a named rule (resolved during build()).
+    // Use this for forward references in recursive grammars.
+    //   expr_ref -> expr
+    common_peg_parser ref(const std::string & name) { return add(common_peg_ref_parser{name}); }
+
+    // Matches zero or more whitespace characters (space, tab, newline).
+    //   S -> [ \t\n]*
+    common_peg_parser space() { return add(common_peg_space_parser{}); }
+
+    // Matches all characters until a delimiter is found (delimiter not consumed).
+    //   S -> (!delim .)*
+    common_peg_parser until(const std::string & delimiter) { return add(common_peg_until_parser{{delimiter}}); }
+
+    // Matches all characters until one of the delimiters in the list is found (delimiter not consumed).
+    //   S -> (!delim .)*
+    common_peg_parser until_one_of(const std::vector<std::string> & delimiters) { return add(common_peg_until_parser{delimiters}); }
+
+    // Matches everything
+    //   S -> .*
+    common_peg_parser rest() { return until_one_of({}); }
+
+    // Matches between min and max repetitions of a parser (inclusive).
+    //   S -> A{m,n}
+    // Use -1 for max to represent unbounded repetition (equivalent to {m,})
+    common_peg_parser repeat(const common_peg_parser & p, int min, int max) { return add(common_peg_repetition_parser{p, min,max}); }
+
+    // Matches exactly n repetitions of a parser.
+    //   S -> A{n}
+    common_peg_parser repeat(const common_peg_parser & p, int n) { return repeat(p, n, n); }
+
+    // Creates a complete JSON parser supporting objects, arrays, strings, numbers, booleans, and null.
+    //   value -> object | array | string | number | true | false | null
+    common_peg_parser json();
+    common_peg_parser json_object();
+    common_peg_parser json_string();
+    common_peg_parser json_array();
+    common_peg_parser json_number();
+    common_peg_parser json_bool();
+    common_peg_parser json_null();
+
+    // Matches JSON string content without the surrounding quotes.
+    // Useful for extracting content within a JSON string.
+    common_peg_parser json_string_content();
+
+    // Matches a JSON object member with a key and associated parser as the
+    // value.
+    common_peg_parser json_member(const std::string & key, const common_peg_parser & p);
+
+    // Wraps a parser with JSON schema metadata for grammar generation.
+    // Used internally to convert JSON schemas to GBNF grammar rules.
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);
+
+    // Creates a named rule, stores it in the grammar, and returns a ref.
+    // If trigger=true, marks this rule as an entry point for lazy grammar generation.
+    //   auto json = p.rule("json", json_obj | json_arr | ...)
+    common_peg_parser rule(const std::string & name, const common_peg_parser & p, bool trigger = false);
+
+    // Creates a named rule using a builder function, and returns a ref.
+    // If trigger=true, marks this rule as an entry point for lazy grammar generation.
+    //   auto json = p.rule("json", [&]() { return json_object() | json_array() | ... })
+    common_peg_parser rule(const std::string & name, const std::function<common_peg_parser()> & builder, bool trigger = false);
+
+    // Creates a trigger rule. When generating a lazy grammar from the parser,
+    // only trigger rules and descendents are emitted.
+    common_peg_parser trigger_rule(const std::string & name, const common_peg_parser & p) { return rule(name, p, true); }
+    common_peg_parser trigger_rule(const std::string & name, const std::function<common_peg_parser()> & builder) { return rule(name, builder, true); }
+
+    // Creates an atomic parser. Atomic parsers do not create an AST node if
+    // the child results in a partial parse, i.e. NEEDS_MORE_INPUT. This is
+    // intended for situations where partial output is undesirable.
+    common_peg_parser atomic(const common_peg_parser & p) { return add(common_peg_atomic_parser{p}); }
+
+    // Tags create nodes in the generated AST for semantic purposes.
+    // Unlike rules, you can tag multiple nodes with the same tag.
+    common_peg_parser tag(const std::string & tag, const common_peg_parser & p) { return add(common_peg_tag_parser{p.id(), tag}); }
+
+    void set_root(const common_peg_parser & p);
+
+    common_peg_arena build();
+};
+
+// Helper function for building parsers
+common_peg_arena build_peg_parser(const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn);

common/unicode.cpp

@@ -0,0 +1,64 @@
+#include "unicode.h"
+
+// implementation adopted from src/unicode.cpp
+
+size_t utf8_sequence_length(unsigned char first_byte) {
+    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    uint8_t highbits = static_cast<uint8_t>(first_byte) >> 4;
+    return lookup[highbits];
+}
+
+utf8_parse_result parse_utf8_codepoint(std::string_view input, size_t offset) {
+    if (offset >= input.size()) {
+        return utf8_parse_result(utf8_parse_result::INCOMPLETE);
+    }
+
+    // ASCII fast path
+    if (!(input[offset] & 0x80)) {
+        return utf8_parse_result(utf8_parse_result::SUCCESS, input[offset], 1);
+    }
+
+    // Invalid: continuation byte as first byte
+    if (!(input[offset] & 0x40)) {
+        return utf8_parse_result(utf8_parse_result::INVALID);
+    }
+
+    // 2-byte sequence
+    if (!(input[offset] & 0x20)) {
+        if (offset + 1 >= input.size()) {
+            return utf8_parse_result(utf8_parse_result::INCOMPLETE);
+        }
+        if ((input[offset + 1] & 0xc0) != 0x80) {
+            return utf8_parse_result(utf8_parse_result::INVALID);
+        }
+        auto result = ((input[offset] & 0x1f) << 6) | (input[offset + 1] & 0x3f);
+        return utf8_parse_result(utf8_parse_result::SUCCESS, result, 2);
+    }
+
+    // 3-byte sequence
+    if (!(input[offset] & 0x10)) {
+        if (offset + 2 >= input.size()) {
+            return utf8_parse_result(utf8_parse_result::INCOMPLETE);
+        }
+        if ((input[offset + 1] & 0xc0) != 0x80 || (input[offset + 2] & 0xc0) != 0x80) {
+            return utf8_parse_result(utf8_parse_result::INVALID);
+        }
+        auto result = ((input[offset] & 0x0f) << 12) | ((input[offset + 1] & 0x3f) << 6) | (input[offset + 2] & 0x3f);
+        return utf8_parse_result(utf8_parse_result::SUCCESS, result, 3);
+    }
+
+    // 4-byte sequence
+    if (!(input[offset] & 0x08)) {
+        if (offset + 3 >= input.size()) {
+            return utf8_parse_result(utf8_parse_result::INCOMPLETE);
+        }
+        if ((input[offset + 1] & 0xc0) != 0x80 || (input[offset + 2] & 0xc0) != 0x80 || (input[offset + 3] & 0xc0) != 0x80) {
+            return utf8_parse_result(utf8_parse_result::INVALID);
+        }
+        auto result = ((input[offset] & 0x07) << 18) | ((input[offset + 1] & 0x3f) << 12) | ((input[offset + 2] & 0x3f) << 6) | (input[offset + 3] & 0x3f);
+        return utf8_parse_result(utf8_parse_result::SUCCESS, result, 4);
+    }
+
+    // Invalid first byte
+    return utf8_parse_result(utf8_parse_result::INVALID);
+}

common/unicode.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include <cstdint>
+#include <string_view>
+
+// UTF-8 parsing utilities for streaming-aware unicode support
+
+struct utf8_parse_result {
+    uint32_t codepoint;      // Decoded codepoint (only valid if status == SUCCESS)
+    size_t bytes_consumed;   // How many bytes this codepoint uses (1-4)
+    enum status { SUCCESS, INCOMPLETE, INVALID } status;
+
+    utf8_parse_result(enum status s, uint32_t cp = 0, size_t bytes = 0)
+        : codepoint(cp), bytes_consumed(bytes), status(s) {}
+};
+
+// Determine the expected length of a UTF-8 sequence from its first byte
+// Returns 0 for invalid first bytes
+size_t utf8_sequence_length(unsigned char first_byte);
+
+// Parse a single UTF-8 codepoint from input
+utf8_parse_result parse_utf8_codepoint(std::string_view input, size_t offset);

convert_hf_to_gguf.py

@@ -1524,6 +1524,79 @@ class TextModel(ModelBase):
         special_vocab._set_special_token("bos", 151643)
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_mistral(self):
+        if not _mistral_common_installed:
+            raise ImportError(_mistral_import_error_msg)
+
+        vocab = MistralVocab(self.dir_model)
+        logger.info(
+            f"Converting tokenizer {vocab.tokenizer_type} of size {vocab.vocab_size}."
+        )
+
+        self.gguf_writer.add_tokenizer_model(vocab.gguf_tokenizer_model)
+
+        tokens = []
+        scores = []
+        toktypes = []
+
+        for text, score, toktype in vocab.all_tokens():
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        assert len(tokens) == vocab.vocab_size, (
+            f"token count ({len(tokens)}) != vocab size ({vocab.vocab_size})"
+        )
+
+        if vocab.tokenizer_type == MistralTokenizerType.tekken:
+            self.gguf_writer.add_tokenizer_pre("tekken")
+            self.gguf_writer.add_token_merges(
+                vocab.extract_vocab_merges_from_model()
+            )
+
+        logger.info(
+            f"Setting bos, eos, unk and pad token IDs to {vocab.bos_id}, {vocab.eos_id}, {vocab.unk_id}, {vocab.pad_id}."
+        )
+
+        self.gguf_writer.add_bos_token_id(vocab.bos_id)
+        self.gguf_writer.add_eos_token_id(vocab.eos_id)
+        self.gguf_writer.add_unk_token_id(vocab.unk_id)
+        self.gguf_writer.add_pad_token_id(vocab.pad_id)
+
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_vocab_size(vocab.vocab_size)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(False)
+
+        local_template_file_path = self.dir_model / "chat_template.jinja"
+
+        if self.is_mistral_format and local_template_file_path.is_file():
+            # Ministral-3 and other new Mistral models come with chat templates.
+            # ref: https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512/tree/main
+            logger.info("Using an existing Mistral local chat template.")
+
+            with open(local_template_file_path, "r", encoding="utf-8") as f:
+                template = f.read()
+        elif not self.is_mistral_format or not self.disable_mistral_community_chat_template:
+            template_dir = Path(__file__).parent / "models/templates/"
+
+            # Log only for Mistral format that the official tokenization and detokenization is via `mistral-common`.
+            if self.is_mistral_format:
+                logger.info(
+                    "Using a Mistral community chat template. These templates can be subject to errors in early days or weeks after a release. "
+                    "Mistral recommends to use `mistral-common` to perform tokenization and detokenization."
+                )
+            template = MistralModel.get_community_chat_template(vocab, template_dir, self.is_mistral_format)
+        else:
+            logger.info("Not using a Mistral local or community chat template. Ensure to perform the tokenization and detokenization via `mistral-common`.")
+            template = None
+
+        if template is not None:
+            self.gguf_writer.add_chat_template(template)
+
 
 class MmprojModel(ModelBase):
     model_type = ModelType.MMPROJ
@@ -1581,10 +1654,27 @@ class MmprojModel(ModelBase):
 
         # load preprocessor config
         self.preprocessor_config = {}
-        if not self.is_mistral_format:
-            with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
+
+        # prefer preprocessor_config.json if possible
+        preprocessor_config_path = self.dir_model / "preprocessor_config.json"
+        if preprocessor_config_path.is_file():
+            with open(preprocessor_config_path, "r", encoding="utf-8") as f:
                 self.preprocessor_config = json.load(f)
 
+        # prefer processor_config.json if possible
+        processor_config_path = self.dir_model / "processor_config.json"
+        if processor_config_path.is_file():
+            with open(processor_config_path, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+                # move image_processor to root level for compat
+                if "image_processor" in cfg:
+                    cfg = {
+                        **cfg,
+                        **cfg["image_processor"],
+                    }
+                # merge configs
+                self.preprocessor_config = {**self.preprocessor_config, **cfg}
+
     def get_vision_config(self) -> dict[str, Any] | None:
         config_name = "vision_config" if not self.is_mistral_format else "vision_encoder"
         return self.global_config.get(config_name)
@@ -2277,67 +2367,6 @@ class LlamaModel(TextModel):
         if self.hf_arch == "VLlama3ForCausalLM":
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
 
-    def _set_vocab_mistral(self):
-        if not _mistral_common_installed:
-            raise ImportError(_mistral_import_error_msg)
-
-        vocab = MistralVocab(self.dir_model)
-        logger.info(
-            f"Converting tokenizer {vocab.tokenizer_type} of size {vocab.vocab_size}."
-        )
-
-        self.gguf_writer.add_tokenizer_model(vocab.gguf_tokenizer_model)
-
-        tokens = []
-        scores = []
-        toktypes = []
-
-        for text, score, toktype in vocab.all_tokens():
-            tokens.append(text)
-            scores.append(score)
-            toktypes.append(toktype)
-
-        assert len(tokens) == vocab.vocab_size, (
-            f"token count ({len(tokens)}) != vocab size ({vocab.vocab_size})"
-        )
-
-        if vocab.tokenizer_type == MistralTokenizerType.tekken:
-            self.gguf_writer.add_tokenizer_pre("tekken")
-            self.gguf_writer.add_token_merges(
-                vocab.extract_vocab_merges_from_model()
-            )
-
-        logger.info(
-            f"Setting bos, eos, unk and pad token IDs to {vocab.bos_id}, {vocab.eos_id}, {vocab.unk_id}, {vocab.pad_id}."
-        )
-
-        self.gguf_writer.add_bos_token_id(vocab.bos_id)
-        self.gguf_writer.add_eos_token_id(vocab.eos_id)
-        self.gguf_writer.add_unk_token_id(vocab.unk_id)
-        self.gguf_writer.add_pad_token_id(vocab.pad_id)
-
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-        self.gguf_writer.add_vocab_size(vocab.vocab_size)
-
-        self.gguf_writer.add_add_bos_token(True)
-        self.gguf_writer.add_add_eos_token(False)
-
-        template_dir = Path(__file__).parent / "models/templates/"
-
-        if not self.is_mistral_format or not self.disable_mistral_community_chat_template:
-            # Log only for Mistral format that the official tokenization and detokenization is via `mistral-common`.
-            if self.is_mistral_format:
-                logger.info(
-                    "Using a Mistral community chat template. These templates can be subject to errors in early days or weeks after a release. "
-                    "Mistral recommends to use `mistral-common` to perform tokenization and detokenization."
-                )
-            template = MistralModel.get_community_chat_template(vocab, template_dir, self.is_mistral_format)
-            self.gguf_writer.add_chat_template(template)
-        else:
-            logger.info("Not using a Mistral community chat template. Ensure to perform the tokenization and detokenization via `mistral-common`.")
-
     def set_vocab(self):
         if self.is_mistral_format:
             return self._set_vocab_mistral()
@@ -2797,9 +2826,38 @@ class Llama4VisionModel(MmprojModel):
 
 @ModelBase.register("Mistral3ForConditionalGeneration")
 class Mistral3Model(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.LLAMA
+    model_arch = gguf.MODEL_ARCH.MISTRAL3
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # for compatibility, we use LLAMA arch for older models
+        # TODO: remove this once everyone has migrated to newer version of llama.cpp
+        if self.hparams.get("model_type") != "ministral3":
+            self.model_arch = gguf.MODEL_ARCH.LLAMA
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
+            self.gguf_writer.add_architecture()
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        rope_params = self.hparams.get("rope_parameters")
+        if self.hparams.get("model_type") == "ministral3":
+            assert rope_params is not None, "ministral3 must have 'rope_parameters' config"
+            assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_params["factor"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_fast(rope_params["beta_fast"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_slow(rope_params["beta_slow"])
+            self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_params["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_freq_base(rope_params["rope_theta"])
+            self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # TODO: probably not worth supporting quantized weight, as official BF16 is also available
+        if name.endswith("weight_scale_inv"):
+            raise ValueError("This is a quantized weight, please use BF16 weight instead")
+
         name = name.replace("language_model.", "")
         if "multi_modal_projector" in name or "vision_tower" in name:
             return []
@@ -4183,6 +4241,36 @@ 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
@@ -5737,9 +5825,11 @@ class Gemma3Model(TextModel):
     norm_shift = 1.0  # Gemma3RMSNorm adds 1.0 to the norm value
 
     def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-        self.gguf_writer.add_add_space_prefix(False)
+        if (self.dir_model / "tokenizer.model").is_file():
+            self._set_vocab_sentencepiece()
+            self.gguf_writer.add_add_space_prefix(False)
+        else:
+            self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
         hparams = self.hparams
@@ -5757,13 +5847,24 @@ class Gemma3Model(TextModel):
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers
         # attn_logit_softcapping is removed in Gemma3
         assert hparams.get("attn_logit_softcapping") is None
-        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        if (final_logit_softcap := hparams.get("final_logit_softcapping")):
+            self.gguf_writer.add_final_logit_softcapping(final_logit_softcap)
+        if hparams.get("sliding_window_pattern") != 1:
+            self.gguf_writer.add_sliding_window(hparams["sliding_window"])
         self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
         if hparams.get("rope_scaling") is not None:
-            assert hparams["rope_scaling"]["rope_type"] == "linear"
-            # important: this rope_scaling is only applied for global layers, and not used by 1B model
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+            rope_scaling = hparams["rope_scaling"]
+            if rope_scaling["rope_type"] == "linear":
+                # important: this rope_scaling is only applied for global layers, and not used by 1B model
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            elif rope_scaling["rope_type"] == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+                self.gguf_writer.add_rope_scaling_yarn_ext_factor(rope_scaling["extrapolation_factor"])
+                self.gguf_writer.add_rope_scaling_yarn_beta_fast(rope_scaling["beta_fast"])
+                self.gguf_writer.add_rope_scaling_yarn_beta_slow(rope_scaling["beta_slow"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -5777,8 +5878,10 @@ class Gemma3Model(TextModel):
 
         # remove OOV (out-of-vocabulary) rows in token_embd
         if "embed_tokens.weight" in name:
-            vocab = self._create_vocab_sentencepiece()
-            tokens = vocab[0]
+            if (self.dir_model / "tokenizer.model").is_file():
+                tokens = self._create_vocab_sentencepiece()[0]
+            else:
+                tokens = self.get_vocab_base()[0]
             data_torch = data_torch[:len(tokens)]
 
