CANN: Improve loading efficiency after converting weights to NZ format. (#14985)

* CANN: Improve loading efficiency after converting weights to NZ format.

* CANN: fix typo

.clang-format

@@ -22,8 +22,8 @@ AllowShortIfStatementsOnASingleLine: Never
 AllowShortLambdasOnASingleLine: Inline
 AllowShortLoopsOnASingleLine: false
 AlwaysBreakBeforeMultilineStrings: true
-BinPackArguments: true
-BinPackParameters: true # OnePerLine
+BinPackArguments: false
+BinPackParameters: false # OnePerLine
 BitFieldColonSpacing: Both
 BreakBeforeBraces: Custom # Attach
 BraceWrapping:
@@ -70,15 +70,18 @@ ExperimentalAutoDetectBinPacking: false
 FixNamespaceComments: true
 IncludeBlocks:   Regroup
 IncludeCategories:
-  - Regex:           '^<.*\.h>'
+  - Regex:           '".*"'
     Priority:        1
     SortPriority:    0
-  - Regex:           '^<.*'
+  - Regex:           '^<.*\.h>'
     Priority:        2
     SortPriority:    0
-  - Regex:           '.*'
+  - Regex:           '^<.*'
     Priority:        3
     SortPriority:    0
+  - Regex:           '.*'
+    Priority:        4
+    SortPriority:    0
 IncludeIsMainRegex: '([-_](test|unittest))?$'
 IncludeIsMainSourceRegex: ''
 IndentAccessModifiers: false

.devops/musa.Dockerfile

@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc4.0.1
+ARG MUSA_VERSION=rc4.2.0
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64
 
-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 

.devops/nix/package.nix

@@ -47,6 +47,7 @@ let
   inherit (lib)
     cmakeBool
     cmakeFeature
+    optionalAttrs
     optionals
     strings
     ;
@@ -197,7 +198,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     ];
 
   # Environment variables needed for ROCm
-  env = optionals useRocm {
+  env = optionalAttrs useRocm {
     ROCM_PATH = "${rocmPackages.clr}";
     HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
   };

.devops/rocm.Dockerfile

@@ -1,8 +1,8 @@
 ARG UBUNTU_VERSION=24.04
 
 # This needs to generally match the container host's environment.
-ARG ROCM_VERSION=6.3
-ARG AMDGPU_VERSION=6.3
+ARG ROCM_VERSION=6.4
+ARG AMDGPU_VERSION=6.4
 
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete

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

@@ -48,98 +48,98 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-24-riscv64-vulkan-cross:
-    runs-on: ubuntu-24.04
+  # ubuntu-24-riscv64-vulkan-cross:
+  #   runs-on: ubuntu-24.04
 
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo dpkg --add-architecture riscv64
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup Riscv
+  #       run: |
+  #         sudo dpkg --add-architecture riscv64
 
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
 
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu \
-                  libvulkan-dev:riscv64
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 glslc \
+  #                 gcc-14-riscv64-linux-gnu \
+  #                 g++-14-riscv64-linux-gnu \
+  #                 libvulkan-dev:riscv64
 
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_VULKAN=ON \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+  #                        -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+  #                        -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-          cmake --build build --config Release -j $(nproc)
+  #         cmake --build build --config Release -j $(nproc)
 
-  ubuntu-24-arm64-vulkan-cross:
-    runs-on: ubuntu-24.04
+  # ubuntu-24-arm64-vulkan-cross:
+  #   runs-on: ubuntu-24.04
 
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup Arm64
-        run: |
-          sudo dpkg --add-architecture arm64
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup Arm64
+  #       run: |
+  #         sudo dpkg --add-architecture arm64
 
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
+  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
 
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  crossbuild-essential-arm64 \
-                  libvulkan-dev:arm64
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 glslc \
+  #                 crossbuild-essential-arm64 \
+  #                 libvulkan-dev:arm64
 
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
-                         -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
-                         -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_VULKAN=ON \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
+  #                        -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
+  #                        -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-          cmake --build build --config Release -j $(nproc)
+  #         cmake --build build --config Release -j $(nproc)
 
   ubuntu-24-ppc64el-cpu-cross:
     runs-on: ubuntu-24.04
@@ -185,52 +185,52 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-24-ppc64el-vulkan-cross:
-    runs-on: ubuntu-24.04
+  # ubuntu-24-ppc64el-vulkan-cross:
+  #   runs-on: ubuntu-24.04
 
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup PowerPC64le
-        run: |
-          sudo dpkg --add-architecture ppc64el
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup PowerPC64le
+  #       run: |
+  #         sudo dpkg --add-architecture ppc64el
 
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
 
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  gcc-14-powerpc64le-linux-gnu \
-                  g++-14-powerpc64le-linux-gnu \
-                  libvulkan-dev:ppc64el
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 glslc \
+  #                 gcc-14-powerpc64le-linux-gnu \
+  #                 g++-14-powerpc64le-linux-gnu \
+  #                 libvulkan-dev:ppc64el
 
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_VULKAN=ON \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+  #                        -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+  #                        -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-          cmake --build build --config Release -j $(nproc)
+  #         cmake --build build --config Release -j $(nproc)
 
   debian-13-loongarch64-cpu-cross:
     runs-on: ubuntu-24.04

.github/workflows/build.yml

@@ -135,6 +135,69 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
+  macOS-latest-cmake-arm64-webgpu:
+    runs-on: macos-14
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-cmake-arm64-webgpu
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+          brew install curl
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          ARTIFACTS_JSON=$(curl -s -L \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
+            -H "X-GitHub-Api-Version: 2022-11-28" \
+            "https://api.github.com/repos/google/dawn/actions/artifacts")
+          echo "Finding latest macos-latest-Release artifact..."
+          DOWNLOAD_URL=$(echo "$ARTIFACTS_JSON" | jq -r '.artifacts
+            | sort_by(.created_at)
+            | reverse
+            | map(select(.name | test("macos-latest-Release$")))
+            | .[0].archive_download_url')
+          if [ "$DOWNLOAD_URL" = "null" ] || [ -z "$DOWNLOAD_URL" ]; then
+            echo "No suitable Dawn artifact found!"
+            exit 1
+          fi
+          echo "Downloading from: $DOWNLOAD_URL"
+          curl -L \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
+            -o artifact.zip "$DOWNLOAD_URL"
+          unzip artifact.zip
+          mkdir dawn
+          tar_file=$(find . -name '*.tar.gz' | head -n 1)
+          echo "Extracting: $tar_file"
+          tar -xvf "$tar_file" -C dawn --strip-components=1
+
+      - name: Build
+        id: cmake_build
+        run: |
+          export CMAKE_PREFIX_PATH=dawn
+          cmake -B build -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
   ubuntu-cpu-cmake:
     strategy:
       matrix:
@@ -344,6 +407,72 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 4200
 
+  ubuntu-22-cmake-webgpu:
+    runs-on: ubuntu-22.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: ubuntu-22-cmake-webgpu
+          evict-old-files: 1d
+
+      - name: Vulkan SDK Dependencies
+        id: vulkan-depends
+        run: |
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
+          ARTIFACTS_JSON=$(curl -s -L \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
+            -H "X-GitHub-Api-Version: 2022-11-28" \
+            "https://api.github.com/repos/google/dawn/actions/artifacts")
+          echo "Finding latest ubuntu-latest-Release artifact..."
+          DOWNLOAD_URL=$(echo "$ARTIFACTS_JSON" | jq -r '.artifacts
+            | sort_by(.created_at)
+            | reverse
+            | map(select(.name | test("ubuntu-latest-Release$")))
+            | .[0].archive_download_url')
+          if [ "$DOWNLOAD_URL" = "null" ] || [ -z "$DOWNLOAD_URL" ]; then
+            echo "No suitable Dawn artifact found!"
+            exit 1
+          fi
+          echo "Downloading from: $DOWNLOAD_URL"
+          curl -L \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
+            -o artifact.zip "$DOWNLOAD_URL"
+          unzip artifact.zip
+          mkdir dawn
+          tar_file=$(find . -name '*.tar.gz' | head -n 1)
+          echo "Extracting: $tar_file"
+          tar -xvf "$tar_file" -C dawn --strip-components=1
+
+      - name: Build
+        id: cmake_build
+        run: |
+          export Dawn_DIR=dawn/lib64/cmake/Dawn
+          cmake -B build -DGGML_WEBGPU=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          # This is using llvmpipe and runs slower than other backends
+          ctest -L main --verbose --timeout 3600
+
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.0.2
@@ -386,7 +515,7 @@ jobs:
 
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
-    container: mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
+    container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
 
     steps:
       - name: Clone

.github/workflows/close-issue.yml

@@ -17,7 +17,7 @@ jobs:
     steps:
       - uses: actions/stale@v5
         with:
-          exempt-issue-labels: "refactor,help wanted,good first issue,research,bug,roadmap"
+          exempt-issue-labels: "refactoring,help wanted,good first issue,research,bug,roadmap"
           days-before-issue-stale: 30
           days-before-issue-close: 14
           stale-issue-label: "stale"

.gitignore

@@ -82,6 +82,7 @@ models/*
 models-mnt
 !models/.editorconfig
 !models/ggml-vocab-*.gguf*
+!models/templates
 
 # Zig
 zig-out/

CODEOWNERS

@@ -9,3 +9,4 @@
 /ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
 /ggml/src/ggml-opt.cpp @JohannesGaessler
 /ggml/src/gguf.cpp @JohannesGaessler
+/ggml/src/ggml-vulkan/ @0cc4m

README.md

@@ -269,6 +269,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
+| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
 ## Obtaining and quantizing models
@@ -434,7 +435,7 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
 ## [`llama-perplexity`](tools/perplexity)
 
-#### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.
+#### A tool for measuring the [perplexity](tools/perplexity/README.md) [^1] (and other quality metrics) of a model over a given text.
 
 - <details open>
     <summary>Measure the perplexity over a text file</summary>
@@ -457,8 +458,7 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
     </details>
 
-[^1]: [tools/perplexity/README.md](./tools/perplexity/README.md)
-[^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
+[^1]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
 
 ## [`llama-bench`](tools/llama-bench)
 

ci/README.md

@@ -54,7 +54,7 @@ docker run --privileged -it \
     -v $HOME/llama.cpp/ci-cache:/ci-cache \
     -v $HOME/llama.cpp/ci-results:/ci-results \
     -v $PWD:/ws -w /ws \
-    mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
+    mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
 ```
 
 Inside the container, execute the following commands:

ci/run.sh

@@ -16,6 +16,9 @@
 # # with VULKAN support
 # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with WebGPU support
+# GG_BUILD_WEBGPU=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 # # with MUSA support
 # GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
@@ -81,6 +84,10 @@ if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 fi
 
+if [ ! -z ${GG_BUILD_WEBGPU} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
+fi
+
 if [ ! -z ${GG_BUILD_MUSA} ]; then
     # Use qy1 by default (MTT S80)
     MUSA_ARCH=${MUSA_ARCH:-21}

common/arg.cpp

@@ -1464,6 +1464,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.swa_full = true;
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
+    add_opt(common_arg(
+        {"--kv-unified", "-kvu"},
+        string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.kv_unified ? "true" : "false"),
+        [](common_params & params) {
+            params.kv_unified = true;
+        }
+    ).set_env("LLAMA_ARG_KV_SPLIT"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1604,7 +1612,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.antiprompt.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-sp", "--special"},
         string_format("special tokens output enabled (default: %s)", params.special ? "true" : "false"),
@@ -2647,6 +2655,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.i_chunk = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--show-statistics"},
+        string_format("show imatrix statistics and then exit (default: %s)", params.show_statistics ? "true" : "false"),
+        [](common_params & params) {
+            params.show_statistics = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--parse-special"},
         string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
@@ -3423,5 +3438,34 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
+    // diffusion parameters
+    add_opt(common_arg(
+        { "--diffusion-steps" }, "N",
+        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
+        [](common_params & params, int value) { params.diffusion.steps = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-eps" }, "F",
+        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
+        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-algorithm" }, "N",
+        string_format("diffusion algorithm: 0=ORIGIN, 1=MASKGIT_PLUS, 2=TOPK_MARGIN, 3=ENTROPY (default: %d)",
+                      params.diffusion.algorithm),
+        [](common_params & params, int value) { params.diffusion.algorithm = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-alg-temp" }, "F",
+        string_format("algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
+        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-visual" },
+        string_format("enable visual diffusion mode (show progressive generation) (default: %s)",
+                      params.diffusion.visual_mode ? "true" : "false"),
+        [](common_params & params) { params.diffusion.visual_mode = true; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+
     return ctx_arg;
 }

common/chat.cpp

@@ -1944,6 +1944,8 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
         }
     }
     auto msg = builder.result();
-    LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+    if (!is_partial) {
+        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+    }
     return msg;
 }

common/common.cpp

@@ -448,6 +448,15 @@ void string_replace_all(std::string & s, const std::string & search, const std::
 bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
     return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
 }
+
+bool string_remove_suffix(std::string & str, const std::string_view & suffix) {
+    bool has_suffix = string_ends_with(str, suffix);
+    if (has_suffix) {
+        str = str.substr(0, str.size() - suffix.size());
+    }
+    return has_suffix;
+}
+
 size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
     if (!str.empty() && !stop.empty()) {
         const char text_last_char = str.back();
@@ -1005,15 +1014,21 @@ struct common_init_result common_init_from_params(common_params & params) {
         params.sampling.ignore_eos = false;
     }
 
-    if (params.sampling.ignore_eos) {
-        for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-            if (llama_vocab_is_eog(vocab, i)) {
-                LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-                params.sampling.logit_bias.push_back({i, -INFINITY});
-            }
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
         }
     }
 
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
     if (params.sampling.penalty_last_n == -1) {
         LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
         params.sampling.penalty_last_n = llama_n_ctx(lctx);
@@ -1157,6 +1172,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
+    cparams.kv_unified        = params.kv_unified;
 
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;

common/common.h

@@ -81,6 +81,7 @@ enum llama_example {
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
+    LLAMA_EXAMPLE_DIFFUSION,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -177,7 +178,8 @@ struct common_params_sampling {
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
 
-    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
+    std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
+    std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
     // print the parameters into a string
     std::string print() const;
@@ -217,6 +219,14 @@ struct common_params_vocoder {
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
 
+struct common_params_diffusion {
+    int32_t steps       = 64;     // number of diffusion steps
+    float   eps         = 1e-3f;  // epsilon for timesteps
+    int32_t algorithm   = 0;      // diffusion algorithm (0=ORIGIN, 1=MASKGIT_PLUS, 2=TOPK_MARGIN, 3=ENTROPY)
+    float   alg_temp    = 0.0f;   // algorithm temperature
+    bool    visual_mode = false;  // show progressive diffusion on screen
+};
+
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
@@ -268,6 +278,7 @@ struct common_params {
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
+    struct common_params_diffusion   diffusion;
 
     struct common_params_model model;
 
@@ -330,6 +341,7 @@ struct common_params {
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+    bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads
@@ -420,9 +432,10 @@ struct common_params {
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
 
-    bool process_output = false; // collect data for the output tensor
-    bool compute_ppl    = true;  // whether to compute perplexity
-    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
+    bool process_output  = false; // collect data for the output tensor
+    bool compute_ppl     = true;  // whether to compute perplexity
+    bool show_statistics = false; // show imatrix statistics per tensor
+    bool parse_special   = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;
@@ -522,6 +535,7 @@ static bool string_starts_with(const std::string & str,
 
 // While we wait for C++20's std::string::ends_with...
 bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+bool string_remove_suffix(std::string & str, const std::string_view & suffix);
 size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);

convert_hf_to_gguf.py

@@ -669,6 +669,36 @@ class TextModel(ModelBase):
         # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
         #       or pull the latest version of the model from Huggingface
         #       don't edit the hashes manually!
+        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"
+        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
+            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
+            res = "minerva-7b"
+        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
+            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
+            res = "hunyuan"
+        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
+            res = "falcon-h1"
+        if chkhsh == "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-1B-Base
+            res = "falcon-h1"
+        if chkhsh == "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-7B-Base
+            res = "falcon-h1"
+        if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
+            res = "falcon-h1"
+        if chkhsh == "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890":
+            # ref: https://huggingface.co/moonshotai/Kimi-K2-Base
+            res = "kimi-k2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -804,45 +834,18 @@ class TextModel(ModelBase):
         if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
             # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
             res = "seed-coder"
-        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
-        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
-        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
-            res = "glm4"
-        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
-            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
-            res = "minerva-7b"
-        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
-            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
-            res = "hunyuan"
         if chkhsh == "b0a6b1c0bd5998ebd9df08611efde34a4ff03faed45ae09c43e6b31ebd4b94cf":
             # ref: https://huggingface.co/skt/A.X-4.0
             res = "a.x-4.0"
-        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
-            res = "falcon-h1"
-        if chkhsh == "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-1B-Base
-            res = "falcon-h1"
-        if chkhsh == "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-7B-Base
-            res = "falcon-h1"
-        if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
-            res = "falcon-h1"
         if chkhsh == "f6791d196f87ce6b56a7d234be618e0d58f8cda3549416635b2bebcd22cd95c4":
             # ref: https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct
             res = "midm-2.0"
         if chkhsh == "169bf0296a13c4d9b7672313f749eb36501d931022de052aad6e36f2bf34dd51":
             # ref: https://huggingface.co/LiquidAI/LFM2-Tokenizer
             res = "lfm2"
-        if chkhsh == "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890":
-            # ref: https://huggingface.co/moonshotai/Kimi-K2-Base
-            res = "kimi-k2"
+        if chkhsh == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb":
+            # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B
+            res = "exaone4"
 
         if res is None:
             logger.warning("\n")
@@ -1897,6 +1900,7 @@ class StableLMModel(TextModel):
     "MixtralForCausalLM",
     "VLlama3ForCausalLM",
     "LlavaForConditionalGeneration",
+    "VoxtralForConditionalGeneration",
     "LlamaModel")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
@@ -1909,6 +1913,11 @@ class LlamaModel(TextModel):
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
 
     def set_vocab(self):
+        path_tekken_json = self.dir_model / "tekken.json"
+        path_tokenizer_json = self.dir_model / "tokenizer.json"
+        if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
+            return self.set_vocab_tekken()
+
         try:
             self._set_vocab_sentencepiece()
         except FileNotFoundError:
@@ -1941,6 +1950,52 @@ class LlamaModel(TextModel):
         if self.hparams.get("vocab_size", 32000) == 49152:
             self.gguf_writer.add_add_bos_token(False)
 
+    def set_vocab_tekken(self):
+        vocab = gguf.vocab.MistralVocab(self.dir_model)
+        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 == gguf.vocab.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)
+
+        script_dir = Path(__file__).parent
+        template_path = script_dir / "models/templates/unsloth-mistral-Devstral-Small-2507.jinja"
+        with open(template_path, "r", encoding="utf-8") as f:
+            template = f.read()
+            self.gguf_writer.add_chat_template(template)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -1968,12 +2023,13 @@ class LlamaModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
-        is_vision_tensor = "vision_tower" in name \
+        is_multimodal_tensor = "vision_tower" in name \
             or "vision_model" in name \
+            or "audio_tower" in name \
             or "model.connector" in name \
             or "multi_modal_projector" in name
 
-        if is_vision_tensor:
+        if is_multimodal_tensor:
             return [] # skip vision tensors
         elif self.hf_arch == "LlamaModel":
             name = "model." + name
@@ -2778,6 +2834,76 @@ class Qwen2Model(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("DreamModel")
+class DreamModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.DREAM
+
+    def get_vocab_base(self) -> tuple[list[str], list[int], str]:
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+
+        vocab_dict = tokenizer.get_vocab()
+        vocab_size = self.hparams.get("vocab_size", len(vocab_dict))
+        assert max(vocab_dict.values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab_dict.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                # Check if it's a special token - treat special tokens as CONTROL tokens
+                if hasattr(tokenizer, 'added_tokens_decoder') and i in tokenizer.added_tokens_decoder:
+                    if tokenizer.added_tokens_decoder[i].special:
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    # Fallback: treat all added vocab as control tokens for special tokens like <|im_start|>
+                    toktypes.append(gguf.TokenType.CONTROL)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        return tokens, toktypes, tokpre
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self._try_set_pooling_type()
+
+        # Dream models use non-causal attention for diffusion
+        self.gguf_writer.add_causal_attention(False)
+        # Handle RoPE scaling similar to Qwen2
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            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"])
+
+        # Add Dream-specific parameters
+        mask_token_id = self.hparams.get("mask_token_id")
+        if mask_token_id is not None:
+            self.gguf_writer.add_mask_token_id(mask_token_id)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Dream model tensors should be mapped directly since it's the base model
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Ernie4_5_ForCausalLM")
 class Ernie4_5Model(TextModel):
     model_arch = gguf.MODEL_ARCH.ERNIE4_5
@@ -2791,7 +2917,8 @@ class Ernie4_5Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