         # ref code in Gemma3RMSNorm
@@ -9779,12 +9882,22 @@ class ApertusModel(LlamaModel):
 
 
 class MistralModel(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.LLAMA
+    model_arch = gguf.MODEL_ARCH.MISTRAL3
     model_name = "Mistral"
     hf_arch = ""
     is_mistral_format = True
     undo_permute = False
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # for compatibility, we use LLAMA arch for older models
+        # TODO: remove this once everyone migrates to newer version of llama.cpp
+        if "llama_4_scaling" not in self.hparams:
+            self.model_arch = gguf.MODEL_ARCH.LLAMA
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
+            self.gguf_writer.add_architecture()
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
     @staticmethod
     def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
         assert TokenizerVersion is not None and Tekkenizer is not None and SentencePieceTokenizer is not None, _mistral_import_error_msg
@@ -9824,6 +9937,112 @@ class MistralModel(LlamaModel):
 
         return template
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        MistralModel.set_mistral_config(self.gguf_writer, self.hparams)
+
+    @staticmethod
+    def set_mistral_config(gguf_writer: gguf.GGUFWriter, hparams: dict):
+        if "yarn" in hparams:
+            yarn_params = hparams["yarn"]
+            gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            gguf_writer.add_rope_scaling_factor(yarn_params["factor"])
+            gguf_writer.add_rope_scaling_yarn_beta_fast(yarn_params["beta"])
+            gguf_writer.add_rope_scaling_yarn_beta_slow(yarn_params["alpha"])
+            gguf_writer.add_rope_scaling_yarn_log_mul(1.0) # mscale_all_dim
+            gguf_writer.add_rope_scaling_orig_ctx_len(yarn_params["original_max_position_embeddings"])
+
+        if "llama_4_scaling" in hparams:
+            gguf_writer.add_attn_temperature_scale(hparams["llama_4_scaling"]["beta"])
+
+
+class MistralMoeModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+    model_name = "Mistral"
+    hf_arch = ""
+    is_mistral_format = True
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        logger.info("Using MistralMoeModel")
+        # remap hparams from Mistral MoE format to DeepseekV2 format
+        # we do this way to be able to reuse DeepseekV2Model set_gguf_parameters logic
+        # ref: https://github.com/vllm-project/vllm/blob/b294e28db2c5dee61bc25157664edcada8b90b31/vllm/transformers_utils/configs/mistral.py
+        config = self.hparams
+        # Mistral key -> HF key
+        config_mapping = {
+            "dim": "hidden_size",
+            "norm_eps": "rms_norm_eps",
+            "n_kv_heads": "num_key_value_heads",
+            "n_layers": "num_hidden_layers",
+            "n_heads": "num_attention_heads",
+            "hidden_dim": "intermediate_size",
+        }
+        # HF key -> (Mistral key, default value)
+        top_level_mapping_with_default = {
+            "model_type": ("model_type", "transformer"),
+            "hidden_act": ("activation", "silu"),
+            "tie_word_embeddings": ("tied_embeddings", False),
+            "max_seq_len": ("max_seq_len", config.get("max_position_embeddings", 128_000)),
+            "max_position_embeddings": ("max_position_embeddings", 128_000),
+        }
+        # mapping top-level keys
+        for key, new_key in config_mapping.items():
+            if key in config:
+                config[new_key] = config[key]
+        for new_key, (key, default_value) in top_level_mapping_with_default.items():
+            config[new_key] = config.get(key, default_value)
+        # mapping MoE-specific keys
+        moe_config_map = {
+            "route_every_n": "moe_layer_freq",
+            "first_k_dense_replace": "first_k_dense_replace",
+            "num_experts_per_tok": "num_experts_per_tok",
+            "num_experts": "n_routed_experts",
+            "expert_hidden_dim": "moe_intermediate_size",
+            "routed_scale": "routed_scaling_factor",
+            "num_shared_experts": "n_shared_experts",
+            "num_expert_groups": "n_group",
+            "num_expert_groups_per_tok": "topk_group",
+        }
+        moe = config["moe"]
+        for key, new_key in moe_config_map.items():
+            if key in moe:
+                config[new_key] = moe[key]
+        # provide missing values
+        config["topk_method"] = None
+        config["norm_topk_prob"] = True
+        config["scoring_func"] = "softmax"
+
+    def set_vocab(self):
+        self._set_vocab_mistral()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        MistralModel.set_mistral_config(self.gguf_writer, self.hparams)
+        yarn_params = self.hparams["yarn"]
+        self.gguf_writer.add_attn_temperature_length(yarn_params["original_max_position_embeddings"])
+        self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1) # mscale_all_dim * 0.1
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.startswith("vision_") or name.startswith("patch_merger.") or "mm_projector" in name:
+            return []
+
+        # rename certain tensors so that we can reuse DeepseekV2Model modify_tensors logic
+        if name.endswith(".qscale_act"):
+            name = name.replace(".qscale_act", ".input_scale")
+        if name.endswith(".qscale_weight"):
+            name = name.replace(".qscale_weight", ".weight_scale")
+        if ".wkv_b." in name:
+            name = name.replace(".wkv_b.", ".kv_b_proj.")
+        if ".experts." in name:
+            name = name.replace(".experts.", ".mlp.experts.")
+            name = name.replace(".w1.", ".gate_proj.")
+            name = name.replace(".w2.", ".down_proj.")
+            name = name.replace(".w3.", ".up_proj.")
+            name = "model." + name
+
+        return super().modify_tensors(data_torch, name, bid)
+
 
 class PixtralModel(LlavaVisionModel):
     model_name = "Pixtral"
@@ -10378,6 +10597,8 @@ def main() -> None:
         elif args.mmproj:
             assert hparams.get("vision_encoder") is not None, "This model does not support multimodal"
             model_class = PixtralModel
+        elif "moe" in hparams:
+            model_class = MistralMoeModel
         else:
             model_class = MistralModel
 

docs/backend/SYCL.md

@@ -42,6 +42,9 @@ 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|
@@ -789,6 +792,8 @@ 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
@@ -835,6 +840,14 @@ 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/backend/ZenDNN.md

@@ -0,0 +1,258 @@
+# llama.cpp for AMD ZenDNN
+
+> [!WARNING]
+> **Note:** ZenDNN is **not** the same as zDNN.
+> - **ZenDNN** (this page): AMD's deep learning library for AMD EPYC CPUs
+> - **zDNN**: IBM's Deep Neural Network acceleration library for IBM Z & LinuxONE Mainframes ([see zDNN documentation](zDNN.md))
+
+- [Background](#background)
+- [OS](#os)
+- [Hardware](#hardware)
+- [Supported Operations](#supported-operations)
+- [DataType Supports](#datatype-supports)
+- [Linux](#linux)
+- [Environment Variable](#environment-variable)
+- [Performance Optimization](#performance-optimization)
+- [Known Issues](#known-issues)
+- [TODO](#todo)
+
+## Background
+
+**ZenDNN** (Zen Deep Neural Network Library) is AMD's high-performance deep learning inference library optimized for AMD EPYC™ CPUs. It provides optimized implementations of key deep learning primitives and operations, delivering significant performance improvements for neural network workloads on AMD Zen-based processor architectures.
+
+**Llama.cpp + ZenDNN**
+
+The llama.cpp ZenDNN backend leverages AMD's optimized matrix multiplication primitives to accelerate inference on AMD CPUs. It utilizes ZenDNN's **LowOHA (Low Overhead Hardware Accelerated)** MatMul operator for efficient GEMM operations with minimal execution overhead, built-in weight caching, and direct access to backend libraries (AOCL BLIS, LibXSMM, OneDNN).
+
+For more information about ZenDNN, visit: https://www.amd.com/en/developer/zendnn.html
+
+## OS
+
+| OS      | Status  | Verified                                       |
+|:-------:|:-------:|:----------------------------------------------:|
+| Linux   | Support | Ubuntu 20.04, 22.04, 24.04                     |
+
+For the latest list of supported operating systems, see the [ZenDNN Supported OS](https://github.com/amd/ZenDNN/blob/zendnnl/README.md#15-supported-os).
+
+## Hardware
+
+### AMD CPUs
+
+**Recommended Processors**
+
+ZenDNN is optimized for AMD EPYC™ processors and AMD Ryzen™ processors based on "Zen" microarchitecture and newer.
+
+| CPU Family                    | Status  | Notes                              |
+|:-----------------------------:|:-------:|:----------------------------------:|
+| AMD EPYC™ 9005 Series (Turin)| Support | 5th Gen - Zen 5 architecture       |
+| AMD EPYC™ 9004 Series (Genoa)| Support | 4th Gen - Zen 4 architecture       |
+| AMD EPYC™ 7003 Series (Milan)| Support | 3rd Gen - Zen 3 architecture       |
+| AMD Ryzen™ AI MAX (Strix Halo)| Support | High-performance mobile processors |
+
+*Notes:*
+
+- Best performance is achieved on AMD EPYC™ processors with high core counts (e.g., EPYC 9005 series).
+- ZenDNN leverages AMD's advanced CPU features including AVX2 and AVX-512 instruction sets.
+- For optimal performance, ensure your system has sufficient memory bandwidth.
+
+## Supported Operations
+
+The ZenDNN backend currently accelerates **matrix multiplication (MUL_MAT)** operations only. Other operations are handled by the standard CPU backend.
+
+| Operation    | Status  | Notes                                          |
+|:-------------|:-------:|:----------------------------------------------:|
+| MUL_MAT      |    ✓    | Accelerated via ZenDNN LowOHA MatMul           |
+
+*Note:* Since only MUL_MAT is accelerated, models will benefit most from ZenDNN when matrix multiplications dominate the computational workload (which is typical for transformer-based LLMs).
+
+## DataType Supports
+
+| DataType               | Status  | Notes                                         |
+|:----------------------:|:-------:|:---------------------------------------------:|
+| FP32                   | Support | Full precision floating point                 |
+| BF16                   | Support | BFloat16 (best performance on Zen 4/Zen 5)    |
+
+*Notes:*
+
+- **BF16** provides best performance on Zen 4 and Zen 5 EPYC™ processors (Genoa, Turin).
+
+## Linux
+
+### I. Setup Environment
+
+You have two options to set up ZenDNN:
+
+#### Option 1: Automatic Download and Build (Recommended)
+
+CMake will automatically download and build ZenDNN for you:
+
+```sh
+# Build llama.cpp - ZenDNN will be automatically downloaded and built
+cmake -B build -DGGML_ZENDNN=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build build --config Release -j $(nproc)
+```
+
+No manual ZenDNN installation required. CMake will handle everything automatically.
+
+#### Option 2: Use Custom ZenDNN Installation
+
+If you want to build ZenDNN yourself or use a specific version:
+
+**Step 1: Build ZenDNN from source**
+
+```sh
+# Clone ZenDNN repository
+git clone https://github.com/amd/ZenDNN.git
+cd ZenDNN
+git checkout zendnnl
+
+# Build and install (requires CMake >= 3.25)
+mkdir build && cd build
+cmake ..
+cmake --build . --target all
+```
+
+Default installation path: `ZenDNN/build/install`
+
+**For detailed build instructions**, refer to the [ZenDNN README](https://github.com/amd/ZenDNN/blob/zendnnl/README.md).
+
+**Step 2: Build llama.cpp with custom ZenDNN path**
+
+```sh
+# Using environment variable
+export ZENDNN_ROOT=/path/to/ZenDNN/build/install
+cmake -B build -DGGML_ZENDNN=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build build --config Release -j $(nproc)
+
+# OR specify path directly in CMake
+cmake -B build -DGGML_ZENDNN=ON -DZENDNN_ROOT=/path/to/ZenDNN/build/install -DCMAKE_BUILD_TYPE=Release
+cmake --build build --config Release -j $(nproc)
+```
+
+### II. Run the Server
+
+#### 1. Download Model
+
+Download LLaMA 3.1 8B Instruct BF16 model:
+
+```sh
+# Download from Hugging Face
+huggingface-cli download meta-llama/Llama-3.1-8B-Instruct-GGUF --local-dir models/
+```
+
+#### 2. Start Server
+
+Run llama.cpp server with ZenDNN acceleration:
+
+```sh
+# Set optimal configuration
+export OMP_NUM_THREADS=64  # Adjust to your CPU core count
+export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS for best performance
+
+# Start server
+./build/bin/llama-server \
+    -m models/Llama-3.1-8B-Instruct.BF16.gguf \
+    --host 0.0.0.0 \
+    --port 8080 \
+    -t 64
+```
+
+Access the server at `http://localhost:8080`.
+
+**Performance tips**:
+- Set `OMP_NUM_THREADS` to match your physical core count
+- Use `ZENDNNL_MATMUL_ALGO=2` for optimal performance
+- For NUMA systems: `numactl --cpunodebind=0 --membind=0 ./build/bin/llama-server ...`
+
+## Environment Variable
+
+### Build Time
+
+| Name               | Value                                 | Function                                    |
+|--------------------|---------------------------------------|---------------------------------------------|
+| GGML_ZENDNN        | ON/OFF                                | Enable ZenDNN backend support               |
+| ZENDNN_ROOT        | Path to ZenDNN installation           | Set ZenDNN installation directory           |
+| GGML_OPENMP        | ON/OFF (recommended: ON)              | Enable OpenMP for multi-threading           |
+
+### Runtime
+
+| Name                    | Value                    | Function                                                          |
+|-------------------------|--------------------------|-------------------------------------------------------------------|
+| OMP_NUM_THREADS         | Number (e.g., 64)        | Set number of OpenMP threads (recommended: physical core count)   |
+| ZENDNNL_MATMUL_ALGO     | 0-5                      | Select MatMul backend algorithm (see Performance Optimization)    |
+| ZENDNNL_PROFILE_LOG_LEVEL | 0-4                    | Profiling log level (0=disabled, 4=verbose)                       |
+| ZENDNNL_ENABLE_PROFILER | 0 or 1                   | Enable detailed profiling (1=enabled)                             |
+| ZENDNNL_API_LOG_LEVEL   | 0-4                      | API log level (0=disabled, 4=verbose)                             |
+
+**Example**:
+
+```sh
+export OMP_NUM_THREADS=64
+export ZENDNNL_MATMUL_ALGO=2  # Use Blocked AOCL BLIS for best performance
+./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Test" -n 100
+```
+
+## Performance Optimization
+
+### MatMul Algorithm Selection
+
+ZenDNN's LowOHA MatMul supports multiple backend algorithms. For **best performance**, use the **Blocked AOCL BLIS** algorithm:
+
+```sh
+export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS (recommended)
+```
+
+**Available algorithms**:
+
+| Value | Algorithm              | Description                                    |
+|:-----:|:-----------------------|:----------------------------------------------|
+| 0     | Dynamic Dispatch       | Automatic backend selection (default)         |
+| 1     | AOCL BLIS              | AOCL BLIS backend                             |
+| 2     | AOCL BLIS Blocked      | **Blocked AOCL BLIS (recommended)**           |
+| 3     | OneDNN                 | OneDNN backend                                |
+| 4     | OneDNN Blocked         | Blocked OneDNN                                |
+| 5     | LibXSMM                | LibXSMM backend                               |
+
+### Profiling and Debugging
+
+For detailed profiling and logging options, refer to the [ZenDNN Logging Documentation](https://github.com/amd/ZenDNN/blob/zendnnl/docs/logging.md).
+
+## Known Issues
+
+- **Limited operation support**: Currently only matrix multiplication (MUL_MAT) is accelerated via ZenDNN. Other operations fall back to the standard CPU backend.
+- **BF16 support**: BF16 operations require AMD Zen 4 or Zen 5 architecture (EPYC 9004/9005 series). On older CPUs, operations will use FP32.
+- **NUMA awareness**: For multi-socket systems, manual NUMA binding may be required for optimal performance.
+
+## Q&A
+
+**Q: How do I verify that ZenDNN backend is being used?**
+
+A: Check the log output when running llama.cpp. You should see messages indicating the ZenDNN backend is initialized. You can also check the backend name in the output.
+
+**Q: What performance improvement can I expect?**
+
+A: Performance gains vary depending on the model size, batch size, and CPU architecture. On AMD EPYC processors, you can typically expect 1.1x-2x speedup compared to standard CPU inference for matrix multiplication operations.
+
+**Q: Can I use ZenDNN on non-AMD processors?**
+
+A: ZenDNN is optimized specifically for AMD processors. While it may work on other x86-64 CPUs, performance benefits are only guaranteed on AMD Zen-based architectures.
+
+**Q: Does ZenDNN support quantized models?**
+
+A: Currently, ZenDNN primarily supports FP32 and BF16 data types. Quantized model support is not available at this time.
+
+**Q: Why is my inference not faster with ZenDNN?**
+
+A: Ensure:
+1. You're using an AMD EPYC or Ryzen processor (Zen 2 or newer)
+2. `OMP_NUM_THREADS` is set appropriately (physical core count)
+3. `ZENDNNL_MATMUL_ALGO=2` is set for best performance (Blocked AOCL BLIS)
+4. You're using a sufficiently large model (small models may not benefit as much)
+5. Enable profiling to verify ZenDNN MatMul is being called
+
+### **GitHub Contribution**:
+Please add the **[ZenDNN]** prefix/tag in issues/PRs titles to help the ZenDNN-team check/address them without delay.
+
+## TODO
+
+- Expand operation support beyond MUL_MAT (attention operations, activations, etc.)

docs/backend/zDNN.md

@@ -1,5 +1,10 @@
 # llama.cpp for IBM zDNN Accelerator
 
+> [!WARNING]
+> **Note:** zDNN is **not** the same as ZenDNN.
+> - **zDNN** (this page): IBM's Deep Neural Network acceleration library for IBM Z & LinuxONE Mainframes
+> - **ZenDNN**: AMD's deep learning library for AMD EPYC CPUs ([see ZenDNN documentation](ZenDNN.md))
+
 ## Background
 
 IBM zDNN (Z Deep Neural Network) is a hardware acceleration library designed specifically to leverage the IBM NNPA (Neural Network Processor Assist) accelerator located within IBM Telum I and II processors. It provides significant performance improvements for neural network inference operations.

docs/build-riscv64-spacemit.md

@@ -19,6 +19,7 @@ cmake -B build \
     -DGGML_RVV=ON \
     -DGGML_RV_ZFH=ON \
     -DGGML_RV_ZICBOP=ON \
+    -DGGML_RV_ZIHINTPAUSE=ON \
     -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
     -DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake \
     -DCMAKE_INSTALL_PREFIX=build/installed

docs/build.md

@@ -431,11 +431,22 @@ docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/ren
 
 ### For Linux users:
 
+#### Using the LunarG Vulkan SDK
+
 First, follow the official LunarG instructions for the installation and setup of the Vulkan SDK in the [Getting Started with the Linux Tarball Vulkan SDK](https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html) guide.
 
 > [!IMPORTANT]
 > After completing the first step, ensure that you have used the `source` command on the `setup_env.sh` file inside of the Vulkan SDK in your current terminal session. Otherwise, the build won't work. Additionally, if you close out of your terminal, you must perform this step again if you intend to perform a build. However, there are ways to make this persistent. Refer to the Vulkan SDK guide linked in the first step for more information about any of this.
 