         num_kv_heads = self.hparams["num_key_value_heads"]
-        head_dim = self.hparams["head_dim"]
+        if (head_dim := self.hparams.get("head_dim")) is None:
+            head_dim = self.hparams["hidden_size"] // num_heads
 
         if "ernie." in name:
             name = name.replace("ernie.", "model.")
@@ -2824,6 +2951,93 @@ class Ernie4_5Model(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Ernie4_5_MoeForCausalLM")
+class Ernie4_5MoeModel(Ernie4_5Model):
+    model_arch = gguf.MODEL_ARCH.ERNIE4_5_MOE
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._experts = [{} for _ in range(self.block_count)]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"])
+        self.gguf_writer.add_expert_used_count(self.hparams["moe_k"])
+        self.gguf_writer.add_interleave_moe_layer_step(self.hparams["moe_layer_interval"])
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["moe_layer_start_index"])
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        if (shared_expert_count := self.hparams.get('moe_num_shared_experts')) is not None:
+            self.gguf_writer.add_expert_shared_count(shared_expert_count)
+            if shared_expert_count > 0 and (shared_expert_intermediate_size := self.hparams.get('intermediate_size')) is not None and (num_key_value_heads := self.hparams.get('num_key_value_heads')) is not None:
+                self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size // num_key_value_heads)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Modify correction bias name as in DeepseekV2
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        # skip Multi-Token Prediction (MTP) layers (again, same as DeepseekV2)
+        match = re.match(r"model.mtp_block.(\d+)", name)
+        if match:
+            return []
+
+        # skip all other MTP tensors for now
+        match = re.match(r"model.mtp_emb_norm.(\d+)", name)
+        if match:
+            return []
+
+        match = re.match(r"model.mtp_hidden_norm.(\d+)", name)
+        if match:
+            return []
+
+        match = re.match(r"model.mtp_linear_proj.(\d+)", name)
+        if match:
+            return []
+
+        # process the experts separately
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["moe_num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["gate_proj", "up_proj", "down_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename_to_retrieve])
+                        del self._experts[bid][ename_to_retrieve]
+
+                    data_torch = torch.stack(datas, dim=0)
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+                    new_name = self.map_tensor_name(merged_name)
+                    tensors.append((new_name, data_torch))
+
+                return tensors
+            else:
+                return []
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register(
     "Qwen2VLModel",
     "Qwen2VLForConditionalGeneration",
@@ -3630,7 +3844,7 @@ class Plamo2Model(TextModel):
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 32))
         self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
-        self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1000000.0))
+        self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 10000))
 
         # Mamba parameters
         self.gguf_writer.add_ssm_state_size(hparams.get("mamba_d_state", 64))
@@ -3641,7 +3855,7 @@ class Plamo2Model(TextModel):
         self.gguf_writer.add_ssm_group_count(0)
 
         # MLP feed forward parameters (for attention layers)
-        self.gguf_writer.add_feed_forward_length(hparams.get("intermediate_size", 16384))
+        self.gguf_writer.add_feed_forward_length(hparams.get("intermediate_size", 13312))
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
@@ -6325,7 +6539,7 @@ class JaisModel(TextModel):
         self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
 
 
-@ModelBase.register("Glm4ForCausalLM")
+@ModelBase.register("Glm4ForCausalLM", "Glm4vForConditionalGeneration")
 class Glm4Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4
 
@@ -6347,7 +6561,8 @@ class Glm4Model(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        rope_dim = self.hparams["head_dim"]
+        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", 0.5)))
         rope_scaling = self.hparams.get("rope_scaling") or {}
         if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
@@ -6355,6 +6570,13 @@ class Glm4Model(TextModel):
             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"])
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("model.visual."): # ignore visual part of Glm4v
+            return []
+        elif name.startswith("model.language_model."):
+            name = name.replace("language_model.", "") # for Glm4v
+        return super().modify_tensors(data_torch, name, bid)
+
 
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(TextModel):
@@ -6622,6 +6844,75 @@ class ExaoneModel(TextModel):
                 yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
 
 
+@ModelBase.register("Exaone4ForCausalLM")
+class Exaone4Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.EXAONE4
+
+    def set_vocab(self):
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        if hparams.get("sliding_window") is not None:
+            self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+            if "layer_types" in hparams:
+                self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in hparams["layer_types"]])
+            elif "sliding_window_pattern" in hparams:
+                sliding_window_pattern = []
+                if isinstance(hparams["sliding_window_pattern"], str):  # e.g. LLLG
+                    for i in range(hparams["num_hidden_layers"]):
+                        sliding_window_pattern.append(hparams["sliding_window_pattern"][i % len(hparams["sliding_window_pattern"])] == "L")
+                if isinstance(hparams["sliding_window_pattern"], int):  # e.g. 4
+                    for i in range(hparams["num_hidden_layers"]):
+                        sliding_window_pattern.append((i + 1) % hparams["sliding_window_pattern"] != 0)
+                if len(sliding_window_pattern) == hparams["num_hidden_layers"]:
+                    self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
+
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
+            if rope_scaling.get("rope_type", '').lower() == "llama3":
+                base = self.hparams.get("rope_theta", 10_000.0)
+                if (dim := self.hparams.get("head_dim")) is None:
+                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+
+                factor = rope_scaling.get("factor", 16.0)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                        rope_factors.append(1 / ((1 - smooth) / factor + smooth))
+
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
+
+
 @ModelBase.register("GraniteForCausalLM")
 class GraniteModel(LlamaModel):
     """Conversion for IBM's GraniteForCausalLM"""
@@ -6993,9 +7284,10 @@ class WhisperEncoderModel(MmprojModel):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.hparams["hidden_size"] = self.hparams["d_model"]
-        self.hparams["intermediate_size"] = self.hparams["encoder_ffn_dim"]
-        self.hparams["num_attention_heads"] = self.hparams["encoder_attention_heads"]
+        if "hidden_size" not in self.hparams and "intermediate_size" not in self.hparams:
+            self.hparams["hidden_size"] = self.hparams["d_model"]
+            self.hparams["intermediate_size"] = self.hparams["encoder_ffn_dim"]
+            self.hparams["num_attention_heads"] = self.hparams["encoder_attention_heads"]
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -7034,9 +7326,21 @@ class UltravoxWhisperEncoderModel(WhisperEncoderModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.ULTRAVOX)
         self.gguf_writer.add_audio_stack_factor(self.global_config["stack_factor"])
 
 
+@ModelBase.register("VoxtralForConditionalGeneration")
+class VoxtralWhisperEncoderModel(WhisperEncoderModel):
+    has_vision_encoder = False # no vision encoder
+    has_audio_encoder = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.VOXTRAL)
+        self.gguf_writer.add_audio_stack_factor(4) # == intermediate_size // hidden_size
+
+
 @ModelBase.register("FalconH1ForCausalLM")
 class FalconH1Model(Mamba2Model):
     model_arch = gguf.MODEL_ARCH.FALCON_H1
@@ -7351,6 +7655,88 @@ class LFM2Model(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("SmallThinkerForCausalLM")
+class SmallThinkerModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.SMALLTHINKER
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (n_experts := self.hparams.get("num_experts", self.hparams.get("moe_num_primary_experts"))) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (n_experts_used := self.hparams.get("num_experts_per_tok", self.hparams.get("moe_num_active_primary_experts"))) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+        if (moe_intermediate_size := self.hparams.get("moe_ffn_hidden_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+            self.gguf_writer.add_feed_forward_length(moe_intermediate_size)
+            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
+        if (self.hparams.get('moe_primary_router_apply_softmax')):
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        else:
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        # YaRN is not enabled by default
+        # To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            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"])
+
+        sliding_window_layout = self.hparams.get("sliding_window_layout")
+        if sliding_window_layout:
+            for i in sliding_window_layout:
+                if i != 0:
+                    sliding_window = self.hparams.get("sliding_window_size")
+                    if sliding_window:
+                        self.gguf_writer.add_sliding_window(sliding_window)
+                    break
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("experts") != -1:
+            n_experts = self.hparams.get("num_experts", self.hparams.get("moe_num_primary_experts"))
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down", "gate", "up"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -7,7 +7,6 @@ import pathlib
 import re
 
 import requests
-import sys
 import json
 import shutil
 import argparse
@@ -69,8 +68,7 @@ args = parser.parse_args()
 hf_token = args.hf_token if args.hf_token is not None else hf_token
 
 if hf_token is None:
-    logger.error("HF token is required. Please provide it as an argument or set it in ~/.cache/huggingface/token")
-    sys.exit(1)
+    logger.warning("HF token not found. You can provide it as an argument or set it in ~/.cache/huggingface/token")
 
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
@@ -131,6 +129,7 @@ models = [
     {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
     {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
+    {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -151,7 +150,7 @@ pre_computed_hashes = [
 
 
 def download_file_with_auth(url, token, save_path):
-    headers = {"Authorization": f"Bearer {token}"}
+    headers = {"Authorization": f"Bearer {token}"} if token else None
     response = sess.get(url, headers=headers)
     response.raise_for_status()
     os.makedirs(os.path.dirname(save_path), exist_ok=True)
@@ -232,7 +231,7 @@ for model in models:
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
-for model in [*all_models, *pre_computed_hashes]:
+for model in [*pre_computed_hashes, *all_models]:
     name = model["name"]
     tokt = model["tokt"]
     chkhsh = model.get("chkhsh")
@@ -240,11 +239,6 @@ for model in [*all_models, *pre_computed_hashes]:
     if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
 
-    # Skip if the tokenizer folder does not exist or there are other download issues previously
-    if not os.path.exists(f"models/tokenizers/{name}"):
-        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
-        continue
-
     # create the tokenizer
     if chkhsh is not None:
         # if the model has a pre-computed hash, use it
@@ -254,15 +248,19 @@ for model in [*all_models, *pre_computed_hashes]:
         chkhsh = existing_models[name]
     else:
         # otherwise, compute the hash of the tokenizer
+
+        # Fail if the tokenizer folder with config does not exist or there are other download issues previously
+        if not os.path.isfile(f"models/tokenizers/{name}/tokenizer_config.json"):
+            raise OSError(f"Config for tokenizer {name} not found. The model may not exist or is not accessible with the provided token.")
+
         try:
             logger.info(f"Loading tokenizer from {f'models/tokenizers/{name}'}...")
             if name == "t5":
                 tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
             else:
                 tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-        except OSError as e:
-            logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
-            continue  # Skip to the next model if the tokenizer can't be loaded
+        except Exception as e:
+            raise OSError(f"Error loading tokenizer for model {name}.") from e
 
         chktok = tokenizer.encode(CHK_TXT)
         chkhsh = sha256(str(chktok).encode()).hexdigest()

docs/backend/CANN.md

@@ -310,5 +310,7 @@ Specifies the memory pool management strategy:
 
 Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
 
-## TODO
-- Support more models and data types.
+### GGML_CANN_WEIGHT_NZ
+
+Converting the matmul weight format from ND to NZ can significantly improve performance on the 310I DUO NPU.
+

docs/build-s390x.md

@@ -42,14 +42,14 @@ cmake --build build --config Release -j $(nproc)
     cmake --build build --config Release -j $(nproc)
     ```
 
--   By default, NNPA is enabled when available. To disable it (not recommended):
+-   By default, NNPA is disabled by default. To enable it:
 
     ```bash
     cmake -S . -B build             \
         -DCMAKE_BUILD_TYPE=Release  \
         -DGGML_BLAS=ON              \
         -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_NNPA=OFF
+        -DGGML_NNPA=ON
 
     cmake --build build --config Release -j $(nproc)
     ```
@@ -84,9 +84,9 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
     ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
 
-    You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
+    You can find popular models pre-converted and verified at [s390x Verified Models](https://huggingface.co/collections/taronaeo/s390x-verified-models-672765393af438d0ccb72a08) or [s390x Runnable Models](https://huggingface.co/collections/taronaeo/s390x-runnable-models-686e951824198df12416017e).
 
-    These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
+    These models have already been converted from `safetensors` to `GGUF` Big-Endian and their respective tokenizers verified to run correctly on IBM z15 and later system.
 
 2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
 
@@ -94,6 +94,14 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
     The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
 
+    Ensure that you have installed the required packages in advance
+
+    ```bash
+    pip3 install -r requirements.txt
+    ```
+
+    Convert the `safetensors` model to `GGUF`
+
     ```bash
     python3 convert_hf_to_gguf.py \
         --outfile model-name-be.f16.gguf \
@@ -116,7 +124,7 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
     ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
 