+#### Using system packages
+
+On Debian / Ubuntu, you can install the required dependencies using:
+```sh
+sudo apt-get install libvulkan-dev glslc
+```
+
+#### Common steps
+
 Second, after verifying that you have followed all of the SDK installation/setup steps, use this command to make sure before proceeding:
 ```bash
 vulkaninfo
@@ -484,6 +495,38 @@ llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB
 
 For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
 
+## ZenDNN
+
+ZenDNN provides optimized deep learning primitives for AMD EPYC™ CPUs. It accelerates matrix multiplication operations for inference workloads.
+
+### Compilation
+
+- Using `CMake` on Linux (automatic build):
+
+    ```bash
+    cmake -B build -DGGML_ZENDNN=ON
+    cmake --build build --config Release
+    ```
+
+    The first build will automatically download and build ZenDNN, which may take 5-10 minutes. Subsequent builds will be much faster.
+
+- Using `CMake` with custom ZenDNN installation:
+
+    ```bash
+    cmake -B build -DGGML_ZENDNN=ON -DZENDNN_ROOT=/path/to/zendnn/install
+    cmake --build build --config Release
+    ```
+
+### Testing
+
+You can test with:
+
+```bash
+./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -n 50
+```
+
+For detailed information about hardware support, setup instructions, and performance optimization, refer to [llama.cpp for ZenDNN](./backend/ZenDNN.md).
+
 ## Arm® KleidiAI™
 KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend.
 

docs/development/parsing.md

@@ -0,0 +1,288 @@
+# Parsing Model Output
+
+The `common` library contains a PEG parser implementation suitable for parsing
+model output.
+
+Types with the prefix `common_peg_*` are intended for general use and may have
+applications beyond parsing model output, such as parsing user-provided regex
+patterns.
+
+Types with the prefix `common_chat_peg_*` are specialized helpers for model
+output.
+
+The parser features:
+
+- Partial parsing of streaming input
+- Built-in JSON parsers
+- AST generation with semantics via "tagged" nodes
+
+## Example
+
+Below is a contrived example demonstrating how to use the PEG parser to parse
+output from a model that emits arguments as JSON.
+
+```cpp
+auto parser = build_chat_peg_native_parser([&](common_chat_peg_native_builder & p) {
+    // Build a choice of all available tools
+    auto tool_choice = p.choice();
+    for (const auto & tool : tools) {
+        const auto & function = tool.at("function");
+        std::string name = function.at("name");
+        const auto & schema = function.at("parameters");
+
+        auto tool_name = p.json_member("name", "\"" + p.literal(name) + "\"");
+        auto tool_args = p.json_member("arguments", p.schema(p.json(), "tool-" + name + "-schema", schema));
+
+        tool_choice |= p.rule("tool-" + name, "{" << tool_name << "," << tool_args << "}");
+    }
+
+    // Define the tool call structure: <tool_call>[{tool}]</tool_call>
+    auto tool_call = p.trigger_rule("tool-call",
+        p.sequence({
+            p.literal("<tool_call>["),
+            tool_choice,
+            p.literal("]</tool_call>")
+        })
+    );
+
+    // Parser accepts content, optionally followed by a tool call
+    return p.sequence({
+        p.content(p.until("<tool_call>")),
+        p.optional(tool_call),
+        p.end()
+    });
+});
+```
+
+For a more complete example, see `test_example_native()` in
+[tests/test-chat-peg-parser.cpp](tests/test-chat-peg-parser.cpp).
+
+## Parsers/Combinators
+
+### Basic Matchers
+
+- **`eps()`** - Matches nothing and always succeeds (epsilon/empty match)
+- **`start()`** - Matches the start of input (anchor `^`)
+- **`end()`** - Matches the end of input (anchor `$`)
+- **`literal(string)`** - Matches an exact literal string
+- **`any()`** - Matches any single character (`.`)
+
+### Combinators
+
+- **`sequence(...)`** - Matches parsers in order; all must succeed
+- **`choice(...)`** - Matches the first parser that succeeds from alternatives (ordered choice)
+- **`one_or_more(p)`** - Matches one or more repetitions (`+`)
+- **`zero_or_more(p)`** - Matches zero or more repetitions (`*`)
+- **`optional(p)`** - Matches zero or one occurrence (`?`)
+- **`repeat(p, min, max)`** - Matches between min and max repetitions (use `-1` for unbounded)
+- **`repeat(p, n)`** - Matches exactly n repetitions
+
+### Lookahead
+
+- **`peek(p)`** - Positive lookahead: succeeds if parser succeeds without consuming input (`&`)
+- **`negate(p)`** - Negative lookahead: succeeds if parser fails without consuming input (`!`)
+
+### Character Classes & Utilities
+
+- **`chars(classes, min, max)`** - Matches repetitions of characters from a character class
+- **`space()`** - Matches zero or more whitespace characters (space, tab, newline)
+- **`until(delimiter)`** - Matches characters until delimiter is found (delimiter not consumed)
+- **`until_one_of(delimiters)`** - Matches characters until any delimiter in the list is found
+- **`rest()`** - Matches everything remaining (`.*`)
+
+### JSON Parsers
+
+- **`json()`** - Complete JSON parser (objects, arrays, strings, numbers, booleans, null)
+- **`json_object()`** - JSON object parser
+- **`json_array()`** - JSON array parser
+- **`json_string()`** - JSON string parser
+- **`json_number()`** - JSON number parser
+- **`json_bool()`** - JSON boolean parser
+- **`json_null()`** - JSON null parser
+- **`json_string_content()`** - JSON string content without surrounding quotes
+- **`json_member(key, p)`** - JSON object member with specific key and value parser
+
+### Grammar Building
+
+- **`ref(name)`** - Creates a lightweight reference to a named rule (for recursive grammars)
+- **`rule(name, p, trigger)`** - Creates a named rule and returns a reference
+- **`trigger_rule(name, p)`** - Creates a trigger rule (entry point for lazy grammar generation)
+- **`schema(p, name, schema, raw)`** - Wraps parser with JSON schema metadata for grammar generation
+
+### AST Control
+
+- **`atomic(p)`** - Prevents AST node creation for partial parses
+- **`tag(tag, p)`** - Creates AST nodes with semantic tags (multiple nodes can share tags)
+
+## GBNF Grammar Generation
+
+The PEG parser also acts as a convenient DSL for generating GBNF grammars, with
+some exceptions.
+
+```cpp
+data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+    foreach_function(params.tools, [&](const json & fn) {
+        builder.resolve_refs(fn.at("parameters"));
+    });
+    parser.build_grammar(builder, data.grammar_lazy);
+});
+```
+
+The notable exception is the `negate(p)` lookahead parser, which cannot be
+defined as a CFG grammar and therefore does not produce a rule. Its usage
+should be limited and preferably hidden behind a `schema()` parser. In many
+cases, `until(delimiter)` or `until_one_of(delimiters)` is a better choice.
+
+Another limitation is that the PEG parser requires an unambiguous grammar. In
+contrast, the `llama-grammar` implementation can support ambiguous grammars,
+though they are difficult to parse.
+
+### Lazy Grammars
+
+During lazy grammar generation, only rules reachable from a `trigger_rule(p)`
+are emitted in the grammar. All trigger rules are added as alternations in the
+root rule. It is still necessary to define trigger patterns, as the parser has
+no interaction with the grammar sampling.
+
+### JSON Schema
+
+The `schema(p, name, schema, raw)` parser will use the `json-schema-to-grammar`
+implementation to generate the grammar instead of the underlying parser.
+
+The `raw` option emits a grammar suitable for a raw string instead of a JSON
+string. In other words, it won't be wrapped in quotes or require escaping
+quotes. It should only be used when `type == "string"`.
+
+The downside is that it can potentially lead to ambiguous grammars. For
+example, if a user provides the pattern `^.*$`, the following grammar may be
+generated:
+
+```
+root ::= "<arg>" .* "</arg>"
+```
+
+This creates an ambiguous grammar that cannot be parsed by the PEG parser. To
+help mitigate this, if `.*` is found in the pattern, the grammar from the
+underlying parser will be emitted instead.
+
+## Common AST Shapes for Chat Parsing
+
+Most model output can be placed in one of the following categories:
+
+- Content only
+- Tool calling with arguments emitted as a single JSON object
+- Tool calling with arguments emitted as separate entities, either XML
+  (Qwen3-Coder, MiniMax M2) or pseudo-function calls (LFM2)
+
+To provide broad coverage,
+[`common/chat-peg-parser.h`](common/chat-peg-parser.h) contains builders and
+mappers that help create parsers and visitors/extractors for these types. They
+require parsers to tag nodes to conform to an AST "shape". This normalization
+makes it easy to extract information and generalize parsing.
+
+### Simple
+
+The `common_chat_peg_builder` builds a `simple` parser that supports
+content-only models with optional reasoning.
+
+- **`reasoning(p)`** - Tag node for extracting `reasoning_content`
+- **`content(p)`** - Tag node for extracting `content`
+
+```cpp
+build_chat_peg_parser([&](common_chat_peg_parser & p) {
+    return p.sequence({
+        p.optional("<think>" + p.reasoning(p.until("</think>")) + "</think>"),
+        p.content(p.until("<tool_call>")),
+        p.end()
+    });
+});
+```
+
+Use `common_chat_peg_mapper` to extract the content. Note that this is already
+done for you in `common_chat_peg_parser` when
+`chat_format == COMMON_CHAT_FORMAT_PEG_SIMPLE`.
+
+```cpp
+auto result = parser.parse(ctx);
+
+common_chat_msg msg;
+auto mapper = common_chat_peg_mapper(msg);
+mapper.from_ast(ctx.ast, result);
+```
+
+### Native
+
+The `common_chat_peg_native_builder` builds a `native` parser suitable for
+models that emit tool arguments as a direct JSON object.
+
+- **`reasoning(p)`** - Tag node for `reasoning_content`
+- **`content(p)`** - Tag node for `content`
+- **`tool(p)`** - Tag entirety of a single tool call
+- **`tool_open(p)`** - Tag start of a tool call
+- **`tool_close(p)`** - Tag end of a tool call
+- **`tool_id(p)`** - Tag the tool call ID (optional)
+- **`tool_name(p)`** - Tag the tool name
+- **`tool_args(p)`** - Tag the tool arguments
+
+```cpp
+build_chat_peg_native_parser([&](common_chat_peg_native_parser & p) {
+    auto get_weather_tool = p.tool(p.sequence({
+        p.tool_open(p.literal("{")),
+        p.json_member("name", "\"" + p.tool_name(p.literal("get_weather")) + "\""),
+        p.literal(","),
+        p.json_member("arguments", p.tool_args(p.json())),
+        p.tool_close(p.literal("}"))
+    }));
+
+    return p.sequence({
+        p.content(p.until("<tool_call>")),
+        p.literal("<tool_call>"),
+        get_weather_tool,
+        p.literal("</tool_call>"),
+        p.end()
+    });
+});
+```
+
+### Constructed
+
+The `common_chat_peg_constructed_builder` builds a `constructed` parser
+suitable for models that emit tool arguments as separate entities, such as XML
+tags.
+
+- **`reasoning(p)`** - Tag node for `reasoning_content`
+- **`content(p)`** - Tag node for `content`
+- **`tool(p)`** - Tag entirety of a single tool call
+- **`tool_open(p)`** - Tag start of a tool call
+- **`tool_close(p)`** - Tag end of a tool call
+- **`tool_name(p)`** - Tag the tool name
+- **`tool_arg(p)`** - Tag a complete tool argument (name + value)
+- **`tool_arg_open(p)`** - Tag start of a tool argument
+- **`tool_arg_close(p)`** - Tag end of a tool argument
+- **`tool_arg_name(p)`** - Tag the argument name
+- **`tool_arg_string_value(p)`** - Tag string value for the argument
+- **`tool_arg_json_value(p)`** - Tag JSON value for the argument
+
+```cpp
+build_chat_peg_constructed_parser([&](common_chat_peg_constructed_builder & p) {
+    auto location_arg = p.tool_arg(
+        p.tool_arg_open("<parameter name=\"" + p.tool_arg_name(p.literal("location")) + "\">"),
+        p.tool_arg_string_value(p.until("</parameter>")),
+        p.tool_arg_close(p.literal("</parameter>"))
+    );
+
+    auto get_weather_tool = p.tool(p.sequence({
+        p.tool_open("<function name=\"" + p.tool_name(p.literal("get_weather")) + "\">"),
+        location_arg,
+        p.tool_close(p.literal("</function>"))
+    }));
+
+    return p.sequence({
+        p.content(p.until("<tool_call>")),
+        p.literal("<tool_call>"),
+        get_weather_tool,
+        p.literal("</tool_call>"),
+        p.end()
+    });
+});
+```

docs/ops.md

@@ -12,110 +12,111 @@ Legend:
 - 🟡 Partially supported by this backend
 - ❌ Not supported by this backend
 
-| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | zDNN |
-|-----------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                           CUMSUM | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                             FILL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
-|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ❌ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
-|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
-|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
-|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ❌ |
-|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
-|                        SOLVE_TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              SUM | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ❌ |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                              TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
-|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
+|-----------|------|------|------|------|------|------|------|------|------|------|------|
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                           CUMSUM | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ |
+|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SUM | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |

docs/ops/Vulkan.csv

@@ -5005,8 +5005,8 @@
 "Vulkan0","DUP","type=f16,ne=[10,10,5,1],permute=[0,2,1,3]","support","1","yes","Vulkan"
 "Vulkan0","DUP","type=f32,ne=[10,10,5,1],permute=[1,0,2,3]","support","1","yes","Vulkan"
 "Vulkan0","DUP","type=f16,ne=[10,10,5,1],permute=[1,0,2,3]","support","1","yes","Vulkan"
-"Vulkan0","DUP","type=i16,ne=[10,8,3,1],permute=[0,2,1,3]","support","0","no","Vulkan"
-"Vulkan0","DUP","type=i16,ne=[10,8,3,1],permute=[1,2,0,3]","support","0","no","Vulkan"
+"Vulkan0","DUP","type=i16,ne=[10,8,3,1],permute=[0,2,1,3]","support","1","yes","Vulkan"
+"Vulkan0","DUP","type=i16,ne=[10,8,3,1],permute=[1,2,0,3]","support","1","yes","Vulkan"
 "Vulkan0","SET","type_src=f32,type_dst=f32,ne=[6,5,4,3],dim=1","support","0","no","Vulkan"
 "Vulkan0","SET","type_src=f32,type_dst=f32,ne=[6,5,4,3],dim=2","support","0","no","Vulkan"
 "Vulkan0","SET","type_src=f32,type_dst=f32,ne=[6,5,4,3],dim=3","support","0","no","Vulkan"
@@ -5032,14 +5032,14 @@
 "Vulkan0","CPY","type_src=f16,type_dst=f16,ne=[3,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=f16,type_dst=f16,ne=[3,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","0","no","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","0","no","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","0","no","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=q4_0,type_dst=q4_0,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=q4_0,type_dst=q4_0,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
 "Vulkan0","CPY","type_src=q4_0,type_dst=q4_0,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3],_src_transpose=0","support","0","no","Vulkan"
@@ -5271,7 +5271,7 @@
 "Vulkan0","CPY","type_src=bf16,type_dst=f16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=f16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=q4_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=q4_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
 "Vulkan0","CPY","type_src=bf16,type_dst=q4_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=0","support","0","no","Vulkan"
@@ -5415,21 +5415,49 @@
 "Vulkan0","CPY","type_src=f16,type_dst=f16,ne=[256,4,3,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=f32,type_dst=f32,ne=[256,4,3,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=f32,type_dst=f32,ne=[256,4,3,3],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,3,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","0","no","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,3,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=f16,type_dst=f16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=f32,type_dst=f32,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
-"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","0","no","Vulkan"
+"Vulkan0","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
+"Vulkan0","CPY","type_src=i32,type_dst=i32,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","Vulkan"
+"Vulkan0","CPY","type_src=i32,type_dst=i32,ne=[256,1,4,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
 "Vulkan0","CPY","type_src=f32,type_dst=f32,ne=[256,1,4,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f32,ne=[10,10,10,1]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f32,ne=[2,1,1,1]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f32,ne=[2,1,3,5]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f32,ne=[2,3,5,7]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f16,ne=[2,1,1,1]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f16,ne=[2,1,3,5]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=f16,ne=[2,3,5,7]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=bf16,ne=[2,1,1,1]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=bf16,ne=[2,1,3,5]","support","1","yes","Vulkan"
-"Vulkan0","CONT","type=bf16,ne=[2,3,5,7]","support","0","no","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[2,1,1,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[2,1,3,5],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[2,3,5,7],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[1,4,4,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[1,8,17,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[10,10,10,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[2,1,1,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[2,1,3,5],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[2,3,5,7],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[1,4,4,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[1,8,17,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f32,ne=[10,10,10,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[2,1,1,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[2,1,3,5],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[2,3,5,7],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[1,4,4,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[1,8,17,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[10,10,10,1],use_view_slice=1","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[2,1,1,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[2,1,3,5],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[2,3,5,7],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[1,4,4,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[1,8,17,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=i32,ne=[10,10,10,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f16,ne=[2,1,1,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f16,ne=[2,1,3,5],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f16,ne=[2,3,5,7],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f16,ne=[1,4,4,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f16,ne=[1,8,17,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=f16,ne=[10,10,10,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=bf16,ne=[2,1,1,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=bf16,ne=[2,1,3,5],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=bf16,ne=[2,3,5,7],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=bf16,ne=[1,4,4,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=bf16,ne=[1,8,17,1],use_view_slice=0","support","1","yes","Vulkan"
+"Vulkan0","CONT","type=bf16,ne=[10,10,10,1],use_view_slice=0","support","1","yes","Vulkan"
 "Vulkan0","ADD","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","Vulkan"
 "Vulkan0","SUB","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","Vulkan"
 "Vulkan0","MUL","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","Vulkan"
@@ -5655,6 +5683,7 @@
 "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=[1024,1024,1,1]","support","1","yes","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"
@@ -8644,9 +8673,13 @@
 "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","FLOOR","type=f16,ne=[1024,1024,1,1]","support","1","yes","Vulkan"
 "Vulkan0","CEIL","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","CEIL","type=f16,ne=[1024,1024,1,1]","support","1","yes","Vulkan"
 "Vulkan0","ROUND","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","ROUND","type=f16,ne=[1024,1024,1,1]","support","1","yes","Vulkan"
 "Vulkan0","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","TRUNC","type=f16,ne=[1024,1024,1,1]","support","1","yes","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"
@@ -8666,9 +8699,13 @@
 "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","FLOOR","type=f32,ne=[1024,1024,1,1]","support","1","yes","Vulkan"
 "Vulkan0","CEIL","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","CEIL","type=f32,ne=[1024,1024,1,1]","support","1","yes","Vulkan"
 "Vulkan0","ROUND","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","ROUND","type=f32,ne=[1024,1024,1,1]","support","1","yes","Vulkan"
 "Vulkan0","TRUNC","type=f32,ne=[7,1,5,3]","support","1","yes","Vulkan"
+"Vulkan0","TRUNC","type=f32,ne=[1024,1024,1,1]","support","1","yes","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"
@@ -9411,28 +9448,405 @@
 "Vulkan0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","Vulkan"
 "Vulkan0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","Vulkan"
 "Vulkan0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[15,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[16,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[31,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[32,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[63,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[64,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[127,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[128,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[255,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[256,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[511,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[512,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1023,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2047,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2048,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[4095,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[4096,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[8191,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[8192,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[16383,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[32767,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[32768,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[65535,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[65536,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[131071,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[131072,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[262143,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[262144,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[524287,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[524288,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1048575,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1048576,1,1,1],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[16,10,10,10],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[60,10,10,10],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[1023,2,1,3],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[1024,2,1,3],order=0","support","1","yes","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[1025,2,1,3],order=0","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[2047,2,1,3],order=0","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[2048,2,1,3],order=0","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[2049,2,1,3],order=0","support","0","no","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1025,2,1,3],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2047,2,1,3],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2048,2,1,3],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2049,2,1,3],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[2,8,8192,1],order=0","support","1","yes","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[8,1,1,1],order=1","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[15,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[16,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[31,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[32,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[63,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[64,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[127,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[128,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[255,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[256,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[511,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[512,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1023,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2047,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2048,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[4095,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[4096,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[8191,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[8192,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[16383,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[32767,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[32768,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[65535,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[65536,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[131071,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[131072,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[262143,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[262144,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[524287,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[524288,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1048575,1,1,1],order=0","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1048576,1,1,1],order=0","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[16,10,10,10],order=1","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[60,10,10,10],order=1","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[1023,2,1,3],order=1","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[1024,2,1,3],order=1","support","1","yes","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[1025,2,1,3],order=1","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[16384,1,1,1],order=1","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[2047,2,1,3],order=1","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","0","no","Vulkan"
-"Vulkan0","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","0","no","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[1025,2,1,3],order=1","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2047,2,1,3],order=1","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","1","yes","Vulkan"
+"Vulkan0","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","1","yes","Vulkan"
 "Vulkan0","ARGSORT","type=f32,ne=[2,8,8192,1],order=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[12,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[13,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[13,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[15,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[15,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[15,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[19,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[19,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[19,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[19,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[27,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[27,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[27,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[27,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[27,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[43,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[43,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[43,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[43,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[43,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[64,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[75,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[64,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[75,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[64,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[75,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[64,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[75,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[64,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[75,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[128,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[139,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[128,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[139,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[128,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[139,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[128,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[139,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[128,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[139,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[128,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[139,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[256,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[267,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[256,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[267,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[256,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[267,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[256,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[267,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[256,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[267,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[256,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[267,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[512,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[523,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1035,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2059,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4096,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[4107,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8192,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[8203,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16395,1,2,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32768,1,1,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[32779,1,2,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65536,1,1,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[65547,1,2,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131072,1,1,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[131083,1,2,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262144,1,1,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[262155,1,2,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=100","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=500","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=1023","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524288,1,1,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[524299,1,2,1],k=9999","support","0","no","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,10,10,10],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[60,10,10,10],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1023,2,1,3],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,2,1,3],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1025,2,1,3],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2047,2,1,3],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,2,1,3],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2049,2,1,3],k=1","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,10,10,10],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[60,10,10,10],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1023,2,1,3],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,2,1,3],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1025,2,1,3],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2047,2,1,3],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,2,1,3],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2049,2,1,3],k=2","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,10,10,10],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[60,10,10,10],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1023,2,1,3],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,2,1,3],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1025,2,1,3],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2047,2,1,3],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,2,1,3],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2049,2,1,3],k=3","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,10,10,10],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[60,10,10,10],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1023,2,1,3],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,2,1,3],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1025,2,1,3],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2047,2,1,3],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,2,1,3],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2049,2,1,3],k=7","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16,10,10,10],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[60,10,10,10],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1023,2,1,3],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1024,2,1,3],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[1025,2,1,3],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[16384,1,1,1],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2047,2,1,3],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2048,2,1,3],k=15","support","1","yes","Vulkan"
+"Vulkan0","TOP_K","type=f32,ne=[2049,2,1,3],k=15","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest,flags=none","support","1","yes","Vulkan"
@@ -9445,6 +9859,10 @@
 "Vulkan0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=none","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic,flags=none","support","1","yes","Vulkan"
+"Vulkan0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=0","support","0","no","Vulkan"
+"Vulkan0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=1","support","0","no","Vulkan"
+"Vulkan0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","Vulkan"
+"Vulkan0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=align_corners","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear,flags=align_corners","support","1","yes","Vulkan"
 "Vulkan0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear,flags=align_corners","support","1","yes","Vulkan"
@@ -9479,23 +9897,37 @@
 "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=1048576.000000,step=1.000000","support","1","yes","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"
+"Vulkan0","CUMSUM","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[127,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[128,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[255,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[256,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[511,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[512,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[1023,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[1024,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[2047,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[2048,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[242004,1,1,1]","support","1","yes","Vulkan"
+"Vulkan0","CUMSUM","type=f32,ne=[375960,1,1,1]","support","1","yes","Vulkan"
 "Vulkan0","XIELU","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
-"Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","0","no","Vulkan"
-"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","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","1","yes","Vulkan"
+"Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","1","yes","Vulkan"
+"Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","1","yes","Vulkan"
+"Vulkan0","TRI","type=f32,ne=[10,10,4,3],tri_type=0","support","1","yes","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","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"
-"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","0","no","Vulkan"
-"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","0","no","Vulkan"
-"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","0","no","Vulkan"
+"Vulkan0","FILL","type=f32,ne=[2048,512,2,2],c=3.500000","support","1","yes","Vulkan"
+"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","1","yes","Vulkan"
+"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","1","yes","Vulkan"
+"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","1","yes","Vulkan"
+"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","1","yes","Vulkan"
+"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","1","yes","Vulkan"
+"Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","1","yes","Vulkan"
 "Vulkan0","SOLVE_TRI","type=f32,ne_lhs=[100,100,4,4],ne_rhs=[41,100,4,4]","support","0","no","Vulkan"
 "Vulkan0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0","support","1","yes","Vulkan"
 "Vulkan0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0","support","1","yes","Vulkan"

examples/CMakeLists.txt

@@ -20,6 +20,7 @@ else()
 
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf)
+    add_subdirectory(idle)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(parallel)

examples/embedding/embedding.cpp

@@ -104,12 +104,16 @@ 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
@@ -123,9 +127,6 @@ 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/idle/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-idle)
+add_executable(${TARGET} idle.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/idle/README.md

@@ -0,0 +1,3 @@
+# llama.cpp/example/idle
+
+https://github.com/ggml-org/llama.cpp/pull/17766

examples/idle/idle.cpp

@@ -0,0 +1,110 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <thread>
+#include <vector>
+
+static void print_usage(int /*argc*/, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers]\n", argv[0]);
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    // init LLM
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = common_model_params_to_llama(params);
+
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
+
+    if (model == NULL) {
+        LOG_ERR("%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    // we need just a dummy token to evaluate
+    std::vector<llama_token> prompt_tokens(1, llama_vocab_bos(vocab));
+
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx   = 512;
+    ctx_params.n_batch = 512;
+    ctx_params.no_perf = false;
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+
+    const int n_iters = 3;
+
+    // warm-up
+    llama_decode(ctx, batch);
+    llama_memory_clear(llama_get_memory(ctx), true);
+    llama_synchronize(ctx);
+
+    for (int64_t t_pause_ms = 0; t_pause_ms <= 4000; t_pause_ms += 800) {
+        double t_sum_us  = 0.0;
+        double t_sum2_us = 0.0;
+
+        for (int i = 0; i < n_iters; i++) {
+            // this pause is important - it simulates "idle GPU"
+            std::this_thread::sleep_for(std::chrono::milliseconds(t_pause_ms));
+
+            const int64_t t_start_us = llama_time_us();
+
+            // this should take constant time
+            llama_decode(ctx, batch);
+            llama_synchronize(ctx);
+
+            const int64_t t_end_us = llama_time_us();
+
+            const double t_cur_us = t_end_us - t_start_us;
+
+#if 1
+            // print individual decode times
+            printf("  - decode time: %8.2f ms\n", t_cur_us / 1000);
+#endif
+
+            t_sum_us  += t_cur_us;
+            t_sum2_us += t_cur_us * t_cur_us;
+
+            llama_memory_clear(llama_get_memory(ctx), true);
+            llama_synchronize(ctx); // just in case
+        }
+
+        const double t_avg_us = t_sum_us / n_iters;
+        const double t_dev_us = sqrt((t_sum2_us / (n_iters - 1)) - (t_avg_us * t_avg_us * n_iters) / (n_iters - 1));
+
+        printf("iters: %4d, pause: %5d ms, avg decode time: %8.2f +/- %4.2f ms\n", n_iters, (int) t_pause_ms, t_avg_us / 1000, t_dev_us / 1000);
+        fflush(stdout);
+    }
+
+    llama_free(ctx);
+    llama_model_free(model);
+
+    return 0;
+}

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/logits.cpp

@@ -144,7 +144,7 @@ int main(int argc, char ** argv) {
             return 1;
         }
         std::string s(buf, n);
-        printf("%s", s.c_str());
+        printf("%s (%d)", s.c_str(), id);
     }
     printf("\n");
 

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

@@ -4,6 +4,11 @@ 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 "$CONVERTED_MODEL" ]; then
@@ -14,7 +19,8 @@ if [ -z "$CONVERTED_MODEL" ]; then
 fi
 
 echo $CONVERTED_MODEL
+echo $MODEL_TESTING_PROMPT
 
 cmake --build ../../build --target llama-logits -j8
 
-../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "$MODEL_TESTING_PROMPT"

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

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

examples/save-load-state/save-load-state.cpp

@@ -241,6 +241,12 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch);
 
+    // this one is managed by common_init_result
+    //llama_free(ctx);
+
+    llama_free(ctx2);
+    llama_free(ctx3);
+
     if (result0 != result2) {
         fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);
         return 1;

examples/simple-cmake-pkg/README.md

@@ -18,6 +18,7 @@ cd llama.cpp
 cmake -S . -B build
 cmake --build build
 cmake --install build --prefix inst
+```
 
 ### Build simple-cmake-pkg
 

examples/sycl/run-llama2.sh

@@ -15,6 +15,9 @@ 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
@@ -18,11 +18,14 @@ 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} -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} -s 0 -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} -c ${CONTEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi

examples/sycl/win-run-llama2.bat

@@ -5,5 +5,7 @@
 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

@@ -5,5 +5,7 @@
 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 -e -ngl 99
+.\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

ggml/CMakeLists.txt

@@ -168,6 +168,7 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        ON)
 option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
 option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
+option(GGML_RV_ZIHINTPAUSE   "ggml: enable riscv zihintpause "  ON)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ${GGML_NATIVE})
 
@@ -175,14 +176,10 @@ option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requi
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
 set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
-
-if (MINGW)
-    set(GGML_WIN_VER "0xA00" CACHE STRING   "ggml: Windows version")
-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)
@@ -225,7 +222,7 @@ option(GGML_WEBGPU                          "ggml: use WebGPU"
 option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
 option(GGML_WEBGPU_CPU_PROFILE              "ggml: enable WebGPU profiling (CPU)"             OFF)
 option(GGML_WEBGPU_GPU_PROFILE              "ggml: enable WebGPU profiling (GPU)"             OFF)
-
+option(GGML_WEBGPU_JSPI                     "ggml: use JSPI for WebGPU"                       ON)
 option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -257,6 +254,9 @@ option(GGML_HEXAGON                         "ggml: enable Hexagon backend"
 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
 
+option(GGML_ZENDNN                          "ggml: use ZenDNN"                                OFF)
+option(ZENDNN_ROOT                          "ggml: path to ZenDNN installation"               "")
+
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
 option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
@@ -318,6 +318,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-sycl.h
     include/ggml-vulkan.h
     include/ggml-webgpu.h
+    include/ggml-zendnn.h
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
@@ -407,62 +408,67 @@ if (MSVC)
         /wd4996  # Disable POSIX deprecation warnings
         /wd4702  # Unreachable code warnings
     )
-    function(disable_msvc_warnings target_name)
+    set(MSVC_COMPILE_OPTIONS
+        "$<$<COMPILE_LANGUAGE:C>:/utf-8>"
+        "$<$<COMPILE_LANGUAGE:CXX>:/utf-8>"
+    )
+    function(configure_msvc_target target_name)
         if(TARGET ${target_name})
             target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+            target_compile_options(${target_name} PRIVATE ${MSVC_COMPILE_OPTIONS})
         endif()
     endfunction()
 
-    disable_msvc_warnings(ggml-base)
-    disable_msvc_warnings(ggml)
-    disable_msvc_warnings(ggml-cpu)
-    disable_msvc_warnings(ggml-cpu-x64)
-    disable_msvc_warnings(ggml-cpu-sse42)
-    disable_msvc_warnings(ggml-cpu-sandybridge)
-    disable_msvc_warnings(ggml-cpu-haswell)
-    disable_msvc_warnings(ggml-cpu-skylakex)
-    disable_msvc_warnings(ggml-cpu-icelake)
-    disable_msvc_warnings(ggml-cpu-alderlake)
+    configure_msvc_target(ggml-base)
+    configure_msvc_target(ggml)
+    configure_msvc_target(ggml-cpu)
+    configure_msvc_target(ggml-cpu-x64)
+    configure_msvc_target(ggml-cpu-sse42)
+    configure_msvc_target(ggml-cpu-sandybridge)
+    configure_msvc_target(ggml-cpu-haswell)
+    configure_msvc_target(ggml-cpu-skylakex)
+    configure_msvc_target(ggml-cpu-icelake)
+    configure_msvc_target(ggml-cpu-alderlake)
 
     if (GGML_BUILD_EXAMPLES)
-        disable_msvc_warnings(common-ggml)
-        disable_msvc_warnings(common)
+        configure_msvc_target(common-ggml)
+        configure_msvc_target(common)
 
-        disable_msvc_warnings(mnist-common)
-        disable_msvc_warnings(mnist-eval)
-        disable_msvc_warnings(mnist-train)
+        configure_msvc_target(mnist-common)
+        configure_msvc_target(mnist-eval)
+        configure_msvc_target(mnist-train)
 
-        disable_msvc_warnings(gpt-2-ctx)
-        disable_msvc_warnings(gpt-2-alloc)
-        disable_msvc_warnings(gpt-2-backend)
-        disable_msvc_warnings(gpt-2-sched)
-        disable_msvc_warnings(gpt-2-quantize)
-        disable_msvc_warnings(gpt-2-batched)
+        configure_msvc_target(gpt-2-ctx)
+        configure_msvc_target(gpt-2-alloc)
+        configure_msvc_target(gpt-2-backend)
+        configure_msvc_target(gpt-2-sched)
+        configure_msvc_target(gpt-2-quantize)
+        configure_msvc_target(gpt-2-batched)
 
-        disable_msvc_warnings(gpt-j)
-        disable_msvc_warnings(gpt-j-quantize)
+        configure_msvc_target(gpt-j)
+        configure_msvc_target(gpt-j-quantize)
 
-        disable_msvc_warnings(magika)
-        disable_msvc_warnings(yolov3-tiny)
-        disable_msvc_warnings(sam)
+        configure_msvc_target(magika)
+        configure_msvc_target(yolov3-tiny)
+        configure_msvc_target(sam)
 
-        disable_msvc_warnings(simple-ctx)
-        disable_msvc_warnings(simple-backend)
+        configure_msvc_target(simple-ctx)
+        configure_msvc_target(simple-backend)
     endif()
 
     if (GGML_BUILD_TESTS)
-        disable_msvc_warnings(test-mul-mat)
-        disable_msvc_warnings(test-arange)
-        disable_msvc_warnings(test-backend-ops)
-        disable_msvc_warnings(test-cont)
-        disable_msvc_warnings(test-conv-transpose)
-        disable_msvc_warnings(test-conv-transpose-1d)
-        disable_msvc_warnings(test-conv1d)
-        disable_msvc_warnings(test-conv2d)
-        disable_msvc_warnings(test-conv2d-dw)
-        disable_msvc_warnings(test-customop)
-        disable_msvc_warnings(test-dup)
-        disable_msvc_warnings(test-opt)
-        disable_msvc_warnings(test-pool)
+        configure_msvc_target(test-mul-mat)
+        configure_msvc_target(test-arange)
+        configure_msvc_target(test-backend-ops)
+        configure_msvc_target(test-cont)
+        configure_msvc_target(test-conv-transpose)
+        configure_msvc_target(test-conv-transpose-1d)
+        configure_msvc_target(test-conv1d)
+        configure_msvc_target(test-conv2d)
+        configure_msvc_target(test-conv2d-dw)
+        configure_msvc_target(test-customop)
+        configure_msvc_target(test-dup)
+        configure_msvc_target(test-opt)
+        configure_msvc_target(test-pool)
     endif ()
 endif()

ggml/include/ggml-rpc.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include "ggml.h"
 #include "ggml-backend.h"
 
 #ifdef  __cplusplus
@@ -8,7 +7,7 @@ extern "C" {
 #endif
 
 #define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    6
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 

ggml/include/ggml-zendnn.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_BACKEND_API ggml_backend_t ggml_backend_zendnn_init(void);
+
+GGML_BACKEND_API bool ggml_backend_is_zendnn(ggml_backend_t backend);
+
+// number of threads used for zendnn operations
+GGML_BACKEND_API void ggml_backend_zendnn_set_n_threads(ggml_backend_t backend_zendnn, int n_threads);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zendnn_reg(void);
+
+#ifdef __cplusplus
+}
+#endif

ggml/include/ggml.h

@@ -204,6 +204,10 @@
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#    define _WIN32_WINNT 0x0A00
+#endif
+
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -2148,7 +2152,8 @@ extern "C" {
     };
 
     enum ggml_scale_flag {
-        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
+        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8),
+        GGML_SCALE_FLAG_ANTIALIAS     = (1 << 9),
     };
 
     // interpolate
@@ -2191,6 +2196,15 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // pad each dimension with values on the other side of the torus (looping around)
+    GGML_API struct ggml_tensor * ggml_pad_circular(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   p0,
+            int                   p1,
+            int                   p2,
+            int                   p3);
+
     GGML_API struct ggml_tensor * ggml_pad_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2204,6 +2218,19 @@ extern "C" {
             int                  rp3
             );
 
+    // pad each dimension with values on the other side of the torus (looping around)
+    GGML_API struct ggml_tensor * ggml_pad_ext_circular(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   lp0,
+            int                   rp0,
+            int                   lp1,
+            int                   rp1,
+            int                   lp2,
+            int                   rp2,
+            int                   lp3,
+            int                   rp3);
+
     // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
     GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
             struct ggml_context * ctx,
@@ -2278,7 +2305,7 @@ extern "C" {
             float                 stop,
             float                 step);
 
-#define GGML_KQ_MASK_PAD 64
+#define GGML_KQ_MASK_PAD 1
 
     // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
     // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]

ggml/src/CMakeLists.txt

@@ -127,10 +127,6 @@ if (NOT MSVC)
     endif()
 endif()
 
-if (MINGW)
-    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
-endif()
-
 #
 # POSIX conformance
 #
@@ -221,6 +217,10 @@ 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)
@@ -270,10 +270,13 @@ function(ggml_add_backend_library backend)
     endif()
 
     # Set versioning properties for all backend libraries
-    set_target_properties(${backend} PROPERTIES
-        VERSION ${GGML_VERSION}
-        SOVERSION ${GGML_VERSION_MAJOR}
-    )
+    # Building a MODULE library with a version is not supported on macOS (https://gitlab.kitware.com/cmake/cmake/-/issues/20782)
+    if (NOT (APPLE AND GGML_BACKEND_DL))
+        set_target_properties(${backend} PROPERTIES
+            VERSION ${GGML_VERSION}
+            SOVERSION ${GGML_VERSION_MAJOR}
+        )
+    endif()
 
     if(NOT GGML_AVAILABLE_BACKENDS)
         set(GGML_AVAILABLE_BACKENDS "${backend}"
@@ -437,6 +440,7 @@ ggml_add_backend(WebGPU)
 ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
 ggml_add_backend(Hexagon)
+ggml_add_backend(ZenDNN)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-alloc.c

@@ -921,10 +921,15 @@ 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;
-            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;
+                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);
+                }
+            }
 #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-reg.cpp