-    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
+    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B GGUF](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
 
     ```bash
     python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
@@ -141,15 +149,15 @@ Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by
 
 ### 2. NNPA Vector Intrinsics Acceleration
 
-Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
+Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
 
 ### 3. zDNN Accelerator
 
-_Only available in IBM z16 or later system. No direction at the moment._
+_Only available in IBM z16 / LinuxONE 4 or later system. No support currently available._
 
 ### 4. Spyre Accelerator
 
-_No direction at the moment._
+_Only available with IBM z17 / LinuxONE 5 or later system. No support currently available._
 
 ## Performance Tuning
 
@@ -189,6 +197,26 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
     Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
 
+4. Failing to install the `sentencepiece` package using GCC 15+
+
+    Answer: The `sentencepiece` team are aware of this as seen in [this issue](https://github.com/google/sentencepiece/issues/1108).
+
+    As a temporary workaround, please run the installation command with the following environment variables.
+
+    ```bash
+    export CXXFLAGS="-include cstdint"
+    ```
+
+    For example,
+
+    ```bash
+    CXXFLAGS="-include cstdint" pip3 install -r requirements.txt
+    ```
+
+5. `-DGGML_NNPA=ON` generates gibberish output
+
+    Answer: We are aware of this as detailed in [this issue](https://github.com/ggml-org/llama.cpp/issues/14877). Please either try reducing the number of threads, or disable the compile option using `-DGGML_NNPA=OFF`.
+
 ## Getting Help on IBM Z & LinuxONE
 
 1. **Bugs, Feature Requests**
@@ -244,3 +272,5 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   ✅ - acceleration available
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
+
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 25, 2025.

docs/build.md

@@ -68,6 +68,9 @@ cmake --build build --config Release
       cmake --build build-x64-windows-llvm-release
       ```
 - Curl usage is enabled by default and can be turned off with `-DLLAMA_CURL=OFF`. Otherwise you need to install development libraries for libcurl.
+  - **Debian / Ubuntu:** `sudo apt-get install libcurl4-openssl-dev`  # (or `libcurl4-gnutls-dev` if you prefer GnuTLS)
+  - **Fedora / RHEL / Rocky / Alma:** `sudo dnf install libcurl-devel`
+  - **Arch / Manjaro:** `sudo pacman -S curl`  # includes libcurl headers
 
 ## BLAS Build
 
@@ -305,9 +308,8 @@ On Linux it is possible to use unified memory architecture (UMA) to share main m
 
 ## Vulkan
 
-**Windows**
-
-### w64devkit
+### For Windows Users:
+**w64devkit**
 
 Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
 
@@ -334,7 +336,7 @@ cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
 ```
 
-### Git Bash MINGW64
+**Git Bash MINGW64**
 
 Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
 
@@ -357,7 +359,8 @@ Now you can load the model in conversation mode using `Vulkan`
 build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
 ```
 
-### MSYS2
+**MSYS2**
+
 Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
 ```sh
 pacman -S git \
@@ -373,9 +376,9 @@ cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
 ```
 
-**With docker**:
+### For Docker users:
 
-You don't need to install Vulkan SDK. It will be installed inside the container.
+You don't need to install the Vulkan SDK. It will be installed inside the container.
 
 ```sh
 # Build the image
@@ -385,32 +388,29 @@ docker build -t llama-cpp-vulkan --target light -f .devops/vulkan.Dockerfile .
 docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
 ```
 
-**Without docker**:
+### For Linux users:
 
-Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
+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.
 
-For example, on Ubuntu 22.04 (jammy), use the command below:
+> [!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.
 
+Second, after verifying that you have followed all of the SDK installation/setup steps, use this command to make sure before proceeding:
 ```bash
-wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
-wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-apt update -y
-apt-get install -y vulkan-sdk
-# To verify the installation, use the command below:
 vulkaninfo
 ```
 
-Alternatively your package manager might be able to provide the appropriate libraries.
-For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
-For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
-
-Then, build llama.cpp using the cmake command below:
-
+Then, assuming you have `cd` into your llama.cpp folder and there are no errors with running `vulkaninfo`, you can proceed to build llama.cpp using the CMake commands below:
 ```bash
 cmake -B build -DGGML_VULKAN=1
 cmake --build build --config Release
-# Test the output binary (with "-ngl 33" to offload all layers to GPU)
-./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+```
+
+Finally, after finishing your build, you should be able to do something like this:
+```bash
+# Test the output binary
+# "-ngl 99" should offload all of the layers to GPU for most (if not all) models.
+./build/bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -ngl 99
 
 # You should see in the output, ggml_vulkan detected your GPU. For example:
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
@@ -557,6 +557,23 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
+## WebGPU [In Progress]
+
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
+
+In the llama.cpp directory, build with CMake:
+
+```
+cmake -B build -DGGML_WEBGPU=ON
+cmake --build build --config Release
+```
+
+### Browser Support
+
+WebGPU allows cross-platform access to the GPU from supported browsers. We utilize [Emscripten](https://emscripten.org/) to compile ggml's WebGPU backend to WebAssembly. Emscripten does not officially support WebGPU bindings yet, but Dawn currently maintains its own WebGPU bindings called emdawnwebgpu.
+
+Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/src/emdawnwebgpu/) to download or build the emdawnwebgpu package (Note that it might be safer to build the emdawbwebgpu package locally, so that it stays in sync with the version of Dawn you have installed above). When building using CMake, the path to the emdawnwebgpu port file needs to be set with the flag `EMDAWNWEBGPU_DIR`.
+
 ## IBM Z & LinuxONE
 
 To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)

docs/development/HOWTO-add-model.md

@@ -23,11 +23,19 @@ The convert script reads the model configuration, tokenizer, tensor names+data a
 
 The required steps to implement for an HF model are:
 
-1. Define the model `Model.register` annotation in a new `Model` subclass, example:
+1. Define the model `ModelBase.register` annotation in a new `TextModel` or `MmprojModel` subclass, example:
 
 ```python
-@Model.register("MyModelForCausalLM")
-class MyModel(Model):
+@ModelBase.register("MyModelForCausalLM")
+class MyModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.MYMODEL
+```
+
+or
+
+```python
+@ModelBase.register("MyModelForConditionalGeneration")
+class MyModel(MmprojModel):
     model_arch = gguf.MODEL_ARCH.MYMODEL
 ```
 
@@ -75,9 +83,10 @@ block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
 `transformer.blocks.{bid}.norm_1` will be mapped to `blk.{bid}.attn_norm` in GGUF.
 
 Depending on the model configuration, tokenizer, code and tensors layout, you will have to override:
-- `Model#set_gguf_parameters`
-- `Model#set_vocab`
-- `Model#write_tensors`
+- `TextModel#set_gguf_parameters`
+- `MmprojModel#set_gguf_parameters`
+- `ModelBase#set_vocab`
+- `ModelBase#modify_tensors`
 
 NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the convention and several tools like `quantize` expect this to proceed the weights.
 

docs/docker.md

@@ -110,7 +110,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
 
 The defaults are:
 
-- `MUSA_VERSION` set to `rc4.0.1`
+- `MUSA_VERSION` set to `rc4.2.0`
 
 The resulting images, are essentially the same as the non-MUSA images:
 

docs/multimodal.md

@@ -97,6 +97,9 @@ NOTE: some models may require large context window, for example: `-c 8192`
 # Qwen2-Audio and SeaLLM-Audio
 # note: no pre-quantized GGUF this model, as they have very poor result
 # ref: https://github.com/ggml-org/llama.cpp/pull/13760
+
+# Mistral's Voxtral
+(tool_name) -hf ggml-org/Voxtral-Mini-3B-2507-GGUF
 ```
 
 **Mixed modalities**:

docs/ops.md

@@ -2,94 +2,101 @@
 
 List of GGML operations and backend support status.
 
+## How to add a backend to this table:
+
+1. Run `test-backend-ops support --output csv` with your backend name and redirect output to a csv file in `docs/ops/` (e.g., `docs/ops/CUDA.csv`)
+2. Regenerate `/docs/ops.md` via `./scripts/create_ops_docs.py`
+
 Legend:
 - ✅ Fully supported by this backend
 - 🟡 Partially supported by this backend
 - ❌ Not supported by this backend
 
-| Operation | BLAS | CPU | CUDA | Metal |
-|-----------|------|------|------|------|
-|                              ABS | ❌ | ✅ | 🟡 | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ |
-|                              ADD | ❌ | ✅ | ✅ | 🟡 |
-|                             ADD1 | ❌ | ✅ | ✅ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ |
-|                            CLAMP | ❌ | ✅ | ✅ | 🟡 |
-|                           CONCAT | ❌ | ✅ | 🟡 | ✅ |
-|                             CONT | ❌ | ✅ | 🟡 | ✅ |
-|                       CONV_2D_DW | ❌ | ✅ | ✅ | ❌ |
-|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ |
-|                CONV_TRANSPOSE_2D | ❌ | ✅ | ✅ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | 🟡 |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ❌ |
-|                              CPY | ❌ | 🟡 | 🟡 | 🟡 |
-|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ✅ | ✅ | ❌ |
-|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | 🟡 |
-|                              DIV | ❌ | ✅ | ✅ | 🟡 |
-|                              DUP | ❌ | ✅ | 🟡 | 🟡 |
-|                              ELU | ❌ | ✅ | ❌ | 🟡 |
-|                              EXP | ❌ | ✅ | 🟡 | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | ✅ | 🟡 | 🟡 |
-|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | 🟡 |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | 🟡 |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | 🟡 |
-|                             GELU | ❌ | ✅ | 🟡 | 🟡 |
-|                         GELU_ERF | ❌ | ✅ | 🟡 | 🟡 |
-|                       GELU_QUICK | ❌ | ✅ | 🟡 | 🟡 |
-|                         GET_ROWS | ❌ | ✅ | 🟡 | ✅ |
-|                    GET_ROWS_BACK | ❌ | 🟡 | 🟡 | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ |
-|                      HARDSIGMOID | ❌ | ✅ | 🟡 | ❌ |
-|                        HARDSWISH | ❌ | ✅ | 🟡 | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | 🟡 |
-|                          L2_NORM | ❌ | ✅ | ✅ | ✅ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ |
-|                              LOG | ❌ | ✅ | ✅ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ |
-|                              MUL | ❌ | ✅ | ✅ | 🟡 |
-|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       MUL_MAT_ID | ❌ | ✅ | ✅ | ✅ |
-|                              NEG | ❌ | ✅ | 🟡 | 🟡 |
-|                             NORM | ❌ | ✅ | ✅ | 🟡 |
-|                   OPT_STEP_ADAMW | ❌ | ✅ | ✅ | ❌ |
-|                         OUT_PROD | 🟡 | 🟡 | 🟡 | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ❌ | ✅ |
-|                          POOL_2D | ❌ | ✅ | ✅ | ✅ |
-|                            REGLU | ❌ | ✅ | ✅ | 🟡 |
-|                             RELU | ❌ | ✅ | 🟡 | 🟡 |
-|                           REPEAT | ❌ | ✅ | 🟡 | ✅ |
-|                      REPEAT_BACK | ❌ | ✅ | ✅ | ❌ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | 🟡 |
-|                    RMS_NORM_BACK | ❌ | ✅ | ✅ | ❌ |
-|                     RMS_NORM_MUL | ❌ | ✅ | ✅ | ✅ |
-|                             ROPE | ❌ | ✅ | ✅ | ✅ |
-|                        ROPE_BACK | ❌ | ✅ | ✅ | ❌ |
-|                        RWKV_WKV6 | ❌ | ✅ | ✅ | ✅ |
-|                        RWKV_WKV7 | ❌ | ✅ | ✅ | ✅ |
-|                            SCALE | ❌ | ✅ | ✅ | ✅ |
-|                              SET | ❌ | ✅ | ❌ | ✅ |
-|                         SET_ROWS | ❌ | 🟡 | ❌ | 🟡 |
-|                              SGN | ❌ | ✅ | 🟡 | ❌ |
-|                          SIGMOID | ❌ | ✅ | 🟡 | 🟡 |
-|                             SILU | ❌ | ✅ | 🟡 | 🟡 |
-|                        SILU_BACK | ❌ | ✅ | ✅ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | 🟡 |
-|                         SOFT_MAX | ❌ | ✅ | ✅ | ✅ |
-|                    SOFT_MAX_BACK | ❌ | 🟡 | 🟡 | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | 🟡 |
-|                             SQRT | ❌ | ✅ | ✅ | 🟡 |
-|                         SSM_CONV | ❌ | ✅ | ✅ | ✅ |
-|                         SSM_SCAN | ❌ | ✅ | ✅ | ✅ |
-|                             STEP | ❌ | ✅ | 🟡 | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | 🟡 |
-|                              SUM | ❌ | ✅ | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ |
-|                           SWIGLU | ❌ | ✅ | ✅ | 🟡 |
-|                             TANH | ❌ | ✅ | 🟡 | 🟡 |
-|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ |
-|                          UPSCALE | ❌ | ✅ | ✅ | 🟡 |
+| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan |
+|-----------|------|------|------|------|------|------|------|------|
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 |
+|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
+|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
+|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
+|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
+|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
+|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                         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 | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
+|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ |

examples/CMakeLists.txt

@@ -33,6 +33,7 @@ else()
     add_subdirectory(speculative-simple)
     add_subdirectory(gen-docs)
     add_subdirectory(training)
+    add_subdirectory(diffusion)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/diffusion/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-diffusion-cli)
+add_executable(${TARGET} diffusion-cli.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/diffusion/diffusion-cli.cpp