@@ -73,6 +73,10 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_ZENDNN
+#include "ggml-zendnn.h"
+#endif
+
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
@@ -203,6 +207,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
+#ifdef GGML_USE_ZENDNN
+        register_backend(ggml_backend_zendnn_reg());
+#endif
 #ifdef GGML_USE_HEXAGON
         register_backend(ggml_backend_hexagon_reg());
 #endif
@@ -534,8 +541,12 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
     fs::path best_path;
 
     for (const auto & search_path : search_paths) {
-        if (!fs::exists(search_path)) {
-            GGML_LOG_DEBUG("%s: search path %s does not exist\n", __func__, path_str(search_path).c_str());
+        if (std::error_code ec; !fs::exists(search_path, ec)) {
+            if (ec) {
+                GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
+            } else {
+                GGML_LOG_DEBUG("%s: search path %s does not exist\n", __func__, path_str(search_path).c_str());
+            }
             continue;
         }
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
@@ -575,8 +586,12 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         for (const auto & search_path : search_paths) {
             fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native();
             fs::path path = search_path / filename;
-            if (fs::exists(path)) {
+            if (std::error_code ec; fs::exists(path, ec)) {
                 return get_reg().load_backend(path, silent);
+            } else {
+                if (ec) {
+                    GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(path).c_str(), ec.message().c_str());
+                }
             }
         }
         return nullptr;
@@ -597,6 +612,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
 #endif
 
     ggml_backend_load_best("blas", silent, dir_path);
+    ggml_backend_load_best("zendnn", silent, dir_path);
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);

ggml/src/ggml-backend.cpp

@@ -723,6 +723,12 @@ struct ggml_backend_sched {
     bool op_offload;
 
     int debug;
+
+    // used for debugging graph reallocations [GGML_SCHED_DEBUG_REALLOC]
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17617
+    int debug_realloc;
+    int debug_graph_size;
+    int debug_prev_graph_size;
 };
 
 #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
@@ -1234,10 +1240,8 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                                 tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
                                 ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
                             }
-                            if (sched->n_copies > 1) {
-                                ggml_set_input(tensor_copy);
-                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
-                            }
+                            ggml_set_input(tensor_copy);
+                            ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
@@ -1289,6 +1293,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
     }
 
     int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+
+    // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
+    sched->debug_prev_graph_size = sched->debug_graph_size;
+    sched->debug_graph_size = graph_size;
+
     if (sched->graph.size < graph_size) {
         sched->graph.size = graph_size;
         sched->graph.nodes = (ggml_tensor **) realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
@@ -1395,14 +1404,27 @@ 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)) {
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
+
+        if (sched->debug_realloc > 0) {
+            // we are interested only in situations where the graph was reallocated even though its size remained the same [GGML_SCHED_DEBUG_REALLOC]
+            // example: https://github.com/ggml-org/llama.cpp/pull/17143
+            const bool unexpected = !backend_ids_changed && sched->debug_prev_graph_size == sched->debug_graph_size;
+
+            if (unexpected || sched->debug_realloc > 1) {
+                GGML_ABORT("%s: unexpected graph reallocation (graph size = %d, nodes = %d, leafs = %d), debug_realloc = %d\n", __func__,
+                        sched->debug_graph_size, sched->graph.n_nodes, sched->graph.n_leafs, sched->debug_realloc);
+            }
+        }
+
         // 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__);
@@ -1614,6 +1636,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
 
     const char * GGML_SCHED_DEBUG = getenv("GGML_SCHED_DEBUG");
     sched->debug = GGML_SCHED_DEBUG ? atoi(GGML_SCHED_DEBUG) : 0;
+
+    sched->debug_realloc = 0;
+#ifdef GGML_SCHED_NO_REALLOC
+    sched->debug_realloc = 1;
+#endif
+    const char * GGML_SCHED_DEBUG_REALLOC = getenv("GGML_SCHED_DEBUG_REALLOC");
+    sched->debug_realloc = GGML_SCHED_DEBUG_REALLOC ? atoi(GGML_SCHED_DEBUG_REALLOC) : sched->debug_realloc;
+
     sched->n_backends = n_backends;
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
@@ -1630,6 +1660,9 @@ ggml_backend_sched_t ggml_backend_sched_new(
     sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
     sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
+    sched->debug_graph_size = 0;
+    sched->debug_prev_graph_size = 0;
+
     sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
     sched->context_buffer = (char *) malloc(sched->context_buffer_size);
 

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

@@ -2500,6 +2500,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 if (op->op_params[0] != GGML_SCALE_MODE_NEAREST) {
                     return false;
                 }
+                if (op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS) {
+                    return false;
+                }
                 return true;
             }
         case GGML_OP_POOL_2D:
@@ -2548,6 +2551,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_ACC:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_PAD:
+            // TODO: add circular padding support for cann, see https://github.com/ggml-org/llama.cpp/pull/16985
+            return ggml_get_op_params_i32(op, 8) == 0;
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
@@ -2561,6 +2566,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
             return true;
         case GGML_OP_OUT_PROD:
             {
+#ifdef ASCEND_310P
+                // Ger is not supported on 310p device
+                return false;
+#endif
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F16:
                     case GGML_TYPE_F32:

ggml/src/ggml-cpu/CMakeLists.txt

@@ -469,6 +469,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_RV_ZICBOP)
                 string(APPEND MARCH_STR "_zicbop")
             endif()
+            if (GGML_RV_ZIHINTPAUSE)
+                string(APPEND MARCH_STR "_zihintpause")
+            endif()
             list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
         else()
             # Begin with the lowest baseline

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

@@ -8,6 +8,10 @@
 #include <sys/sysctl.h>
 #endif
 
+#if !defined(HWCAP2_SVE2)
+#define HWCAP2_SVE2 (1 << 1)
+#endif
+
 #if !defined(HWCAP2_I8MM)
 #define HWCAP2_I8MM (1 << 13)
 #endif

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

@@ -505,7 +505,6 @@ void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     constexpr int blocklen          = 8;
 
     assert(n % qk == 0);
-    assert(nr % 4 == 0);
     assert(nc % ncols_interleaved == 0);
 
     UNUSED(nb);
@@ -645,7 +644,6 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
     constexpr int blocklen          = 8;
 
     assert(n % qk == 0);
-    assert(nr % 4 == 0);
     assert(nc % ncols_interleaved == 0);
 
     UNUSED(nb);

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

@@ -1,20 +1,23 @@
 #include "ggml-backend-impl.h"
 
 #if defined(__riscv) && __riscv_xlen == 64
-#include <sys/auxv.h>
-
-//https://github.com/torvalds/linux/blob/master/arch/riscv/include/uapi/asm/hwcap.h#L24
-#ifndef COMPAT_HWCAP_ISA_V
-#define COMPAT_HWCAP_ISA_V (1 << ('V' - 'A'))
-#endif
+#include <asm/hwprobe.h>
+#include <asm/unistd.h>
+#include <unistd.h>
 
 struct riscv64_features {
     bool has_rvv = false;
 
     riscv64_features() {
-        uint32_t hwcap = getauxval(AT_HWCAP);
+        struct riscv_hwprobe probe;
+        probe.key = RISCV_HWPROBE_KEY_IMA_EXT_0;
+        probe.value = 0;
 
-        has_rvv = !!(hwcap & COMPAT_HWCAP_ISA_V);
+        int ret = syscall(__NR_riscv_hwprobe, &probe, 1, 0, NULL, 0);
+
+        if (0 == ret) {
+            has_rvv = !!(probe.value & RISCV_HWPROBE_IMA_V);
+        }
     }
 };
 

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

@@ -490,6 +490,15 @@ static inline void ggml_thread_cpu_relax(void) {
 static inline void ggml_thread_cpu_relax(void) {
     _mm_pause();
 }
+#elif defined(__riscv)
+static inline void ggml_thread_cpu_relax(void) {
+    #ifdef __riscv_zihintpause
+        __asm__ __volatile__ ("pause");
+    #else
+        /* Encoding of the pause instruction */
+        __asm__ __volatile__ (".4byte 0x100000F");
+    #endif
+}
 #else
 static inline void ggml_thread_cpu_relax(void) {;}
 #endif
@@ -683,22 +692,14 @@ bool ggml_is_numa(void) {
 }
 
 #if defined(__ARM_ARCH)
-
-#if defined(__linux__) && defined(__aarch64__)
-#include <sys/auxv.h>
-#endif
-
-static void ggml_init_arm_arch_features(void) {
 #if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
-#if defined(__linux__)
-    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-#else
-    // TODO: add support of SVE for non-linux systems
-#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
-#endif
-#endif
+#include <arm_sve.h>
+static void ggml_init_arm_arch_features(void) {
+    ggml_arm_arch_features.sve_cnt = svcntb();
 }
-
+#else
+static void ggml_init_arm_arch_features(void) {}
+#endif
 #endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
@@ -2706,6 +2707,11 @@ struct ggml_cplan ggml_graph_plan(
         n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS;
     }
 
+#if defined(__EMSCRIPTEN__) && !defined(__EMSCRIPTEN_PTHREADS__)
+    // Emscripten without pthreads support can only use a single thread
+    n_threads = 1;
+#endif
+
     size_t work_size = 0;
 
     struct ggml_cplan cplan;

ggml/src/ggml-cpu/llamafile/sgemm-ppc.h

@@ -0,0 +1,333 @@
+#pragma once
+
+typedef vector unsigned char vec_t;
+typedef __vector_quad acc_t;
+
+template <typename TA>
+class tinyBLAS_Q0_PPC {
+  public:
+    tinyBLAS_Q0_PPC(int64_t k,
+                    const TA *A, int64_t lda,
+                    const block_q8_0 *B, int64_t ldb,
+                    float *C, int64_t ldc,
+                    int ith, int nth);
+
+    void matmul(int64_t m, int64_t n);
+    void matmul_tiled_q0(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        vec_t A_pack[mc*kc*2];
+        vec_t B_pack[nc*kc*2];
+        int comparray[mc*kc];
+        constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        int64_t tiles  = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles) {
+            end = tiles;
+        }
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                if constexpr(is_Ablock_q4) {
+                    packNormalInt4_large(A + ii*lda + kk, lda, mc, 4, (int8_t*)A_pack, comparray);
+                } else {
+                    packNormal_large<int8_t, vector signed char>(A + ii*lda + kk, lda, mc, 8, (int8_t*)A_pack, false, comparray);
+                }
+                packNormal_large<uint8_t, vector unsigned char>(B + jj*ldb + kk, ldb, nc, 8, (uint8_t*)B_pack, true);
+                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack, comparray);
+            }
+        }
+    }
+
+  private:
+    inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
+        for (int I = 0; I < RM; I++) {
+            for (int J = 0; J < RN; J++) {
+                *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
+            }
+        }
+    }
+
+    inline void add_save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
+        for (int I = 0; I < RM; I++) {
+            for (int J = 0; J < RN; J++) {
+                float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
+                *c_ptr += *((float*)&fin_res[idx+I]+J);
+            }
+        }
+    }
+
+    template<typename ArrayType>
+    inline void compute(acc_t* ACC, int c_idx, int s_idx, ArrayType& comparray, vector float* vs, vector float* fin_res) {
+        vector signed int vec_C[4];
+        vector float CA[4] = {0};
+        vector float res[4] = {0};
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int i = 0; i < 4; i++) {
+            CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
+            res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
+            fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
+        }
+    }
+
+    inline void process_q4_elements(vector signed char (&c)[2], int* ca) {
+        const vector signed char lowMask = vec_splats((signed char)0xF);
+        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+        const vector signed char v8 = vec_splats((signed char)0x8);
+        vector signed int vsum = {0};
+        vector signed int vsum2 = {0};
+        c[0] = vec_and(c[1], lowMask);
+        c[1] = vec_sr(c[1], v4);
+        c[0] = vec_sub(c[0], v8);
+        c[1] = vec_sub(c[1], v8);
+        vsum = vec_sum4s(c[0], vsum);
+        vsum2 = vec_sum4s(c[1], vsum2);
+        vsum = vec_add(vsum, vsum2);
+        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+    }
+
+    template <typename V1, typename V2>
+    inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
+        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
+        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
+        V2 t1, t2, t3, t4, t5, t6, t7, t8;
+        vector unsigned char xor_vector;
+        uint8_t flip_vec = 0x80;
+        xor_vector = vec_splats(flip_vec);
+        t1 = vec_perm(s1, s2, swiz1);
+        t2 = vec_perm(s1, s2, swiz2);
+        t3 = vec_perm(s3, s4, swiz1);
+        t4 = vec_perm(s3, s4, swiz2);
+        t5 = vec_perm(t1, t3, swiz3);
+        t6 = vec_perm(t1, t3, swiz4);
+        t7 = vec_perm(t2, t4, swiz3);
+        t8 = vec_perm(t2, t4, swiz4);
+        if (flip == true) {
+            t5 = vec_xor(t5, xor_vector);
+            t6 = vec_xor(t6, xor_vector);
+            t7 = vec_xor(t7, xor_vector);
+            t8 = vec_xor(t8, xor_vector);
+        }
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset+16);
+        vec_xst(t7, 0, vecOffset+32);
+        vec_xst(t8, 0, vecOffset+48);
+    }
+
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+        if constexpr(RM == 4 && RN == 8) {
+            KERNEL_4x8(ii,jj);
+        } else if constexpr(RM == 8 && RN == 4) {
+            KERNEL_8x4(ii,jj);
+        } else if constexpr(RM == 8 && RN == 8) {
+            KERNEL_8x8(ii,jj);
+        } else {
+            assert(false && "RN/RM values not supported");
+        }
+    }
+    template<int size>
+    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray);
+    template<typename VA, typename VB>
+    void packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip);
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n);
+    void KERNEL_4x8(int64_t ii, int64_t jj);
+    void KERNEL_8x4(int64_t ii, int64_t jj);
+    void KERNEL_8x8(int64_t ii, int64_t jj);
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN);
+    template <int RM, int RN>
+    void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n);
+
+    void compute_scale(int64_t ii, int64_t jj, int blk, vector float* vs){
+        for (int I = 0; I<8; I++) {
+            float a_scale = unhalf((A+((ii+I)*lda)+blk)->d);
+            for (int J = 0; J<4; J++) {
+                *((float*)&vs[I]+J) = (a_scale * unhalf((B+((jj+J)*ldb)+blk)->d));
+                *((float*)&vs[I+8]+J) = (a_scale * unhalf((B+((jj+J+4)*ldb)+blk)->d));
+             }
+         }
+    }
+
+    inline void process_q8_elements(const int8_t *qs, int *ca) {
+        vector signed char c1 = vec_xl(0, qs);
+        vector signed char c2 = vec_xl(16, qs);
+        vector signed int vsum1 = {0};
+        vector signed int vsum2 = {0};
+        vsum1 = vec_sum4s(c1, vsum1);
+        vsum2 = vec_sum4s(c2, vsum2);
+        vector signed int vsum = vec_add(vsum1, vsum2);
+        *ca = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+    }
+
+    template<typename VA, typename VB>
+    void packNormal_large(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip, int* comparray=nullptr) {
+        int64_t i, j;
+        block_q8_0 *aoffset = NULL;
+        VA *vecOffset = NULL;
+        block_q8_0* aoffsets[8];
+        __vector_pair arr[8];
+        VB c[8][2] = {0};
+        VB c1[8] = {0}; VB c2[8] = {0};
+        aoffset = const_cast<block_q8_0*>(a);
+        vecOffset = vec;
+        j = (rows >> 3);
+        int index = 0;
+        if (j > 0) {
+            do {
+                for (int it = 0; it < 8; it++)
+                    aoffsets[it] = aoffset + it*lda;
+                aoffset += 8 * lda;
+                for (int blk = 0; blk < kc; blk++) {
+                    for (int it = 0; it < 8; it++) {
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)(aoffsets[it]+blk)->qs);
+                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        c1[it] = c[it][0];
+                        c2[it] = c[it][1];
+                        if (comparray){
+                            process_q8_elements((aoffsets[it]+ blk)->qs, &comparray[index + 8*blk + it]);
+                        }
+                    }
+                    vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
+                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
+                    vecOffset += 256;
+                }
+                j--;
+                index += 8*kc;
+            } while(j > 0);
+        }
+
+    }
+
+    void packNormalInt4_large(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, int*comparray) {
+        int64_t i, j;
+        TA *aoffset = NULL;
+        int8_t *vecOffset = NULL;
+        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
+        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+        vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
+        vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
+        aoffset = const_cast<TA*>(a);
+        vecOffset = vec;
+        int index = 0;
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;
+                for (int blk = 0; blk < kc; blk++) {
+                    c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset1+blk)->qs));
+                    c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset2+blk)->qs));
+                    c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset3+blk)->qs));
+                    c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset4+blk)->qs));
+                    c5[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset5+blk)->qs));
+                    c6[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset6+blk)->qs));
+                    c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset7+blk)->qs));
+                    c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset8+blk)->qs));
+
+                    process_q4_elements(c1, &comparray[index + 8*blk+0]);
+                    process_q4_elements(c2, &comparray[index + 8*blk+1]);
+                    process_q4_elements(c3, &comparray[index + 8*blk+2]);
+                    process_q4_elements(c4, &comparray[index + 8*blk+3]);
+                    process_q4_elements(c5, &comparray[index + 8*blk+4]);
+                    process_q4_elements(c6, &comparray[index + 8*blk+5]);
+                    process_q4_elements(c7, &comparray[index + 8*blk+6]);
+                    process_q4_elements(c8, &comparray[index + 8*blk+7]);
+                    vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
+                    vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
+                    vecOffset += 256;
+                }
+                j--;
+                index += 8*kc;
+            } while (j > 0);
+        }
+    }
+
+    void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t *vec_A, vec_t *vec_B, int *comparray) {
+        acc_t acc[8];
+        for (int i = 0; i < mc ; i += 8) {
+            for (int j = 0; j < nc; j += 8) {
+                vector float fin_res[16] = {0};
+                vector float vs[16] = {0};
+                for (int64_t kk = 0; kk < kc; kk+=2) {
+                    for (int x = 0; x < 8; x++) {
+                        __builtin_mma_xxsetaccz(&acc[x]);
+                    }
+                    int A_block_idx = (i/8)*(16*kc) + kk*16;
+                    int B_block_idx = (j/8)*(16*kc)+ kk*16;
+                    vec_t *A_block = &vec_A[A_block_idx];
+                    vec_t *B_block = &vec_B[B_block_idx];
+                    for (int x = 0; x < 8; x++) {
+                        __builtin_mma_xvi8ger4pp(&acc[0], A_block[x],     B_block[x]);
+                        __builtin_mma_xvi8ger4pp(&acc[1], A_block[x + 8], B_block[x]);
+                        __builtin_mma_xvi8ger4pp(&acc[2], A_block[x],     B_block[x+8]);
+                        __builtin_mma_xvi8ger4pp(&acc[3], A_block[x+8],   B_block[x+8]);
+                    }
+                    compute_scale(ii+i, jj+j, l+kk, vs);
+                    int c_index = (i/8)*(8*kc)+ kk*8;
+                    int* c_block = &comparray[c_index];
+                    compute(&acc[0], 0,  0,  c_block, vs, fin_res);
+                    compute(&acc[1], 4,  4,  c_block, vs, fin_res);
+                    compute(&acc[2], 0,  8,  c_block, vs, fin_res);
+                    compute(&acc[3], 4, 12,  c_block, vs, fin_res);
+
+                    A_block_idx = (i/8)*(16*kc) + (kk+1)*16;
+                    B_block_idx = (j/8)*(16*kc)+ (kk+1)*16;
+                    A_block = &vec_A[A_block_idx];
+                    B_block = &vec_B[B_block_idx];
+                    for (int x = 0; x < 8; x++) {
+                        __builtin_mma_xvi8ger4pp(&acc[4], A_block[x],     B_block[x]);
+                        __builtin_mma_xvi8ger4pp(&acc[5], A_block[x + 8], B_block[x]);
+                        __builtin_mma_xvi8ger4pp(&acc[6], A_block[x],     B_block[x+8]);
+                        __builtin_mma_xvi8ger4pp(&acc[7], A_block[x+8],   B_block[x+8]);
+                    }
+                    compute_scale(ii+i, jj+j, l+kk+1, vs);
+                    c_index = (i/8)*(8*kc)+ (kk+1)*8;
+                    c_block = &comparray[c_index];
+                    compute(&acc[4], 0,  0,  c_block, vs, fin_res);
+                    compute(&acc[5], 4,  4,  c_block, vs, fin_res);
+                    compute(&acc[6], 0,  8,  c_block, vs, fin_res);
+                    compute(&acc[7], 4, 12,  c_block, vs, fin_res);
+
+                }
+                if (l == 0) {
+                    save_res(ii+i,   jj+j,    0,  fin_res);
+                    save_res(ii+i+4, jj+j,    4,  fin_res);
+                    save_res(ii+i,   jj+j+4,  8,  fin_res);
+                    save_res(ii+i+4, jj+j+4, 12,  fin_res);
+                } else {
+                    add_save_res(ii+i,   jj+j,    0,  fin_res);
+                    add_save_res(ii+i+4, jj+j,    4,  fin_res);
+                    add_save_res(ii+i,   jj+j+4,  8,  fin_res);
+                    add_save_res(ii+i+4, jj+j+4, 12,  fin_res);
+                }
+            }
+        }
+    }
+
+    const TA *const A;
+    const block_q8_0 *const B;
+    float *C;
+    const int64_t k;
+    int64_t kc;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -117,8 +117,7 @@ inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
 #endif
 
 #if defined(__MMA__)
-typedef vector unsigned char vec_t;
-typedef __vector_quad acc_t;
+#include "sgemm-ppc.h"
 #endif
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED FUSED MULTIPLY ADD
@@ -1573,95 +1572,35 @@ class tinyBLAS_BF16_PPC {
     const int nth;
 };
 
-template <typename TA>
-class tinyBLAS_Q0_PPC {
-  public:
-    tinyBLAS_Q0_PPC(int64_t k,
-                const TA *A, int64_t lda,
-                const block_q8_0 *B, int64_t ldb,
-                float *C, int64_t ldc,
-                int ith, int nth)
+    template <typename TA>
+    tinyBLAS_Q0_PPC<TA>::tinyBLAS_Q0_PPC(int64_t k,
+        const TA *A, int64_t lda,
+        const block_q8_0 *B, int64_t ldb,
+        float *C, int64_t ldc,
+        int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+                kc = 64;
     }
 
-    void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
-    }
-
-  private:
-
-    inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-       for (int I = 0; I < RM; I++) {
-          for (int J = 0; J < RN; J++) {
-             *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
-          }
-       }
-    }
-
-    template<int size>
-    inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array<int, size>& comparray, vector float* vs, vector float* fin_res) {
-       vector signed int vec_C[4];
-       vector float CA[4] = {0};
-       vector float res[4] = {0};
-       __builtin_mma_disassemble_acc(vec_C, ACC);
-       for (int i = 0; i < 4; i++) {
-          CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
-          res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
-          fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
-       }
-    }
-    /* This function processes quantized data from block_q4_0 elements.
-     * First the we try to extract the two int4 values stored in single int8_t into two signed int8.
-     * And then we subtract each of the resultant element with 8, to convert signed int8 to unsigned int8.
-     * Also compute the rowsum which is required to compensate the above conversion. */
-    inline void process_q4_elements(vector signed char (&c)[2], int* ca) {
-        const vector signed char lowMask = vec_splats((signed char)0xF);
-        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-        const vector signed char v8 = vec_splats((signed char)0x8);
-        vector signed int vsum = {0};
-        vector signed int vsum2 = {0};
-        c[0] = vec_and(c[1], lowMask);
-        c[1] = vec_sr(c[1], v4);
-        c[0] = vec_sub(c[0], v8);
-        c[1] = vec_sub(c[1], v8);
-        vsum = vec_sum4s(c[0], vsum);
-        vsum2 = vec_sum4s(c[1], vsum2);
-        vsum = vec_add(vsum, vsum2);
-        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
-    }
-
-    template <typename V1, typename V2>
-    inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
-        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
-        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
-        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
-        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
-        V2 t1, t2, t3, t4, t5, t6, t7, t8;
-        vector unsigned char xor_vector;
-        uint8_t flip_vec = 0x80;
-        xor_vector = vec_splats(flip_vec);
-        t1 = vec_perm(s1, s2, swiz1);
-        t2 = vec_perm(s1, s2, swiz2);
-        t3 = vec_perm(s3, s4, swiz1);
-        t4 = vec_perm(s3, s4, swiz2);
-        t5 = vec_perm(t1, t3, swiz3);
-        t6 = vec_perm(t1, t3, swiz4);
-        t7 = vec_perm(t2, t4, swiz3);
-        t8 = vec_perm(t2, t4, swiz4);
-        if (flip == true) {
-            t5 = vec_xor(t5, xor_vector);
-            t6 = vec_xor(t6, xor_vector);
-            t7 = vec_xor(t7, xor_vector);
-            t8 = vec_xor(t8, xor_vector);
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::matmul(int64_t m, int64_t n) {
+        int mc = 64; int nc = 64;
+        if (n % 8 == 0 && n < nc) {
+                nc = n;
+                mc = 32 ;
+                kc = 32;
+        }
+        const bool is_aligned = ((m & (mc - 1)) == 0) & ((n & (nc - 1)) == 0) & ((k & (kc - 1)) == 0);
+        if (is_aligned) {
+            this->matmul_tiled_q0(m, n, mc, nc, kc);
+        } else {
+            mnpack(0, m, 0, n);
         }
-        vec_xst(t5, 0, vecOffset);
-        vec_xst(t6, 0, vecOffset+16);
-        vec_xst(t7, 0, vecOffset+32);
-        vec_xst(t8, 0, vecOffset+48);
     }
 
-    template<int size>
-    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
+   template<typename TA>
+   template<int size>
+   void tinyBLAS_Q0_PPC<TA>::packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
         int64_t i, j;
         TA *aoffset = NULL;
         int8_t *vecOffset = NULL;
@@ -1781,8 +1720,10 @@ class tinyBLAS_Q0_PPC {
             }
         }
     }
+
+    template<typename TA>
     template<typename VA, typename VB>
-    void packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    void tinyBLAS_Q0_PPC<TA>::packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
         int64_t i, j;
         block_q8_0 *aoffset = NULL;
         VA *vecOffset = NULL;
@@ -1822,7 +1763,6 @@ class tinyBLAS_Q0_PPC {
             j--;
         } while(j > 0);
     }
-
     if (rows & 4) {
             aoffsets[0]  = aoffset;
             for (int it = 1; it < 4; it++ )
@@ -1878,7 +1818,8 @@ class tinyBLAS_Q0_PPC {
         }
     }
 
-    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int m_rem = MIN(m - m0, 16);
         int n_rem = MIN(n - n0, 16);
 
@@ -1915,7 +1856,8 @@ class tinyBLAS_Q0_PPC {
     }
 
 
-    void KERNEL_4x8(int64_t ii, int64_t jj) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::KERNEL_4x8(int64_t ii, int64_t jj) {
         vec_t vec_A[8], vec_B[16] = {0};
         acc_t acc_0, acc_1;
         std::array<int, 4> comparray {};
@@ -1953,14 +1895,15 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
+            compute(&acc_0, 0, 0, comparray, vs, fin_res);
+            compute(&acc_1, 0, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
         save_res(ii, jj+4, 4, fin_res);
     }
 
-    void KERNEL_8x4(int64_t ii, int64_t jj) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x4(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[8] = {0};
         acc_t acc_0, acc_1;
         std::array<int, 8> comparray {};
@@ -1997,16 +1940,18 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(&acc_0, 0, 0, comparray, vs, fin_res);
+            compute(&acc_1, 4, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
         save_res(ii+4, jj, 4, fin_res);
     }
 
-    void KERNEL_8x8(int64_t ii, int64_t jj) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x8(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[16] = {0};
         acc_t acc_0, acc_1, acc_2, acc_3;
+        acc_t acc_4, acc_5, acc_6, acc_7;
         std::array<int, 8> comparray {};
         vector float fin_res[16] = {0};
         vector float vs[16] = {0};
@@ -2046,10 +1991,10 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
-            compute<8>(&acc_2, 0, 8, comparray, vs, fin_res);
-            compute<8>(&acc_3, 4, 12, comparray, vs, fin_res);
+            compute(&acc_0, 0, 0, comparray, vs, fin_res);
+            compute(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(&acc_2, 0, 8, comparray, vs, fin_res);
+            compute(&acc_3, 4, 12, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
         save_res(ii+4, jj, 4, fin_res);
@@ -2057,7 +2002,8 @@ class tinyBLAS_Q0_PPC {
         save_res(ii+4, jj+4, 12, fin_res);
     }
 
-    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2125,21 +2071,9 @@ class tinyBLAS_Q0_PPC {
         }
     }
 
-    template<int RM, int RN>
-    inline void kernel(int64_t ii, int64_t jj) {
-       if constexpr(RM == 4 && RN == 8) {
-          KERNEL_4x8(ii,jj);
-       } else if constexpr(RM == 8 && RN == 4) {
-          KERNEL_8x4(ii,jj);
-       } else if constexpr(RM == 8 && RN == 8) {
-          KERNEL_8x8(ii,jj);
-       } else {
-          assert(false && "RN/RM values not supported");
-       }
-    }
-
+    template<typename TA>
     template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    NOINLINE void tinyBLAS_Q0_PPC<TA>::gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2151,20 +2085,12 @@ class tinyBLAS_Q0_PPC {
         for (int64_t job = start; job < end; ++job) {
             int64_t ii = m0 + job / xtiles * RM;
             int64_t jj = n0 + job % xtiles * RN;
-            kernel<RM, RN>(ii, jj);
+            this->kernel<RM, RN>(ii, jj);
         }
     }
 
-    const TA *const A;
-    const block_q8_0 *const B;
-    float *C;
-    const int64_t k;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};
+template class tinyBLAS_Q0_PPC<block_q4_0>;
+template class tinyBLAS_Q0_PPC<block_q8_0>;
 
 class tinyBLAS_PPC {
   public:

ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -6,6 +6,12 @@
 #include <vecintrin.h>
 #endif
 
+#ifdef _MSC_VER
+#define NOINLINE __declspec(noinline)
+#else
+#define NOINLINE __attribute__((__noinline__))
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

ggml/src/ggml-cpu/ops.cpp

@@ -6383,7 +6383,7 @@ static void ggml_compute_forward_im2col_3d_f16(
                                         const int64_t iih = ioh*s1 + ikh*d1 - p1;
                                         const int64_t iid = iod*s2 + ikd*d2 - p2;
 
-                                        if (iid < 0 || iid >= ID || iih < 0 || iih >= IH || iiw < 0 || iiw >= IW || iid < 0 || iid >= ID) {
+                                        if (iid < 0 || iid >= ID || iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
                                             dst_data[iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw] = 0;
                                         } else {
                                             const float * const s = (const float *) ((const char *)src_data + iid*nb12 + iih*nb11 + iiw*nb10); // [ID, IH, IW]
@@ -6554,8 +6554,13 @@ static void ggml_call_mul_mat(ggml_type type, const ggml_compute_params * params
     ggml_compute_forward_mul_mat(params, &dst);
 }
 
+static inline int64_t ggml_wrap_around(int64_t coord, int64_t size) {
+    return (coord  + size) % size; // adding size avoids negative number weirdness
+}
+
 // ggml_compute_forward_conv_2d
 
+
 static void ggml_compute_forward_conv_2d_impl(const ggml_compute_params * params,
                                               const ggml_tensor *         kernel,  // [KW, KH, IC, OC]
                                               const ggml_tensor *         src,     // [W, H, C, N]
@@ -7420,6 +7425,65 @@ static void ggml_compute_forward_upscale_f32(
                 }
             }
         }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR && (mode_flags & GGML_SCALE_FLAG_ANTIALIAS)) {
+        // Similar to F.interpolate(..., mode="bilinear", align_corners=False, antialias=True)
+        // https://github.com/pytorch/pytorch/blob/8871ff29b743948d1225389d5b7068f37b22750b/aten/src/ATen/native/cpu/UpSampleKernel.cpp
+        auto triangle_filter = [](float x) -> float {
+            return std::max(1.0f - fabsf(x), 0.0f);
+        };
+
+        // support and invscale, minimum 1 pixel for bilinear
+        const float support1  = std::max(1.0f, 1.0f / sf1);
+        const float invscale1 = 1.0f / support1;
+        const float support0  = std::max(1.0f, 1.0f / sf0);
+        const float invscale0 = 1.0f / support0;
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float) i1 + pixel_offset) / sf1;
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float) i0 + pixel_offset) / sf0;
+
+                        // the range of source pixels that contribute
+                        const int64_t x_min = std::max<int64_t>(x - support0 + pixel_offset, 0);
+                        const int64_t x_max = std::min<int64_t>(x + support0 + pixel_offset, ne00);
+                        const int64_t y_min = std::max<int64_t>(y - support1 + pixel_offset, 0);
+                        const int64_t y_max = std::min<int64_t>(y + support1 + pixel_offset, ne01);
+
+                        // bilinear filter with antialiasing
+                        float val = 0.0f;
+                        float total_weight = 0.0f;
+
+                        for (int64_t sy = y_min; sy < y_max; sy++) {
+                            const float weight_y = triangle_filter((sy - y + pixel_offset) * invscale1);
+
+                            for (int64_t sx = x_min; sx < x_max; sx++) {
+                                const float weight_x = triangle_filter((sx - x + pixel_offset) * invscale0);
+                                const float weight = weight_x * weight_y;
+
+                                if (weight <= 0.0f) {
+                                    continue;
+                                }
+
+                                const float pixel = *(const float *)((const char *)src0->data + sx*nb00 + sy*nb01 + i02*nb02 + i03*nb03);
+                                val += pixel * weight;
+                                total_weight += weight;
+                            }
+                        }
+
+                        if (total_weight > 0.0f) {
+                            val /= total_weight;
+                        }
+
+                        float * dst_ptr = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *dst_ptr = val;
+                    }
+                }
+            }
+        }
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -7532,6 +7596,7 @@ void ggml_compute_forward_upscale(
 
 // ggml_compute_forward_pad
 
+template<bool circular_t>
 static void ggml_compute_forward_pad_f32(
     const ggml_compute_params * params,
           ggml_tensor * dst) {
@@ -7556,23 +7621,40 @@ static void ggml_compute_forward_pad_f32(
     const int32_t lp3 = ggml_get_op_params_i32(dst, 6);
     const int32_t rp3 = ggml_get_op_params_i32(dst, 7);
 
-
     // TODO: optimize
 
     for (int64_t i2 = 0; i2 < ne2; ++i2) {
         for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
             for (int64_t i0 = 0; i0 < ne0; ++i0) {
                 for (int64_t i3 = 0; i3 < ne3; ++i3) {
-                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
-                    if ((i0 >= lp0 && i0 < ne0 - rp0) \
-                         && (i1 >= lp1 && i1 < ne1 - rp1) \
-                         && (i2 >= lp2 && i2 < ne2 - rp2) \
-                         && (i3 >= lp3 && i3 < ne3 - rp3)) {
-                        const int64_t src_idx = (i3 - lp3)*nb03 + (i2 - lp2)*nb02 + (i1 - lp1)*nb01 + (i0 - lp0)*nb00;
+                    // circular means wrap around on a torus, so x and y loop around
+                    if constexpr (circular_t) {
+                        const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+                        const int64_t src_i0 = ggml_wrap_around(i0 - lp0, ne00);
+                        const int64_t src_i1 = ggml_wrap_around(i1 - lp1, ne01);
+                        const int64_t src_i2 = ggml_wrap_around(i2 - lp2, ne02);
+                        const int64_t src_i3 = ggml_wrap_around(i3 - lp3, ne03);
+
+                        const int64_t src_idx =
+                            src_i3*nb03 +
+                            src_i2*nb02 +
+                            src_i1*nb01 +
+                            src_i0*nb00;
+
                         const float * src_ptr = (const float *)((char *) src0->data + src_idx);
                         dst_ptr[dst_idx] = *src_ptr;
                     } else {
-                        dst_ptr[dst_idx] = 0;
+                        const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+                        if ((i0 >= lp0 && i0 < ne0 - rp0) \
+                            && (i1 >= lp1 && i1 < ne1 - rp1) \
+                            && (i2 >= lp2 && i2 < ne2 - rp2) \
+                            && (i3 >= lp3 && i3 < ne3 - rp3)) {
+                            const int64_t src_idx = (i3 - lp3)*nb03 + (i2 - lp2)*nb02 + (i1 - lp1)*nb01 + (i0 - lp0)*nb00;
+                            const float * src_ptr = (const float *)((char *) src0->data + src_idx);
+                            dst_ptr[dst_idx] = *src_ptr;
+                        } else {
+                            dst_ptr[dst_idx] = 0;
+                        }
                     }
                 }
             }
@@ -7580,16 +7662,20 @@ static void ggml_compute_forward_pad_f32(
     }
 }
 
+
 void ggml_compute_forward_pad(
     const ggml_compute_params * params,
     ggml_tensor * dst) {
-
     const ggml_tensor * src0 = dst->src[0];
-
+    const bool circular = (bool) ggml_get_op_params_i32(dst, 8);
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_pad_f32(params, dst);
+                if (circular) {
+                    ggml_compute_forward_pad_f32<true>(params, dst);
+                } else {
+                    ggml_compute_forward_pad_f32<false>(params, dst);
+                }
             } break;
         default:
             {
@@ -9766,7 +9852,8 @@ static void ggml_compute_forward_solve_tri_f32(const struct ggml_compute_params
             }
 
             const float diag = A_batch[i00 * n + i00];
-            GGML_ASSERT(diag != 0.0f && "Zero diagonal in triangular matrix");
+            assert(diag != 0.0f && "Zero diagonal in triangular matrix");
+
             X_batch[i00 * k + i01] = (B_batch[i00 * k + i01] - sum) / diag;
         }
     }

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);
 
-    cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream);
+    CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));
 
     size_t temp_storage_bytes = 0;
 

ggml/src/ggml-cuda/common.cuh

@@ -21,10 +21,12 @@
 #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)
@@ -84,12 +86,12 @@
 
 #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_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_PH1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // MTT S5000
 
 #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_NG)
-#define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
+#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)
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
 #    define GGML_CUDA_USE_CUB
@@ -212,9 +214,9 @@ 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) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
@@ -224,7 +226,7 @@ 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)
+#if defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA3))
 #define AMD_WMMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(RDNA4)
 
@@ -250,12 +252,14 @@ 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;
+    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
 }
 
 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_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc));
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -272,7 +276,9 @@ 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;
+    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);
 }
 
 static bool fp32_mma_hardware_available(const int cc) {
@@ -288,7 +294,7 @@ static bool amd_mfma_available(const int cc) {
 }
 
 static bool amd_wmma_available(const int cc) {
-    return GGML_CUDA_CC_IS_RDNA4(cc);
+    return (GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_RDNA3(cc));
 }
 
 static bool volta_mma_available(const int cc) {
@@ -457,6 +463,53 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
+template<typename T, int width = WARP_SIZE>
+static __device__ __forceinline__ T warp_prefix_inclusive_sum(T x) {
+    const int lane_id = threadIdx.x % width;
+#pragma unroll
+    for (int offset = 1; offset < width; offset <<= 1) {
+        const T t = __shfl_up_sync(0xffffffff, x, offset, width);
+        if (lane_id >= offset) {
+            x += t;
+        }
+    }
+    return x;
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ float2 warp_prefix_inclusive_sum(float2 a) {
+    const int lane_id = threadIdx.x % width;
+#pragma unroll
+    for (int offset = 1; offset < width; offset <<= 1) {
+        const float t_x = __shfl_up_sync(0xffffffff, a.x, offset, width);
+        const float t_y = __shfl_up_sync(0xffffffff, a.y, offset, width);
+        if (lane_id >= offset) {
+            a.x += t_x;
+            a.y += t_y;
+        }
+    }
+    return a;
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ half2 warp_prefix_inclusive_sum(half2 a) {
+#ifdef FP16_AVAILABLE
+    const int lane_id = threadIdx.x % width;
+#pragma unroll
+    for (int offset = 1; offset < width; offset <<= 1) {
+        const half2 t = __shfl_up_sync(0xffffffff, a, offset, width);
+        if (lane_id >= offset) {
+            a = __hadd2(a, t);
+        }
+    }
+    return a;
+
+#else
+    NO_DEVICE_CODE;
+    return a;
+#endif // FP16_AVAILABLE
+}
+
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
@@ -558,8 +611,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
     acc += v.y*u.y;
 }
 
-static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
 #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
     asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
 #else
 #ifdef FAST_FP16_AVAILABLE
@@ -571,7 +628,7 @@ 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 // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
+#endif // V_DOT2_F32_F16_AVAILABLE
 }
 
 static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {
@@ -972,6 +1029,157 @@ 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;
+
+    // Original order of nodes in this concurrent region (before interleaving)
+    // Used to restore grouping for fusion within streams
+    std::vector<const ggml_tensor *> original_order;
+
+    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))
+    , original_order(std::move(other.original_order))
+    , 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::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
+
+    void reset() {
+        concurrent_events.clear();
+    }
+};
+
 struct ggml_backend_cuda_context {
     int device;
     std::string name;
@@ -982,11 +1190,15 @@ 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) {
@@ -997,9 +1209,9 @@ struct ggml_backend_cuda_context {
         return streams[device][stream];
     }
 
-    cudaStream_t stream() {
-        return stream(device, 0);
-    }
+    cudaStream_t stream() { return stream(device, curr_stream_no); }
+
+    ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }
 
     cublasHandle_t cublas_handle(int device) {
         if (cublas_handles[device] == nullptr) {
@@ -1015,15 +1227,15 @@ struct ggml_backend_cuda_context {
     }
 
     // pool
-    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
+    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];
 
-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, int stream_no);
 
     ggml_cuda_pool & pool(int device) {
-        if (pools[device] == nullptr) {
-            pools[device] = new_pool_for_device(device);
+        if (pools[device][curr_stream_no] == nullptr) {
+            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no);
         }
-        return *pools[device];
+        return *pools[device][curr_stream_no];
     }
 
     ggml_cuda_pool & pool() {

ggml/src/ggml-cuda/cpy.cu

@@ -86,6 +86,9 @@ 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) {
@@ -202,7 +205,7 @@ static void ggml_cpy_scalar_cuda(
             ne00n = ne00;
             ne01n = ne01;
             ne02n = ne02;
-        } else if (nb00 > nb02) {
+        } else {
             ne00n = ne00;
             ne01n = ne01*ne02;
             ne02n = 1;

ggml/src/ggml-cuda/cumsum.cu

@@ -0,0 +1,237 @@
+#include <algorithm>
+#include "cumsum.cuh"
+#include "convert.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
+
+#ifdef GGML_CUDA_USE_CUB
+#   include <cub/device/device_scan.cuh>
+#endif // GGML_CUDA_USE_CUB
+
+template<typename T, int BLOCK_SIZE>
+static __global__ void cumsum_cub_kernel(
+        const T * __restrict__ src,
+        T * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t  s01, const int64_t  s02, const int64_t  s03,
+        const int64_t   s1,  const int64_t   s2,  const int64_t   s3) {
+#ifdef GGML_CUDA_USE_CUB
+    using BlockScan = cub::BlockScan<T, BLOCK_SIZE>;
+
+    __shared__ typename BlockScan::TempStorage temp_storage;
+    __shared__ T block_carry;      // carry from previous tile
+
+    const int tid = threadIdx.x;
+
+    const int64_t i1 = blockIdx.x;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i3 = blockIdx.z;
+
+    if (i1 >= ne01 || i2 >= ne02 || i3 >= ne03) {
+        return;
+    }
+
+    const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
+    T *       dst_row = dst + i1 * s1  + i2 * s2  + i3 * s3;
+
+    if (tid == 0) {
+        block_carry = 0;
+    }
+    __syncthreads();
+
+    for (int64_t start = 0; start < ne00; start += BLOCK_SIZE) {
+        int64_t idx = start + tid;
+        T x = (idx < ne00) ? src_row[idx] : T(0);
+
+        T inclusive;
+        T block_total;
+        BlockScan(temp_storage).InclusiveSum(x, inclusive, block_total);
+
+        __syncthreads();
+
+        T final_val = inclusive + block_carry;
+
+        // store result
+        if (idx < ne00) {
+            dst_row[idx] = final_val;
+        }
+
+        __syncthreads();
+
+        if (tid == 0) {
+            block_carry += block_total;
+        }
+
+        __syncthreads();
+    }
+#else
+    NO_DEVICE_CODE;
+#endif // GGML_CUDA_USE_CUB
+}
+
+// Fallback kernel implementation (original)
+template<typename T>
+static __global__ void cumsum_kernel(
+        const T * src, T * dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t  s00, const int64_t  s01, const int64_t  s02, const int64_t  s03,
+        const int64_t   s0, const int64_t   s1, const int64_t   s2, const int64_t   s3) {
+
+    GGML_UNUSED_VARS(s00, s0);
+
+    const int tid = threadIdx.x;
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int lane = tid % warp_size;
+    const int warp = tid / warp_size;
+    const int warps_per_block = blockDim.x / warp_size;
+
+    extern __shared__ float smem[];
+    float * s_vals = smem;
+    float * s_warp_sums = smem + blockDim.x;
+    float * s_carry = smem + blockDim.x + warps_per_block;
+    float * s_chunk_total = s_carry + 1;
+
+    // Initialize carry
+    if (tid == 0) {
+        *s_carry = 0.0f;
+    }
+    __syncthreads();
+
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
+    T       * dst_row = dst + i1 * s1  + i2 * s2  + i3 * s3;
+
+    for (int64_t start = 0; start < ne00; start += blockDim.x) {
+        int64_t idx = start + tid;
+        float val = (idx < ne00) ? ggml_cuda_cast<float, T>(src_row[idx]) : 0.0f;
+
+        // 1. Warp inclusive scan
+        val = warp_prefix_inclusive_sum<T, warp_size>(val);
+        s_vals[tid] = val;
+
+        // Store warp total
+        if (lane == warp_size - 1) {
+            s_warp_sums[warp] = val;
+        }
+        __syncthreads();
+
+        // 2. Exclusive scan of warp sums (warp 0 only)
+        if (warp == 0) {
+            float w = (tid < warps_per_block) ? s_warp_sums[tid] : 0.0f;
+            float inc = warp_prefix_inclusive_sum<T, warp_size>(w);
+            if (tid < warps_per_block) {
+                s_warp_sums[tid] = inc - w;   // exclusive sum
+            }
+            if (tid == warps_per_block - 1) {
+                *s_chunk_total = inc;          // total sum of this chunk
+            }
+        }
+        __syncthreads();
+
+        float carry = *s_carry;
+        float final_val = s_vals[tid] + s_warp_sums[warp] + carry;
+        if (idx < ne00) {
+            dst_row[idx] = ggml_cuda_cast<T, float>(final_val);
+        }
+        __syncthreads();
+
+        // Update carry for next chunk
+        if (tid == 0) {
+            *s_carry += *s_chunk_total;
+        }
+        __syncthreads();
+    }
+}
+
+template<typename T>
+static void cumsum_cuda(
+        const T * src, T * dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+        const int64_t  nb0,  const int64_t nb1, const int64_t  nb2, const int64_t  nb3,
+        cudaStream_t stream) {
+
+    const size_t type_size = sizeof(T);
+    bool use_cub = false;
+#ifdef GGML_CUDA_USE_CUB
+    // Check if we can use CUB (data must be contiguous along innermost dimension)
+    const bool is_contiguous = (nb00 == type_size) && (nb0 == type_size);
+
+    if (is_contiguous) {
+        use_cub = true;
+    }
+#endif // GGML_CUDA_USE_CUB
+    dim3 grid_dims(ne01, ne02, ne03);
+    const auto &info = ggml_cuda_info().devices[ggml_cuda_get_device()];
+    const int warp_size = info.warp_size;
+    const int num_warps = (ne00 + warp_size - 1) / warp_size;
+    int block_size = num_warps * warp_size;
+    block_size = std::min(block_size, CUDA_CUMSUM_BLOCK_SIZE);
+    dim3 block_dims(block_size, 1, 1);
+    const int warps_per_block = block_size / warp_size;
+    const size_t shmem_size = (block_size + warps_per_block + 2) * sizeof(float);
+
+    if (use_cub) {
+        cumsum_cub_kernel<T, CUDA_CUMSUM_BLOCK_SIZE><<<grid_dims, CUDA_CUMSUM_BLOCK_SIZE, 0, stream>>>(
+            src, dst,
+            ne00, ne01, ne02, ne03,
+            nb01 / type_size, nb02 / type_size, nb03 / type_size,
+            nb1 / type_size,  nb2 / type_size,  nb3 / type_size
+        );
+    } else {
+        cumsum_kernel<<<grid_dims, block_dims, shmem_size, stream>>>(
+            src, dst,
+            ne00, ne01, ne02, ne03,
+            nb00 / type_size, nb01 / type_size, nb02 / type_size, nb03 / type_size,
+            nb0 / type_size, nb1 / type_size, nb2 / type_size, nb3 / type_size
+        );
+    }
+}
+
+void ggml_cuda_op_cumsum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == dst->type);
+    switch(src0->type) {
+        case GGML_TYPE_F32:
+            {
+                cumsum_cuda(
+                    (const float *)src0->data, (float *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    stream