@@ -0,0 +1,507 @@
+#include "arg.h"
+#include "chat.h"
+#include "common.h"
+#include "llama.h"
+#include "log.h"
+
+#include <limits.h>
+#include <string>
+#include <vector>
+#include <algorithm>
+#include <cmath>
+#include <limits>
+#include <random>
+
+typedef bool (*diffusion_step_callback_t)(int32_t step,
+                                          int32_t total_steps,
+                                          const llama_token * tokens,
+                                          int32_t n_tokens,
+                                          void * user_data);
+
+enum diffusion_alg {
+    DIFFUSION_ALG_ORIGIN       = 0,
+    DIFFUSION_ALG_MASKGIT_PLUS = 1,
+    DIFFUSION_ALG_TOPK_MARGIN  = 2,
+    DIFFUSION_ALG_ENTROPY      = 3,
+};
+
+struct diffusion_params {
+    int32_t                   steps;
+    float                     eps;
+    float                     temperature;
+    float                     top_p;
+    int32_t                   top_k;
+    llama_token               mask_token_id;
+    enum diffusion_alg        algorithm;
+    float                     alg_temp;
+    diffusion_step_callback_t step_callback;
+    void *                    step_callback_user_data;
+    int32_t                   seed;
+};
+
+
+static diffusion_params diffusion_default_params() {
+    diffusion_params params        = {};
+    params.steps                   = 64;
+    params.eps                     = 1e-3f;
+    params.temperature             = 0.2f;
+    params.top_p                   = 0.95f;
+    params.top_k                   = 0;
+    params.mask_token_id           = LLAMA_TOKEN_NULL;
+    params.algorithm               = DIFFUSION_ALG_ORIGIN;
+    params.alg_temp                = 0.0f;
+    params.step_callback           = nullptr;
+    params.step_callback_user_data = nullptr;
+    params.seed                    = 0;
+    return params;
+}
+
+static void diffusion_generate(llama_context * ctx,
+                        const llama_token * input_tokens,
+                        llama_token * output_tokens,
+                        int32_t n_input,
+                        int32_t max_length,
+                        struct diffusion_params params,
+                        int32_t & n_generated) {
+
+    n_generated = 0;
+    if (!ctx || !input_tokens || !output_tokens || n_input <= 0 || max_length <= n_input) {
+        return;
+    }
+
+    const llama_model * model = llama_get_model(ctx);
+
+    // Initialize with input and pad with mask tokens
+    std::copy(input_tokens, input_tokens + n_input, output_tokens);
+    std::fill(output_tokens + n_input, output_tokens + max_length, params.mask_token_id);
+
+    std::mt19937 rng(params.seed);
+
+    std::vector<float> timesteps(params.steps + 1);
+    for (int32_t i = 0; i <= params.steps; i++) {
+        timesteps[i] = 1.0f - (float) i / params.steps * (1.0f - params.eps);
+    }
+
+    llama_set_causal_attn(ctx, false);
+
+    int32_t n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(model));
+
+    std::vector<llama_token_data> candidates(n_vocab);
+
+    std::vector<llama_token_data> conf_candidates;
+    conf_candidates.reserve(max_length);
+
+    std::vector<int32_t> mask_positions;
+    mask_positions.reserve(max_length);
+
+    struct llama_sampler * sampler = llama_sampler_chain_init(llama_sampler_chain_default_params());
+    if (params.top_k > 0) {
+        llama_sampler_chain_add(sampler, llama_sampler_init_top_k(params.top_k));
+    }
+    if (params.top_p < 1.0f) {
+        llama_sampler_chain_add(sampler, llama_sampler_init_top_p(params.top_p, 1));
+    }
+    if (params.temperature > 0.0f) {
+        llama_sampler_chain_add(sampler, llama_sampler_init_temp(params.temperature));
+    }
+    llama_sampler_chain_add(sampler, llama_sampler_init_dist(params.seed));
+
+    struct llama_sampler * dist_sampler = llama_sampler_init_dist(params.seed);
+
+    llama_batch batch = llama_batch_init(max_length, 0, 1);
+    batch.n_tokens    = max_length;
+
+    int64_t total_sampling_time = 0;
+    int64_t total_time = 0;
+
+    int64_t time_start = ggml_time_us();
+    for (int32_t step = 0; step < params.steps; step++) {
+        if (params.step_callback) {
+            if (!params.step_callback(step, params.steps, output_tokens, max_length, params.step_callback_user_data)) {
+                break;
+            }
+        }
+
+        for (int32_t i = 0; i < max_length; i++) {
+            batch.token[i]     = output_tokens[i];
+            batch.pos[i]       = i;
+            batch.n_seq_id[i]  = 1;
+            batch.seq_id[i][0] = 0;
+            batch.logits[i]    = 1;
+        }
+
+        int ret = llama_decode(ctx, batch);
+        if (ret != 0) {
+            LOG_ERR("%s: failed to decode at step %d, ret = %d\n", __func__, step, ret);
+            break;
+        }
+
+        float * raw_logits = llama_get_logits(ctx);
+        if (!raw_logits) {
+            LOG_ERR("%s: failed to get logits at step %d\n", __func__, step);
+            break;
+        }
+
+        auto get_logits_for_pos = [&](int32_t pos) -> const float * {
+            return pos == 0 ? raw_logits : raw_logits + (pos - 1) * n_vocab;
+        };
+
+        int64_t time_start_sampling = ggml_time_us();
+
+        mask_positions.clear();
+        for (int32_t i = 0; i < max_length; i++) {
+            if (output_tokens[i] == params.mask_token_id) {
+                mask_positions.push_back(i);
+            }
+        }
+
+        if (mask_positions.empty()) {
+            break;
+        }
+
+        float t = timesteps[step];
+        float s = timesteps[step + 1];
+
+        if (params.algorithm == DIFFUSION_ALG_ORIGIN) {
+            float p_transfer = (step < params.steps - 1) ? (1.0f - s / t) : 1.0f;
+
+            for (int32_t pos : mask_positions) {
+                if (std::uniform_real_distribution<float>(0.0f, 1.0f)(rng) < p_transfer) {
+                    const float * pos_logits = get_logits_for_pos(pos);
+                    for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
+                        candidates[token_id].id    = token_id;
+                        candidates[token_id].logit = pos_logits[token_id];
+                        candidates[token_id].p     = 0.0f;
+                    }
+
+                    llama_token_data_array cur_p = {
+                        /* .data       = */ candidates.data(),
+                        /* .size       = */ (size_t) n_vocab,  // Reset size to full vocab
+                        /* .selected   = */ -1,
+                        /* .sorted     = */ false,
+                    };
+
+                    llama_sampler_apply(sampler, &cur_p);
+                    output_tokens[pos] = cur_p.data[cur_p.selected].id;
+                }
+            }
+        } else {
+            std::vector<std::pair<float, int32_t>> confidences;
+            std::vector<llama_token>               sampled_tokens(mask_positions.size());
+
+            for (size_t i = 0; i < mask_positions.size(); i++) {
+                int32_t       pos        = mask_positions[i];
+                const float * pos_logits = get_logits_for_pos(pos);
+
+                for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
+                    candidates[token_id].logit = pos_logits[token_id];
+                    candidates[token_id].p     = 0.0f;
+                    candidates[token_id].id    = token_id;
+                }
+
+                llama_token_data_array cur_p = {
+                    /* .data       = */ candidates.data(),
+                    /* .size       = */ candidates.size(),
+                    /* .selected   = */ -1,
+                    /* .sorted     = */ false,
+                };
+
+                llama_sampler_apply(sampler, &cur_p);
+
+                llama_token sampled_token = cur_p.data[cur_p.selected].id;
+
+                float confidence = 0.0f;
+                if (params.algorithm == DIFFUSION_ALG_ENTROPY) {
+                    const float epsilon = 1e-10f;
+                    for (size_t j = 0; j < cur_p.size; j++) {
+                        float prob = cur_p.data[j].p;
+                        confidence += prob * logf(prob + epsilon);
+                    }
+                } else if (params.algorithm == DIFFUSION_ALG_TOPK_MARGIN) {
+                    confidence = cur_p.data[0].p - cur_p.data[1].p;
+                } else {
+                    confidence = cur_p.data[cur_p.selected].p;
+                }
+
+                sampled_tokens[i] = sampled_token;
+                confidences.emplace_back(confidence, i);
+            }
+
+            int32_t num_transfer =
+                (step < params.steps - 1) ? (int32_t) (mask_positions.size() * (1.0f - s / t)) : mask_positions.size();
+
+            if (num_transfer > 0) {
+                if (params.alg_temp == 0.0f) {
+                    std::partial_sort(confidences.begin(), confidences.begin() + num_transfer, confidences.end(),
+                                      [](const std::pair<float, int32_t> & a, const std::pair<float, int32_t> & b) {
+                                          if (a.first != b.first) {
+                                              return a.first > b.first;
+                                          }
+                                          return a.second < b.second;
+                                      });
+                } else {
+                    conf_candidates.clear();
+
+                    for (int32_t pos = 0; pos < max_length; pos++) {
+                        float conf_logit = -std::numeric_limits<float>::infinity();
+
+                        auto it = std::find(mask_positions.begin(), mask_positions.end(), pos);
+                        if (it != mask_positions.end()) {
+                            size_t mask_idx = std::distance(mask_positions.begin(), it);
+                            conf_logit = confidences[mask_idx].first / params.alg_temp;  // Apply temperature scaling
+                        }
+
+                        conf_candidates.emplace_back(llama_token_data{ pos, conf_logit, 0.0f });
+                    }
+
+                    llama_token_data_array conf_array = {
+                        /* .data       = */ conf_candidates.data(),
+                        /* .size       = */ conf_candidates.size(),
+                        /* .selected   = */ -1,
+                        /* .sorted     = */ false,
+                    };
+
+                    for (int32_t i = 0; i < num_transfer; i++) {
+                        // Apply distribution sampler to get selected index
+                        llama_sampler_apply(dist_sampler, &conf_array);
+                        int selected_idx      = conf_array.selected;
+                        confidences[i].second = conf_candidates[selected_idx].id;
+
+                        conf_candidates[selected_idx].p = 0.0f;
+                        conf_array.selected             = -1;
+                    }
+                }
+
+                if (params.alg_temp == 0.0f) {
+                    // Deterministic - use confidence order
+                    for (int32_t i = 0; i < num_transfer; i++) {
+                        int32_t     mask_idx = confidences[i].second;
+                        int32_t     pos      = mask_positions[mask_idx];
+                        llama_token token    = sampled_tokens[mask_idx];
+                        output_tokens[pos]   = token;
+                    }
+                } else {
+                    for (int32_t i = 0; i < num_transfer; i++) {
+                        int32_t pos = confidences[i].second;
+                        auto    it  = std::find(mask_positions.begin(), mask_positions.end(), pos);
+                        if (it != mask_positions.end()) {
+                            int32_t mask_idx   = std::distance(mask_positions.begin(), it);
+                            output_tokens[pos] = sampled_tokens[mask_idx];
+                        }
+                    }
+                }
+            }
+        }
+        int64_t time_end_sampling = ggml_time_us();
+        total_sampling_time += time_end_sampling - time_start_sampling;
+    }
+    int64_t time_end = ggml_time_us();
+    total_time += time_end - time_start;
+
+    LOG_INF("\ntotal time: %0.2fms, time per step: %0.2fms, sampling time per step: %0.2fms\n",
+            total_time / 1000.0, total_time / 1000.0 / params.steps, total_sampling_time / 1000.0 / params.steps);
+
+
+    llama_batch_free(batch);
+    llama_sampler_free(sampler);
+    llama_sampler_free(dist_sampler);
+
+    n_generated = max_length;
+}
+
+
+
+
+static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
+    if (!use_chat_template) {
+        return prompt;
+    }
+
+    auto chat_templates = common_chat_templates_init(model, "");
+
+    common_chat_templates_inputs inputs;
+    common_chat_msg              user_msg;
+    user_msg.role                = "user";
+    user_msg.content             = prompt;
+    inputs.add_generation_prompt = true;
+    inputs.messages.push_back(user_msg);
+
+    auto result = common_chat_templates_apply(chat_templates.get(), inputs);
+
+    return result.prompt;
+}
+
+struct callback_data {
+    const common_params_diffusion * diff_params;
+    const llama_vocab *             vocab;
+    int32_t                         n_input;
+};
+
+static bool diffusion_step_callback(int32_t step,
+                                    int32_t total_steps,
+                                    const llama_token * tokens,
+                                    int32_t n_tokens,
+                                    void * user_data) {
+    (void)user_data;
+
+    callback_data * data = static_cast<callback_data *>(user_data);
+
+    auto print_progress_bar = [](int32_t step, int32_t total_steps) {
+        int progress_percent = (step * 100) / total_steps;
+        int progress_bars    = (step * 50) / total_steps;
+        LOG_INF("\rdiffusion step: %d/%d [%s%s] %d%%",
+            step,
+            total_steps,
+            std::string(progress_bars, '=').c_str(),
+            std::string(50 - progress_bars, ' ').c_str(),
+            progress_percent);
+    };
+
+    if (data->diff_params->visual_mode) {
+        // Visual mode: clear
+        LOG_INF("\033[2J\033[H");  // Clear screen and move cursor to top-left
+
+        print_progress_bar(step, total_steps);
+
+        LOG_INF("\n");
+
+        std::string current_text = " ";
+
+        for (int32_t i = data->n_input; i < n_tokens; i++) {
+            std::string token_str;
+            if (tokens[i] != llama_vocab_mask(data->vocab)) {
+                char piece[256];
+                int  n_chars = llama_token_to_piece(data->vocab, tokens[i], piece, sizeof(piece), 0, false);
+                if (n_chars > 0) {
+                    piece[n_chars] = '\0';
+                    token_str      = piece;
+                }
+            } else {
+                token_str = " ";
+            }
+
+            current_text += token_str;
+        }
+
+        LOG_INF("%s\n", current_text.c_str());
+    } else {
+        print_progress_bar(step, total_steps);
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DIFFUSION)) {
+        return 1;
+    }
+
+    const char * alg_names[] = { "ORIGIN", "MASKGIT_PLUS", "TOPK_MARGIN", "ENTROPY" };
+    const char * alg_name    = (params.diffusion.algorithm >= 0 && params.diffusion.algorithm <= 3) ?
+                                   alg_names[params.diffusion.algorithm] :
+                                   "UNKNOWN";
+
+    common_init();
+    llama_backend_init();
+
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers       = params.n_gpu_layers;
+    model_params.devices            = params.devices.data();
+    model_params.use_mmap           = params.use_mmap;
+    model_params.use_mlock          = params.use_mlock;
+    model_params.check_tensors      = params.check_tensors;
+
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
+    if (!model) {
+        LOG_ERR("error: failed to load model '%s'\n", params.model.path.c_str());
+        return 1;
+    }
+
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx                = params.n_ctx;
+    ctx_params.n_batch              = params.n_batch;
+    ctx_params.n_ubatch             = params.n_ubatch;
+    ctx_params.flash_attn           = params.flash_attn;
+    ctx_params.no_perf              = params.no_perf;
+    ctx_params.type_k               = params.cache_type_k;
+    ctx_params.type_v               = params.cache_type_v;
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    if (!ctx) {
+        LOG_ERR("error: failed to create context\n");
+        llama_model_free(model);
+        return 1;
+    }
+
+    llama_set_n_threads(ctx, params.cpuparams.n_threads, params.cpuparams_batch.n_threads);
+
+    const llama_vocab * vocab            = llama_model_get_vocab(model);
+    std::string         formatted_prompt = format_input_text(params.prompt, params.enable_chat_template, model);
+
+    std::vector<llama_token> input_tokens = common_tokenize(vocab, formatted_prompt,
+                                                            /*add special tokens*/ true,
+                                                            /*parse special*/ true);
+    int                      n_input      = input_tokens.size();
+
+    if (n_input >= params.n_ctx) {
+        LOG_ERR("error: input too long (%d tokens), max context is %d\n", n_input, params.n_ctx);
+        llama_free(ctx);
+        llama_model_free(model);
+        return 1;
+    }
+
+    struct diffusion_params ldiff_params = diffusion_default_params();
+    ldiff_params.steps                   = params.diffusion.steps;
+    ldiff_params.eps                     = params.diffusion.eps;
+    ldiff_params.temperature             = params.sampling.temp;
+    ldiff_params.top_p                   = params.sampling.top_p;
+    ldiff_params.top_k                   = params.sampling.top_k;
+    ldiff_params.algorithm               = static_cast<enum diffusion_alg>(params.diffusion.algorithm);
+    ldiff_params.alg_temp                = params.diffusion.alg_temp;
+    ldiff_params.seed                    = params.sampling.seed;
+
+    llama_token mask_token_id = llama_vocab_mask(vocab);
+    GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
+
+    LOG_INF("diffusion_params: - %-25s llama_token      = %d\n", "mask_token_id", mask_token_id);
+    LOG_INF("diffusion_params: - %-25s u32              = %d\n", "steps", params.diffusion.steps);
+    LOG_INF("diffusion_params: - %-25s f32              = %.6f\n", "eps", params.diffusion.eps);
+    LOG_INF("diffusion_params: - %-25s u32              = %d (%s)\n", "algorithm", params.diffusion.algorithm,
+            alg_name);
+    LOG_INF("diffusion_params: - %-25s f32              = %.3f\n", "alg_temp", params.diffusion.alg_temp);
+
+    ldiff_params.mask_token_id = mask_token_id;
+
+    callback_data cb_data = { &params.diffusion, vocab, n_input };
+
+    ldiff_params.step_callback           = diffusion_step_callback;
+    ldiff_params.step_callback_user_data = &cb_data;
+
+    int32_t n_generated = 0;
+
+    std::vector<llama_token> output_tokens(params.n_ubatch);
+    diffusion_generate(ctx, input_tokens.data(), output_tokens.data(), n_input, params.n_ubatch,
+                       ldiff_params, n_generated);
+
+    if (n_generated > 0) {
+        if (params.diffusion.visual_mode) {
+            //clear screen and move cursor to top-left
+            LOG_INF("\033[2J\033[H");
+        }
+        output_tokens.erase(output_tokens.begin(), output_tokens.begin() + n_input);
+        std::string output_data = common_detokenize(vocab, output_tokens, false);
+        LOG_INF("\n%s\n", output_data.c_str());
+    } else {
+        LOG_INF("Error: diffusion generation failed\n");
+    }
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_backend_free();
+
+    return 0;
+}

examples/embedding/embedding.cpp

@@ -81,6 +81,14 @@ int main(int argc, char ** argv) {
 
     params.embedding = true;
 
+    // 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;
+    }
+
     // utilize the full context
     if (params.n_batch < params.n_ctx) {
         LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);
@@ -107,7 +115,7 @@ int main(int argc, char ** argv) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     const int n_ctx_train = llama_model_n_ctx_train(model);
-    const int n_ctx = llama_n_ctx(ctx);
+    const int n_ctx       = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 

examples/parallel/parallel.cpp

@@ -184,6 +184,9 @@ int main(int argc, char ** argv) {
     // extra text to insert in each client's prompt in order to make it larger
     const int32_t n_junk = std::max(1, params.n_junk);
 
+    // signed seed, use negative values to indicate different seeds for the different clients
+    const int32_t & sseed = params.sampling.seed;
+
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -219,11 +222,21 @@ int main(int argc, char ** argv) {
 
     const int n_ctx = llama_n_ctx(ctx);
 
+    if (sseed >= 0) {
+        LOG_INF("%s: initializing all samplers with the same RNG seed: %d (use a negative seed to have different seeds)\n", __func__, sseed);
+    } else {
+        LOG_INF("%s: initializing samplers with different RNG seeds, starting from %d\n", __func__, sseed);
+    }
+
     std::vector<client> clients(n_clients);
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
         client.smpl = common_sampler_init(model, params.sampling);
+
+        if (sseed < 0) {
+            params.sampling.seed--;
+        }
     }
 
     std::vector<llama_token> tokens_system;
@@ -345,7 +358,7 @@ int main(int argc, char ** argv) {
                     client.n_decoded = 0;
                     client.i_batch   = batch.n_tokens - 1;
 
-                    LOG_INF("\033[31mClient %3d, seq %4d, junk = %4d, started decoding ...\033[0m\n", client.id, client.seq_id, n_junk_cur);
+                    LOG_INF("\033[31mClient %3d, seq %4d, junk = %4d, prompt = %d, started decoding ...\033[0m\n", client.id, client.seq_id, n_junk_cur, client.n_prompt);
 
                     g_seq_id += 1;
 