+                );
+            } break;
+        // We do not support those on CPU for now anyway, so comment them out because they cause errors on some CI platforms
+        /*case GGML_TYPE_F16:
+            {
+                cumsum_cuda(
+                    (const half *)src0->data, (half *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    stream
+                );
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                cumsum_cuda(
+                    (const nv_bfloat16 *)src0->data, (nv_bfloat16 *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    stream
+                );
+            } break;*/
+        default:
+            GGML_ABORT("fatal error");
+    }
+}

ggml/src/ggml-cuda/cumsum.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CUMSUM_BLOCK_SIZE 256
+
+void ggml_cuda_op_cumsum(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -10,6 +10,12 @@
 #define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
 #define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
 
+// log(2) = 0.6931, by adding this to the KQ maximum used for the softmax the numerical range representable
+//     by the VKQ accumulators is effectively being shifted up by a factor of 8.
+// This reduces issues with numerical overflow but also causes larger values to be flushed to zero.
+// However, as the output from FlashAttention will usually be used as an input for a matrix multiplication this should be negligible.
+#define FATTN_KQ_MAX_OFFSET 0.6931f
+
 typedef void (* fattn_kernel_t)(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -25,7 +31,7 @@ typedef void (* fattn_kernel_t)(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
@@ -55,11 +61,11 @@ 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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
@@ -621,7 +627,8 @@ static __global__ void flash_attn_mask_to_KV_max(
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
-        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11) {
+        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11,
+        const int nbatch_fa) {
     constexpr int ncols = ncols1*ncols2;
 
     const int bidx0 = blockIdx.x;
@@ -632,8 +639,8 @@ static __global__ void flash_attn_stream_k_fixup(
 
     const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
 
-    const int iter_k = ne11 / FATTN_KQ_STRIDE;
-    const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
+    const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
+    const int iter_j = (ne01 + (ncols1    - 1)) / ncols1;
 
     const int kbc0      = (bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
     const int kbc0_stop = (bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
@@ -765,7 +772,7 @@ static __global__ void flash_attn_combine_results(
 template <int DV, int ncols1, int ncols2>
 void launch_fattn(
     ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
-    const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
+    const int nbatch_fa, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
@@ -790,8 +797,6 @@ void launch_fattn(
     GGML_ASSERT(!V || V->nb[0] == ggml_element_size(V));
 
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
-    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
-        "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
 
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
@@ -915,7 +920,7 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
-        const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 1) / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+        const int ntiles_KQ = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by tensor size.
 
         // parallel_blocks must not be larger than what the tensor size allows:
         parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
@@ -970,6 +975,9 @@ void launch_fattn(
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
+    // TODO other tensor dimensions after removal of WMMA kernel:
+    const uint3 ne01 = init_fastdiv_values(Q->ne[1]);
+
     GGML_ASSERT(block_dim.x % warp_size == 0);
     fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
         (const char *) Q->data,
@@ -980,7 +988,7 @@ void launch_fattn(
         KV_max.ptr,
         !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
-        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
+        Q->ne[0], ne01,     Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3], nb11, nb12, nb13,
         nb21, nb22, nb23,
         mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0, mask ? mask->ne[3] : 0,
@@ -995,7 +1003,7 @@ void launch_fattn(
 
             flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1]);
+                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], nbatch_fa);
         }
     } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(DV, 1, 1);

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

@@ -501,6 +501,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
         const half2 * const __restrict__ K_h2,
         const half2 * const __restrict__ V_h2,
         const half  * const __restrict__ mask,
+        const uint3 ne01,
         const float logit_softcap,
         const float slope,
         T_KQ      * const KQ,
@@ -512,7 +513,8 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
         float * const KQ_sum,
         T_acc * const VKQ,
         const int k_VKQ_0,
-        const int k_VKQ_max) {
+        const int k_VKQ_max,
+        const int col_Q_0) {
     constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
     constexpr int cpy_ne = cpy_nb / 4;
 
@@ -556,12 +558,18 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
     // Apply logit softcap + mask, update KQ_max:
 #pragma unroll
     for (int jc0 = 0; jc0 < cpw; ++jc0) {
-        const int j = (jc0 + (threadIdx.y / np)*cpw)/ncols2;
+        const int j = fastmodulo(col_Q_0 + (jc0 + (threadIdx.y / np)*cpw)/ncols2, ne01);
 
 #pragma unroll
         for (int i_KQ_0 = 0; i_KQ_0 < nbatch_fa; i_KQ_0 += np*warp_size) {
             const int i_KQ = i_KQ_0 + (threadIdx.y % np)*warp_size + threadIdx.x;
 
+#if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+            // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
+            // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
+            KQ_acc[i_KQ_0/(np*warp_size)*cpw + jc0] *= 4.0f;
+#endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+
             if (use_logit_softcap) {
                 KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] = logit_softcap * tanhf(KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
             }
@@ -570,7 +578,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
                 KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] += (ncols2 > 1 || mask) ?
                     slope*__half2float(mask[j*stride_mask + k_VKQ_0 + i_KQ]) : 0.0f;
 
-                KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
+                KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] + FATTN_KQ_MAX_OFFSET);
             }
         }
 
@@ -609,7 +617,7 @@ 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 < k_VKQ_sup ?
+                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < static_cast<uint32_t>(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;
@@ -736,7 +744,7 @@ static __global__ void flash_attn_tile(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
@@ -781,11 +789,11 @@ static __global__ void flash_attn_tile(
     const int sequence = blockIdx.z / (ne02/ncols2);
     const int head0 = blockIdx.z*ncols2 - sequence*ne02; // == blockIdx.z % (ne02/ncols2)
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f  = (const float *) (Q + nb03*sequence + nb02* head0              + nb01*col_Q_0);
+    const float * Q_f  = (const float *) (Q + nb03*sequence + nb02* head0);
     const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
     const half2 * V_h2 = (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio)); // K and V have same shape
 
-    const half * maskh = mask ? (const half *) (mask + nb33*(sequence % ne33) + nb31*col_Q_0) : nullptr;
+    const half * maskh = mask ? (const half *) (mask + nb33*(sequence % ne33)) : nullptr;
 
     const int stride_K2   = nb11 / sizeof(half2);
     const int stride_V2   = nb21 / sizeof(half2);
@@ -842,11 +850,9 @@ static __global__ void flash_attn_tile(
         for (int i0 = 0; i0 < DKQp; i0 += np*warp_size*cpy_ne_D) {
             if (i0 + np*warp_size*cpy_ne_D <= DKQ || i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D < DKQ) {
                 float tmp_f[cpy_ne_D] = {0.0f};
-                if (ncols1 == 1 || col_Q_0 + j < ne01) {
-                    ggml_cuda_memcpy_1<sizeof(tmp_f)>
-                        (tmp_f, &Q_f[c*(nb02/sizeof(float)) + j*(nb01/sizeof(float))
-                                     + i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D]);
-                }
+                ggml_cuda_memcpy_1<sizeof(tmp_f)>
+                    (tmp_f, &Q_f[c*(nb02/sizeof(float)) + fastmodulo(col_Q_0 + j, ne01)*(nb01/sizeof(float))
+                                 + i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D]);
 
 #pragma unroll
                 for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
@@ -858,6 +864,11 @@ static __global__ void flash_attn_tile(
 #pragma unroll
                 for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
                     tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
+#if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+                    // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
+                    // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
+                    tmp_h2[i1/2] *= make_half2(0.25f, 0.25f);
+#endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
                 }
                 ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
                     &Q_tmp[jc*(DKQ/2) + i0/2 + (threadIdx.y % np)*(warp_size*cpy_ne_D/2) + threadIdx.x*(cpy_ne_D/2)],
@@ -881,23 +892,23 @@ static __global__ void flash_attn_tile(
         while (k_VKQ_0 < k_VKQ_max - nbatch_fa) {
             constexpr bool oob_check = false;
             flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
-                (Q_tmp, K_h2, V_h2, maskh, logit_softcap, slope, KQ, KV_tmp,
-                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max);
+                (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
+                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
             k_VKQ_0 += gridDim.y*nbatch_fa;
         }
         if (k_VKQ_0 < k_VKQ_max) {
             constexpr bool oob_check = true;
             flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
-                (Q_tmp, K_h2, V_h2, maskh, logit_softcap, slope, KQ, KV_tmp,
-                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max);
+                (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
+                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
         }
     } else {
         // Branch without out-of-bounds checks.
         for (int k_VKQ_0 = blockIdx.y*nbatch_fa; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*nbatch_fa) {
             constexpr bool oob_check = false;
             flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
-                (Q_tmp, K_h2, V_h2, maskh, logit_softcap, slope, KQ, KV_tmp,
-                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max);
+                (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
+                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
         }
     }
 
@@ -1010,13 +1021,13 @@ static __global__ void flash_attn_tile(
         const int j = jc / ncols2;
         const int c = jc % ncols2;
 
-        if (ncols1 > 1 && col_Q_0 + j >= ne01) {
+        if (ncols1 > 1 && col_Q_0 + j >= int(ne01.z)) {
             return;
         }
 
         const float scale = gridDim.y == 1 ? 1.0f/KQ_sum[jc0] : 1.0f;
 
-        const int j_dst_unrolled = ((sequence*ne01 + col_Q_0 + j)*ne02 + head0 + c)*gridDim.y + blockIdx.y;
+        const int j_dst_unrolled = ((sequence*int(ne01.z) + col_Q_0 + j)*ne02 + head0 + c)*gridDim.y + blockIdx.y;
 
 #ifdef FAST_FP16_AVAILABLE
         constexpr int cpy_ne_D = cpy_ne/2 < (DVp/2)/warp_size ? cpy_ne/2 : (DVp/2)/warp_size;

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

@@ -33,7 +33,7 @@ static __global__ void flash_attn_ext_vec(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
@@ -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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_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) {
@@ -150,12 +150,12 @@ static __global__ void flash_attn_ext_vec(
             float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
 
             // Set memory to zero if out of bounds:
-            if (ncols > 1 && ic0 + j >= ne01) {
+            if (ncols > 1 && ic0 + j >= int(ne01.z)) {
 #pragma unroll
                 for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
                     const int i = i0 + threadIdx.x;
 
-                    if (i0 + WARP_SIZE <= D/sizeof(int) || i < D/sizeof(int)) {
+                    if (i0 + WARP_SIZE <= int(D/sizeof(int)) || i < int(D/sizeof(int))) {
                         tmp_q_i32[i] = 0;
                     }
                 }
@@ -191,7 +191,7 @@ static __global__ void flash_attn_ext_vec(
 
         __syncthreads();
     } else {
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
         const half2 scale_h2 = make_half2(scale, scale);
 #pragma unroll
         for (int j = 0; j < ncols; ++j) {
@@ -201,7 +201,7 @@ static __global__ void flash_attn_ext_vec(
                 const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
 
                 float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
-                if (ncols == 1 || ic0 + j < ne01) {
+                if (ncols == 1 || ic0 + j < int(ne01.z)) {
                     ggml_cuda_memcpy_1<cpy_nb>(tmp,            &Q_j[i]);
                     ggml_cuda_memcpy_1<cpy_nb>(tmp + cpy_ne/2, &Q_j[i + cpy_ne/2]);
                 }
@@ -222,7 +222,7 @@ static __global__ void flash_attn_ext_vec(
 #pragma unroll
             for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
                 const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
-                if (ncols == 1 || ic0 + j < ne01) {
+                if (ncols == 1 || ic0 + j < int(ne01.z)) {
                     ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ],            &Q_j[i]);
                     ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ + cpy_ne/2], &Q_j[i + cpy_ne/2]);
                 }
@@ -233,7 +233,7 @@ static __global__ void flash_attn_ext_vec(
                 Q_reg[j][k].y *= scale;
             }
         }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
     }
 
     const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
@@ -266,13 +266,13 @@ static __global__ void flash_attn_ext_vec(
                     sum = logit_softcap*tanhf(sum);
                 }
 
-                if (mask) {
+                if (mask && (ncols == 1 || ic0 + j < int(ne01.z))) {
                     sum += slope*__half2float(maskh[j*ne11 + i_KQ]);
                 }
 
-                KQ_max_new[j] = fmaxf(KQ_max_new[j], sum);
+                KQ_max_new[j] = fmaxf(KQ_max_new[j], sum + FATTN_KQ_MAX_OFFSET);
 
-                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == i_KQ_0) {
+                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == uint32_t(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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
@@ -412,7 +412,7 @@ static __global__ void flash_attn_ext_vec(
 
 #pragma unroll
     for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        if (ncols > 1 && ic0 + j_VKQ >= ne01) {
+        if (ncols > 1 && ic0 + j_VKQ >= int(ne01.z)) {
             break;
         }
 
@@ -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 FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_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 // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
         KQ_sum[j_VKQ] *= kqmax_scale;
         KQ_sum[j_VKQ] = warp_reduce_sum(KQ_sum[j_VKQ]);
@@ -479,7 +479,7 @@ static __global__ void flash_attn_ext_vec(
                 if (gridDim.y == 1) {
                     dst_val /= KQ_sum[j_VKQ];
                 }
-                dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
+                dst[(((sequence*int(ne01.z) + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
             }
         }
 
@@ -489,8 +489,8 @@ static __global__ void flash_attn_ext_vec(
 
     }
 
-    if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < ne01)) {
-        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
+    if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < int(ne01.z))) {
+        dst_meta[((sequence*int(ne01.z) + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
     }
 #else
     GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,

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

@@ -38,14 +38,14 @@ static __global__ void flash_attn_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
+#if defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(D == 128 || D == 256)) {
         NO_DEVICE_CODE;
@@ -149,7 +149,7 @@ static __global__ void flash_attn_ext_f16(
             if (i0 + warp_size > D && i >= D) {
                 break;
             }
-            KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f;
+            KQ[j*D_padded + i] = ic0 + j < int(ne01.z) ? Q_f[j*stride_Q + i] * scale : 0.0f;
         }
     }
 
@@ -218,8 +218,9 @@ static __global__ void flash_attn_ext_f16(
                 for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += warp_size) {
                     const int k = k0 + threadIdx.x;
 
-                    KQ_f_tmp[k0/warp_size] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
-                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/warp_size]);
+                    KQ_f_tmp[k0/warp_size] += mask && ic0 + j < int(ne01.z) ?
+                        __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
+                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/warp_size] + FATTN_KQ_MAX_OFFSET);
                 }
                 KQ_max_new = warp_reduce_max<warp_size>(KQ_max_new);
 
@@ -270,7 +271,7 @@ static __global__ void flash_attn_ext_f16(
                 for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += warp_size) {
                     const int k = k0 + threadIdx.x;
 
-                    KQ2_tmp[k0/warp_size] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
+                    KQ2_tmp[k0/warp_size] += mask && ic0 + j < int(ne01.z) ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
                     KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/warp_size]);
                 }
                 KQ_max_new = __half2half2(warp_reduce_max<warp_size>(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
@@ -431,7 +432,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
     for (int j0 = 0; j0 < ncols; j0 += nwarps) {
         const int j_VKQ = j0 + threadIdx.y;
-        if (ic0 + j_VKQ >= ne01) {
+        if (ic0 + j_VKQ >= int(ne01.z)) {
             return;
         }
 
@@ -442,7 +443,7 @@ static __global__ void flash_attn_ext_f16(
             KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
         }
 
-        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
+        const int j_dst_unrolled = ((sequence*int(ne01.z) + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
 
 #pragma unroll
         for (int i0 = 0; i0 < D; i0 += warp_size) {
@@ -481,7 +482,7 @@ static __global__ void flash_attn_ext_f16(
               ne31, ne32, ne33,
               nb31, nb32, nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
+#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
 }
 
 constexpr int get_max_power_of_2(int x) {

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

@@ -2,9 +2,9 @@
 
 #include "common.cuh"
 
-#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+#if defined(GGML_USE_MUSA)
 #define GGML_USE_WMMA_FATTN
-#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+#endif // defined(GGML_USE_MUSA)
 
 #if defined(GGML_HIP_ROCWMMA_FATTN)
 #if defined(CDNA) && (ROCWMMA_VERSION_MAJOR < 2 || ROCWMMA_VERSION_MINOR > 0 || ROCWMMA_VERSION_PATCH > 0)

ggml/src/ggml-cuda/fattn.cu

@@ -12,13 +12,13 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
     const ggml_tensor * Q = dst->src[0];
 
     if constexpr (ncols2 <= 8) {
-        if (Q->ne[1] <= 8/ncols2) {
+        if (turing_mma_available(cc) && Q->ne[1] <= 8/ncols2) {
             ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
             return;
         }
     }
 
-    if (Q->ne[1] <= 16/ncols2) {
+    if (turing_mma_available(cc) && Q->ne[1] <= 16/ncols2) {
         ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
         return;
     }
@@ -41,7 +41,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     float max_bias = 0.0f;
     memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
 
-    const bool use_gqa_opt = mask && max_bias == 0.0f;
+    const bool use_gqa_opt = mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
@@ -275,8 +275,8 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
-    // If Turing tensor cores available, use them:
-    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72) {
+    // If Turing tensor cores are available, use them:
+    if (turing_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if (can_use_vector_kernel) {
             if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
@@ -297,7 +297,21 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
                 return BEST_FATTN_KERNEL_VEC;
             }
         }
+        return BEST_FATTN_KERNEL_MMA_F16;
+    }
 
+    if (volta_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
+        int gqa_ratio_eff = 1;
+        const int ncols2_max = Q->ne[0] == 576 ? 16 : 8;
+        while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {
+            gqa_ratio_eff *= 2;
+        }
+        if (can_use_vector_kernel && Q->ne[1] * gqa_ratio_eff <= 2) {
+            return BEST_FATTN_KERNEL_VEC;
+        }
+        if (Q->ne[1] * gqa_ratio_eff <= 16) {
+            return BEST_FATTN_KERNEL_TILE; // On Volta tensor cores are only faster for sufficiently large matrices.
+        }
         return BEST_FATTN_KERNEL_MMA_F16;
     }
 

ggml/src/ggml-cuda/fill.cu

@@ -0,0 +1,37 @@
+#include "fill.cuh"
+#include "convert.cuh"
+
+#define CUDA_FILL_BLOCK_SIZE 256
+
+template <typename T>
+static __global__ void fill_kernel(T * __restrict__ dst, const int64_t k, const T value) {
+    const int64_t i = (int64_t)blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = value;
+}
+
+void ggml_cuda_op_fill(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    void * dst_d = dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    float value;
+    memcpy(&value, dst->op_params, sizeof(float));
+
+    const int64_t k = ggml_nelements(dst);
+    const int64_t num_blocks = (k + CUDA_FILL_BLOCK_SIZE - 1) / CUDA_FILL_BLOCK_SIZE;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            fill_kernel<<<num_blocks, CUDA_FILL_BLOCK_SIZE, 0, stream>>>((float *)dst_d, k, value);
+            break;
+        case GGML_TYPE_F16:
+            fill_kernel<<<num_blocks, CUDA_FILL_BLOCK_SIZE, 0, stream>>>((half *)dst_d, k, ggml_cuda_cast<half>(value));
+            break;
+        default:
+            GGML_ABORT("unsupported type");
+    }
+}

ggml/src/ggml-cuda/fill.cuh

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

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

@@ -53,6 +53,10 @@
 #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-cuda/tri.cuh"
+#include "ggml-cuda/cumsum.cuh"
+#include "ggml-cuda/fill.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -521,7 +525,8 @@ 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) {
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int                  device,
+                                                                               [[maybe_unused]] int stream_no) {
 #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));
@@ -2699,6 +2704,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CROSS_ENTROPY_LOSS:
             ggml_cuda_cross_entropy_loss(ctx, dst);
             break;
+        case GGML_OP_CUMSUM:
+            ggml_cuda_op_cumsum(ctx, dst);
+            break;
+        case GGML_OP_TRI:
+            ggml_cuda_op_tri(ctx, dst);
+            break;
         case GGML_OP_RWKV_WKV6:
             ggml_cuda_op_rwkv_wkv6(ctx, dst);
             break;
@@ -2717,6 +2728,12 @@ 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;
+        case GGML_OP_FILL:
+            ggml_cuda_op_fill(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3046,7 +3063,12 @@ 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);
 
-    if (ops.size() == topk_moe_ops_with_norm.size() &&
+    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) &&
         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];
@@ -3056,8 +3078,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
+    if (is_equal(topk_moe_ops, ops) && 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)) {
@@ -3065,7 +3086,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
+    if (is_equal(topk_moe_ops_delayed_softmax, ops) &&
         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];
@@ -3081,9 +3102,8 @@ 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 (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}))) {
-
+    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 })) {
         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];
@@ -3095,9 +3115,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    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}))) {
-
+    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 })) {
         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];
@@ -3107,7 +3126,9 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == 3 && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+    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 })) {
         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];
@@ -3192,27 +3213,141 @@ 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 original node order within each concurrent region to enable fusion within streams
+
+                std::unordered_map<const ggml_tensor *, int> node_to_idx;
+                node_to_idx.reserve(cgraph->n_nodes);
+                for (int i = 0; i < cgraph->n_nodes; ++i) {
+                    node_to_idx[cgraph->nodes[i]] = i;
+                }
+
+                for (auto & [fork_node, event] : stream_ctx.concurrent_events) {
+                    // Find positions of all nodes from this event in the current graph
+                    std::vector<int> positions;
+                    positions.reserve(event.original_order.size());
+
+                    bool all_found = true;
+                    for (const ggml_tensor * orig_node : event.original_order) {
+                        auto it = node_to_idx.find(orig_node);
+                        if (it != node_to_idx.end()) {
+                            positions.push_back(it->second);
+                        } else {
+                            all_found = false;
+                            break;
+                        }
+                    }
+
+                    if (!all_found || positions.size() != event.original_order.size()) {
+                        continue;
+                    }
+
+                    // Sort positions to get contiguous range
+                    std::vector<int> sorted_positions = positions;
+                    std::sort(sorted_positions.begin(), sorted_positions.end());
+
+                    bool is_contiguous = true;
+                    for (size_t i = 1; i < sorted_positions.size(); ++i) {
+                        if (sorted_positions[i] != sorted_positions[i-1] + 1) {
+                            is_contiguous = false;
+                            break;
+                        }
+                    }
+
+                    if (!is_contiguous) {
+                        continue;
+                    }
+
+                    // Restore original order at the sorted positions
+                    int start_pos = sorted_positions[0];
+                    for (size_t i = 0; i < event.original_order.size(); ++i) {
+                        cgraph->nodes[start_pos + i] = const_cast<ggml_tensor *>(event.original_order[i]);
+                    }
+                }
+            }
+
             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);
+                    }
+                }
+
 #ifdef GGML_CUDA_DEBUG
                 const int nodes_fused = i - prev_i - 1;
-                prev_i = i;
                 if (nodes_fused > 0) {
                     GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
                 }
 #endif
+                prev_i = i;
 
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
                 }
 
+
+                // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
 
@@ -3505,13 +3640,17 @@ 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);
+               }
             }
         }
 
@@ -3651,6 +3790,234 @@ 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;
+
+    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;
+                }
+
+                // Save the original order of nodes in this region before interleaving
+                // This is used later to restore grouping for fusion within streams
+                concurrent_event.original_order.reserve(total_branch_nodes);
+                for (int i = fork_node_idx + 1; i < join_node_idx; ++i) {
+                    concurrent_event.original_order.push_back(cgraph->nodes[i]);
+                }
+
+                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,
@@ -3665,7 +4032,7 @@ 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          = */ NULL,
+    /* .graph_optimize          = */ ggml_backend_cuda_graph_optimize,
 };
 
 static ggml_guid_t ggml_backend_cuda_guid() {
@@ -4254,7 +4621,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
+        case GGML_OP_FILL:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
             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

@@ -68,10 +68,31 @@ static __device__ __forceinline__ half2 ggml_cuda_movmatrix(const half2 x) {
 
 namespace ggml_cuda_mma {
 
+    // Some architectures like Volta or CDNA3 perform multiple matrix multiplications per warp in parallel,
+    //     effectively the warp is being split into subgroups of threads that each perform a single mma instruction.
+    // In those cases the data can be split in different ways across the warp.
+    enum data_layout {
+        // By default the data uses the I direction as its major dimension and the J direction as its minor dimension.
+        // For the A/C matrices this means I major == row major, J major == column major.
+        // For the B matrix this means I major == column major, J major == row major.
+        // MIRRORED == Each data value is held exactly once per thread subgroup.
+        DATA_LAYOUT_I_MAJOR           =  0, // Always used for Turing, Ampere, Ada Lovelace, consumer Blackwell.
+        DATA_LAYOUT_I_MAJOR_MIRRORED  = 10,
+        DATA_LAYOUT_J_MAJOR_MIRRORED  = 20,
+    };
+    // Implemented mma combinations are:
+    //   - (I_MAJOR, I_MAJOR)          -> I_MAJOR
+    //   - (I_MAJOR, I_MAJOR_MIRRORED) -> I_MAJOR
+    //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
+
+    template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
+    struct tile {};
+
     template <int I_, int J_, typename T>
-    struct tile {
-        static constexpr int I  = I_;
-        static constexpr int J  = J_;
+    struct tile<I_, J_, T, DATA_LAYOUT_I_MAJOR> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
 #if defined(AMD_MFMA_AVAILABLE)
         static constexpr int ne = I * J / 64;
@@ -131,9 +152,9 @@ namespace ggml_cuda_mma {
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 32 && J == 8) {
 #ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
-                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+                return (((threadIdx.x % 16) / 4) * 8) + ((threadIdx.x / 16) * 4) + (l & 2) + (threadIdx.x % 2);
 #else
-                return (l & 2) | (threadIdx.x & ~2);
+                return (l & 2) + (threadIdx.x & ~2);
 #endif // GGML_CUDA_MMA_NO_VOLTA_PERM
             } else {
                 NO_DEVICE_CODE;
@@ -143,7 +164,7 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 32 && J == 8) {
-                return (threadIdx.x & 2) | (l & (4 + 1));
+                return (threadIdx.x & 2) + (l & (4 + 1));
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -152,6 +173,9 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_WMMA_AVAILABLE)
 #if defined(RDNA4)
         static constexpr int ne = I * J / 32;
+#elif defined(RDNA3)
+        static constexpr int ne = (I == 16 && J == 16) ? I * J / 32 : I * J / 16;
+#endif // defined(RDNA4)
         T x[ne] = {0};
 
         static constexpr __device__ bool supported() {
@@ -161,7 +185,11 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 16 && J == 16) {
+#if defined(RDNA4)
                 return 8 * (threadIdx.x / 16) + l;
+#elif defined(RDNA3)
+                return 2 * l + (threadIdx.x / 16);
+#endif // defined(RDNA4)
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -176,7 +204,6 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
@@ -196,9 +223,9 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return ((l / 2) * 8) | (threadIdx.x / 4);
+                return ((l / 2) * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+                return (((l / 2) % 2) * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 32 && J == 8) {
                 return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
@@ -211,11 +238,11 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return (l * 4) | (threadIdx.x % 4);
+                return (l * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return ((threadIdx.x % 4) * 2) | (l % 2);
+                return ((threadIdx.x % 4) * 2) + (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+                return ((l / 4) * 8) + ((threadIdx.x % 4) * 2) + (l % 2);
             } else if constexpr (I == 32 && J == 8) {
                 return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
@@ -227,26 +254,24 @@ namespace ggml_cuda_mma {
     };
 
     template <int I_, int J_>
-    struct tile<I_, J_, half2> {
-        static constexpr int I  = I_;
-        static constexpr int J  = J_;
+    struct tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
 #if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
         static constexpr __device__ bool supported() {
-            if (I ==  8 && J ==  8) return true;
-            if (I == 32 && J ==  8) return true;
+            if (I == 32 && J ==  4) return true;
             return false;
         }
 
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && J == 8) {
-                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
-            } else if constexpr (I == 32 && J == 8) {
+            if constexpr (I == 32 && J == 4) {
 #ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
-                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+                return (((threadIdx.x % 16) / 4) * 8) + ((threadIdx.x / 16) * 4) + (threadIdx.x % 4);
 #else
                 return threadIdx.x;
 #endif // GGML_CUDA_MMA_NO_VOLTA_PERM
@@ -257,7 +282,7 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr ((I == 8 || I == 32) && J == 8) {
+            if constexpr (I == 32 && J == 4) {
                 return l;
             } else {
                 NO_DEVICE_CODE;
@@ -265,6 +290,7 @@ namespace ggml_cuda_mma {
             }
         }
 #elif defined(AMD_WMMA_AVAILABLE)
+
         static constexpr int ne = I * J / 32;
         half2 x[ne] = {{0.0f, 0.0f}};
 
@@ -307,11 +333,11 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return (l * 8) | (threadIdx.x / 4);
+                return (l * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return ((l % 2) * 8) | (threadIdx.x / 4);
+                return ((l % 2) * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 32 && J == 8) {
-                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
+                return ((l / 4) * 16) + ((l % 2) * 8) + (threadIdx.x / 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -320,13 +346,13 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return (l * 4) | (threadIdx.x % 4);
+                return (l * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return ((l / 2) * 4) | (threadIdx.x % 4);
+                return ((l / 2) * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 32 && J == 8) {
-                return ((l & 2) * 2) | (threadIdx.x % 4);
+                return ((l & 2) * 2) + (threadIdx.x % 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -336,14 +362,15 @@ namespace ggml_cuda_mma {
     };
 
     template <int I_, int J_>
-    struct tile<I_, J_, nv_bfloat162> {
-        static constexpr int I  = I_;
-        static constexpr int J  = J_;
+    struct tile<I_, J_, nv_bfloat162, DATA_LAYOUT_I_MAJOR> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
+        static constexpr int         ne = I * J / WARP_SIZE;
 
-#if defined(AMD_WMMA_AVAILABLE)
-        static constexpr int ne = I * J / 32;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+#if defined(AMD_WMMA_AVAILABLE)
         static constexpr __device__ bool supported() {
             if (I == 16 && J == 8) return true;
             return false;
@@ -367,9 +394,6 @@ namespace ggml_cuda_mma {
             }
         }
 #else
-        static constexpr int ne = I * J / WARP_SIZE;
-        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
-
         static constexpr __device__ bool supported() {
             if (I ==  8 && J ==  8) return true;
             if (I == 16 && J ==  4) return true;
@@ -381,9 +405,9 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return (l * 8) | (threadIdx.x / 4);
+                return (l * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return ((l % 2) * 8) | (threadIdx.x / 4);
+                return ((l % 2) * 8) + (threadIdx.x / 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -392,11 +416,11 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return (l * 4) | (threadIdx.x % 4);
+                return (l * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return ((l / 2) * 4) | (threadIdx.x % 4);
+                return ((l / 2) * 4) + (threadIdx.x % 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -405,6 +429,73 @@ namespace ggml_cuda_mma {
 #endif  // defined(AMD_WMMA_AVAILABLE)
     };
 
+    template <int I_, int J_>
+    struct tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR_MIRRORED;
+        static constexpr int         ne = I * J / (WARP_SIZE/4);
+
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int /*l*/) {
+            if constexpr (I == 8 && J == 4) {
+                return ((threadIdx.x / 16) * 4) + (threadIdx.x % 4);
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 8 && J == 4) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+    };
+
+    template <int I_, int J_>
+    struct tile<I_, J_, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_J_MAJOR_MIRRORED;
+        static constexpr int         ne = I * J / (WARP_SIZE/4);
+
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 4) {
+                return ((l / 2) * 4) + (threadIdx.x % 4);
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 8 && J == 4) {
+                return ((threadIdx.x / 16) * 2) + (l % 2);
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+    };
+
+#if defined(TURING_MMA_AVAILABLE)
     template <int I, int J>
     static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
         tile<I, J/2, half2> ret;
@@ -422,9 +513,26 @@ namespace ggml_cuda_mma {
 
         return ret;
     }
+#else // Volta
+    template <int I, int J>
+    static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
+        tile<I, J/2, half2> ret;
+#pragma unroll
+        for (int l0 = 0; l0 < tile_float.ne; l0 += 4) {
+            ret.x[l0/2 + 0] = make_half2(tile_float.x[l0 + 0], tile_float.x[l0 + 1]);
+            ret.x[l0/2 + 1] = make_half2(tile_float.x[l0 + 2], tile_float.x[l0 + 3]);
 
-    template <int I, int J, typename T>
-    static __device__ __forceinline__ void load_generic(tile<I, J, T> & t, const T * __restrict__ xs0, const int stride) {
+            // On Volta FP16 and FP32 tiles have a different memory layout,
+            //     for the conversion threads with an offset of 2 need to exchange half their values:
+            ret.x[l0/2 + (((threadIdx.x % 4) / 2) ^ 1)] = __shfl_xor_sync(
+                0xFFFFFFFF, ret.x[l0/2 + (((threadIdx.x % 4) / 2) ^ 1)], 2, WARP_SIZE);
+        }
+        return ret;
+    }
+#endif // defined(TURING_MMA_AVAILABLE)
+
+    template <int I, int J, typename T, data_layout dl>
+    static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #if defined(AMD_MFMA_AVAILABLE)
         if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
 #pragma unroll
@@ -443,18 +551,34 @@ namespace ggml_cuda_mma {
         } else if constexpr (std::is_same_v<T, int>) {
             if constexpr (I == 16 && J == 4) {
                 int64_t * xi = (int64_t *) t.x;
+#if defined(RDNA4)
                 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) {
+#elif defined(RDNA3)
+                static_assert(tile<I,J,T>::ne >= 4, "fragment too small");
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride);
+                xi[0] = xs[0];
+                xi[1] = xs[1];
+#endif // defined(RDNA4)
+            } else if constexpr (I == 16 && J == 8) {
                 int64_t * xi = (int64_t *) t.x;
+#if defined(RDNA4)
                 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{
+#elif defined(RDNA3)
+                static_assert(tile<I,J,T>::ne >= 8, "fragment too small");
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride);
+                // contiguous four 64-bit chunks per lane for the wider RDNA3 fragment
+                xi[0] = xs[0];
+                xi[1] = xs[1];
+                const int64_t * xs1 = xs + 2;
+                xi[2] = xs1[0];
+                xi[3] = xs1[1];
+#endif // defined(RDNA4)
+            } else {
                 NO_DEVICE_CODE;
             }
         } else {
@@ -511,18 +635,6 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
 #else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-        GGML_UNUSED_VARS(t, xs0, stride);
-        NO_DEVICE_CODE;
-#else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-#endif // TURING_MMA_AVAILABLE
-    }
-
-    template <typename T>
-    static __device__ __forceinline__ void load_ldmatrix(
-            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
 #if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #if 1
         // TODO: more generic handling
@@ -533,9 +645,31 @@ namespace ggml_cuda_mma {
         load_generic(t, xs0, stride);
 #endif // 1
 #else
-        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
-        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
-        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+        load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // TURING_MMA_AVAILABLE
+    }
+
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & t, const half2 * __restrict__ xs0, const int stride) {
+        ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
+    }
+
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & t, const half2 * __restrict__ xs0, const int stride) {
+#pragma unroll
+        for (int l0 = 0; l0 < t.ne; l0 += 2) {
+            ggml_cuda_memcpy_1<2*sizeof(half2)>(t.x + l0, xs0 + t.get_i(l0)*stride + t.get_j(l0));
+        }
+    }
+
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 4, half2> & t, const half2 * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
+#else
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     }
 
@@ -747,12 +881,14 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #elif defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
         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);
+#endif // RDNA4
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -762,12 +898,14 @@ namespace ggml_cuda_mma {
     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)
+#if defined(RDNA4)
         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);
+#endif // RDNA4
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -796,14 +934,14 @@ namespace ggml_cuda_mma {
 #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)
+        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+        int32x2_t * a_vec = (int32x2_t *) A.x;
+        int32x2_t * b_vec = (int32x2_t *) B.x;
 
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
             true,
@@ -822,7 +960,30 @@ namespace ggml_cuda_mma {
             acc[0],
             true
         );
-#endif // defined(RDNA4)
+
+#elif defined(RDNA3)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * a_vec = (int32x4_t *) A.x;
+        int32x4_t * b_vec = (int32x4_t *) B.x;
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            true
+        );
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+            true,
+            a_vec[1],
+            true,
+            b_vec[1],
+            acc[0],
+            true
+        );
+#endif // RDNA4
 
 #else
         GGML_UNUSED_VARS(D, A, B);
@@ -860,14 +1021,14 @@ namespace ggml_cuda_mma {
     template <typename T1, typename T2, int J, int K>
     static __device__ __forceinline__ void mma(
             tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
-        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
-        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        tile      <16, J, T1> * D16 = reinterpret_cast<      tile<16, J, T1> *>(&D);
+        const tile<16, K, T2> * A16 = reinterpret_cast<const tile<16, K, T2> *>(&A);
         mma(D16[0], A16[0], B);
         mma(D16[1], A16[1], B);
     }
 
     static __device__ __forceinline__ void mma(
-            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+            tile<32, 8, float> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
 #if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
@@ -880,46 +1041,69 @@ namespace ggml_cuda_mma {
             "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
-        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
-            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
-            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
-            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
-        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
-            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
-            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
-            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
 #else
-        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
-        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
-        mma(D16[0], A16[0], B);
-        mma(D16[1], A16[1], B);
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 4, half2> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.row.f16.f16.f16.f16 "
+            "{%0, %1, %2, %3}, {%4, %5}, {%6, %7}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.row.f16.f16.f16.f16 "
+            "{%0, %1, %2, %3}, {%4, %5}, {%6, %7}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     }
 
 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;
+#if defined(RDNA4)
+        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+        int32x2_t * a_vec = (int32x2_t *) A.x;
+        int32x2_t * b_vec = (int32x2_t *) B.x;
 
-    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
+        );
+#elif defined(RDNA3)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * a_vec = (int32x4_t *) A.x;
+        int32x4_t * b_vec = (int32x4_t *) B.x;
 
-    acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-        true,
-        a_vec[0],
-        true,
-        b_vec[0],
-        acc[0],
-        false
-    );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            false
+        );
+#endif // RDNA4
 #else
         GGML_UNUSED(D);
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif
+#endif // AMD_WMMA_AVAILABLE
     }
 }
-

ggml/src/ggml-cuda/mmf.cu

@@ -151,7 +151,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
             return false;
         }
     } else {
-        if (src1_ncols > 16 || GGML_CUDA_CC_IS_RDNA4(cc)) {
+        if (src1_ncols > 16) {
             return false;
         }
     }
@@ -160,9 +160,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc) || amd_wmma_available(cc);
+            return ampere_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
         default:
             return false;
     }

ggml/src/ggml-cuda/mmf.cuh

@@ -37,23 +37,19 @@ static __global__ void mul_mat_f(
     typedef tile<16,       8, T>     tile_A;
     typedef tile<tile_B_I, 8, T>     tile_B;
     typedef tile<16,       tile_C_J, float> tile_C;
-
-    constexpr bool a_supported = tile_A::supported();
-    constexpr bool b_supported = tile_B::supported();
-    constexpr bool c_supported = tile_C::supported();
-    constexpr bool supported = a_supported && b_supported && c_supported;
 #else
-    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
-    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
-    constexpr bool supported = I_16_supported || I_32_supported;
-
-    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;
+#ifdef VOLTA_MMA_AVAILABLE
+    if constexpr (!std::is_same_v<T, half2>) {NO_DEVICE_CODE;} else {
+    typedef tile<32, 4, T,     DATA_LAYOUT_I_MAJOR>          tile_A;
+    typedef tile< 8, 4, T,     DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B;
+    typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR>          tile_C;
+#else
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile<8,  8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+#endif // VOLTA_MMA_AVAILABLE
 #endif // defined(AMD_WMMA_AVAILABLE)
-    if constexpr (!supported) {
+    if constexpr (!tile_A::supported() || !tile_B::supported() || !tile_C::supported()) {
         NO_DEVICE_CODE;
         return;
     }
@@ -248,6 +244,9 @@ static __global__ void mul_mat_f(
             }
         }
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    }
+#endif //VOLTA_MMA_AVAILABLE
 #else
     GGML_UNUSED_VARS(x, y, ids, dst,
         ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
@@ -278,27 +277,24 @@ static __global__ void mul_mat_f_ids(
     typedef tile<16,       8, T>     tile_A;
     typedef tile<tile_B_I, 8, T>     tile_B;
     typedef tile<16,       tile_C_J, float> tile_C;
-
-    constexpr bool a_supported = tile_A::supported();
-    constexpr bool b_supported = tile_B::supported();
-    constexpr bool c_supported = tile_C::supported();
-    constexpr bool supported = a_supported && b_supported && c_supported;
 #else
-    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
-    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
-    constexpr bool supported = I_16_supported || I_32_supported;
-
-    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;
+#ifdef VOLTA_MMA_AVAILABLE
+    if constexpr (!std::is_same_v<T, half2>) {NO_DEVICE_CODE;} else {
+    typedef tile<32, 4, T,     DATA_LAYOUT_I_MAJOR>          tile_A;
+    typedef tile< 8, 4, T,     DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B;
+    typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR>          tile_C;
+#else
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile<8,  8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+#endif // VOLTA_MMA_AVAILABLE
 #endif // defined(AMD_WMMA_AVAILABLE)
-    if constexpr (!supported) {
+    if constexpr (!tile_A::supported() || !tile_B::supported() || !tile_C::supported()) {
         NO_DEVICE_CODE;
         return;
     }
 
+
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
     constexpr int ntA = rows_per_block / tile_A::I;
@@ -517,6 +513,9 @@ static __global__ void mul_mat_f_ids(
             }
         }
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    }
+#endif // VOLTA_MMA_AVAILABLE
 #else
     GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
         ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,

ggml/src/ggml-cuda/mmq.cu

@@ -307,10 +307,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
     }
 
     if (amd_wmma_available(cc)) {
-        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
-            return true;
-        }
+        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_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+
 }

ggml/src/ggml-cuda/mmq.cuh

@@ -1542,8 +1542,10 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
             tile_C Cm;
             if (k01 >= MMQ_TILE_NE_K * 3/4) {
                 tile_A A1;
-                A1.x[0] = 0x01010101;
-                A1.x[1] = 0x01010101;
+#pragma unroll
+                for (int l = 0; l < tile_A::ne; ++l) {
+                    A1.x[l] = 0x01010101;
+                }
                 mma(Cm, A1, B);
             }