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

@@ -15,6 +15,12 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    if (params.n_parallel == 1) {
+        // the example uses 2 sequences, so when n_parallel == 1, we need to enable unified kv cache
+        printf("%s: n_parallel == 1, enabling unified kv cache\n", __func__);
+        params.kv_unified = true;
+    }
+
     common_init();
 
     if (params.n_predict < 0) {

ggml/CMakeLists.txt

@@ -131,7 +131,7 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
-option(GGML_NNPA             "ggml: enable nnpa"             ON)
+option(GGML_NNPA             "ggml: enable nnpa"             OFF)  # temp disabled by default, see: https://github.com/ggml-org/llama.cpp/issues/14877
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
@@ -174,6 +174,9 @@ option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental,
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
 option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
+option(GGML_HIP_MMQ_MFMA                    "ggml: enable MFMA MMA for CDNA in MMQ"           ON)
+option(GGML_MUSA_GRAPHS                     "ggml: use MUSA graph, experimental, unstable"    OFF)
+option(GGML_MUSA_MUDNN_COPY                 "ggml: enable muDNN for accelerated copy"         OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -181,6 +184,8 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
+option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
+option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -270,6 +275,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-rpc.h
     include/ggml-sycl.h
     include/ggml-vulkan.h
+    include/ggml-webgpu.h
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

ggml/cmake/ggml-config.cmake.in

@@ -1,152 +1,189 @@
+@PACKAGE_INIT@
 
 @GGML_VARIABLES_EXPANDED@
 
-@PACKAGE_INIT@
-
-set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
-set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
-#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
-
-find_package(Threads REQUIRED)
-
-find_library(GGML_LIBRARY ggml
-    REQUIRED
-    HINTS ${GGML_LIB_DIR}
-    NO_CMAKE_FIND_ROOT_PATH)
-
-add_library(ggml::ggml UNKNOWN IMPORTED)
-set_target_properties(ggml::ggml
-    PROPERTIES
-        IMPORTED_LOCATION "${GGML_LIBRARY}")
-
-find_library(GGML_BASE_LIBRARY ggml-base
-    REQUIRED
-    HINTS ${GGML_LIB_DIR}
-    NO_CMAKE_FIND_ROOT_PATH)
-
-add_library(ggml::ggml-base UNKNOWN IMPORTED)
-set_target_properties(ggml::ggml-base
-    PROPERTIES
-        IMPORTED_LOCATION "${GGML_BASE_LIBRARY}")
-
+# Find all dependencies before creating any target.
+include(CMakeFindDependencyMacro)
+find_dependency(Threads)
 if (NOT GGML_SHARED_LIB)
+    set(GGML_CPU_INTERFACE_LINK_LIBRARIES "")
+    set(GGML_CPU_INTERFACE_LINK_OPTIONS   "")
+
     if (APPLE AND GGML_ACCELERATE)
-        find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
+        if(NOT ACCELERATE_FRAMEWORK)
+            set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
+            return()
+        endif()
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${ACCELERATE_FRAMEWORK})
     endif()
 
-    if (GGML_OPENMP)
-        find_package(OpenMP REQUIRED)
+    if (GGML_OPENMP_ENABLED)
+        find_dependency(OpenMP)
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
     endif()
 
     if (GGML_CPU_HBM)
-        find_library(memkind memkind REQUIRED)
+        find_library(memkind memkind)
+        if(NOT memkind)
+            set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
+            return()
+        endif()
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES memkind)
     endif()
 
     if (GGML_BLAS)
-        find_package(BLAS REQUIRED)
-        list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
-        list(APPEND GGML_CPU_INTERFACE_LINK_OPTIONS   ${BLAS_LINKER_FLAGS})
+        find_dependency(BLAS)
+        list(APPEND GGML_BLAS_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
+        list(APPEND GGML_BLAS_INTERFACE_LINK_OPTIONS   ${BLAS_LINKER_FLAGS})
     endif()
 
     if (GGML_CUDA)
-        find_package(CUDAToolkit REQUIRED)
+        set(GGML_CUDA_INTERFACE_LINK_LIBRARIES "")
+        find_dependency(CUDAToolkit)
+        if (GGML_STATIC)
+            list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cudart_static>)
+            if (WIN32)
+                list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas> $<LINK_ONLY:CUDA::cublasLt>)
+            else()
+                list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas_static> $<LINK_ONLY:CUDA::cublasLt_static>)
+            endif()
+        endif()
+        if (NOT GGML_CUDA_NO_VMM)
+            list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cuda_driver>)
+        endif()
     endif()
 
     if (GGML_METAL)
-        find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-        find_library(METAL_FRAMEWORK    Metal REQUIRED)
-        find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+        find_library(FOUNDATION_LIBRARY Foundation)
+        find_library(METAL_FRAMEWORK    Metal)
+        find_library(METALKIT_FRAMEWORK MetalKit)
+        if(NOT FOUNDATION_LIBRARY OR NOT METAL_FRAMEWORK OR NOT METALKIT_FRAMEWORK)
+            set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
+            return()
+        endif()
+        set(GGML_METAL_INTERFACE_LINK_LIBRARIES
+            ${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
+    endif()
 
-        list(APPEND GGML_METAL_INTERFACE_LINK_LIBRARIES
-                    ${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
+    if (GGML_OPENCL)
+        find_dependency(OpenCL)
+        set(GGML_OPENCL_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:OpenCL::OpenCL>)
     endif()
 
     if (GGML_VULKAN)
-        find_package(Vulkan REQUIRED)
-        list(APPEND GGML_VULKAN_INTERFACE_LINK_LIBRARIES Vulkan::Vulkan)
+        find_dependency(Vulkan)
+        set(GGML_VULKAN_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:Vulkan::Vulkan>)
     endif()
 
     if (GGML_HIP)
-        find_package(hip     REQUIRED)
-        find_package(hipblas REQUIRED)
-        find_package(rocblas REQUIRED)
-        list(APPEND GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
+        find_dependency(hip)
+        find_dependency(hipblas)
+        find_dependency(rocblas)
+        set(GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
     endif()
 
     if (GGML_SYCL)
+        set(GGML_SYCL_INTERFACE_LINK_LIBRARIES "")
         find_package(DNNL)
         if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
             list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES DNNL::dnnl)
         endif()
         if (WIN32)
-            find_package(IntelSYCL REQUIRED)
-            find_package(MKL       REQUIRED)
+            find_dependency(IntelSYCL)
+            find_dependency(MKL)
             list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
         endif()
     endif()
 endif()
 
-set(_ggml_all_targets "")
-foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
-    string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
-    string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
+set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
+set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
+#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
 
-    find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
+if(NOT TARGET ggml::ggml)
+    find_package(Threads REQUIRED)
+
+    find_library(GGML_LIBRARY ggml
         REQUIRED
         HINTS ${GGML_LIB_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
-    message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
-
-    add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
-    set_target_properties(ggml::${_ggml_backend}
+    add_library(ggml::ggml UNKNOWN IMPORTED)
+    set_target_properties(ggml::ggml
         PROPERTIES
-            INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
-            IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
-            IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
-            INTERFACE_COMPILE_FEATURES c_std_90
-            POSITION_INDEPENDENT_CODE ON)
+            IMPORTED_LOCATION "${GGML_LIBRARY}")
 
-    string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
-    if(is_cpu_variant)
-        list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
-        set_target_properties(ggml::${_ggml_backend}
-           PROPERTIES
-               INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
+    find_library(GGML_BASE_LIBRARY ggml-base
+        REQUIRED
+        HINTS ${GGML_LIB_DIR}
+        NO_CMAKE_FIND_ROOT_PATH)
 
-        if(GGML_CPU_INTERFACE_LINK_OPTIONS)
-            set_target_properties(ggml::${_ggml_backend}
-                PROPERTIES
-                    INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
-        endif()
+    add_library(ggml::ggml-base UNKNOWN IMPORTED)
+    set_target_properties(ggml::ggml-base
+        PROPERTIES
+            IMPORTED_LOCATION "${GGML_BASE_LIBRARY}")
 
-    else()
-        list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
+    set(_ggml_all_targets "")
+    foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
+        string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
+        string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
+
+        find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
+            REQUIRED
+            HINTS ${GGML_LIB_DIR}
+            NO_CMAKE_FIND_ROOT_PATH)
+
+        message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
+
+        add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
         set_target_properties(ggml::${_ggml_backend}
             PROPERTIES
-                INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
+                INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
+                IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+                IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
+                INTERFACE_COMPILE_FEATURES c_std_90
+                POSITION_INDEPENDENT_CODE ON)
 
-        if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
+        string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
+        if(is_cpu_variant)
+            list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
+            set_target_properties(ggml::${_ggml_backend}
+            PROPERTIES
+                INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
+
+            if(GGML_CPU_INTERFACE_LINK_OPTIONS)
+                set_target_properties(ggml::${_ggml_backend}
+                    PROPERTIES
+                        INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
+            endif()
+
+        else()
+            list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
             set_target_properties(ggml::${_ggml_backend}
                 PROPERTIES
-                    INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
+                    INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
+
+            if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
+                set_target_properties(ggml::${_ggml_backend}
+                    PROPERTIES
+                        INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
+            endif()
         endif()
-    endif()
 
-    list(APPEND _ggml_all_targets ggml::${_ggml_backend})
-endforeach()
+        list(APPEND _ggml_all_targets ggml::${_ggml_backend})
+    endforeach()
 
-list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
-set_target_properties(ggml::ggml
-    PROPERTIES
-        INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}")
+    list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
+    set_target_properties(ggml::ggml
+        PROPERTIES
+            INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}")
 
-add_library(ggml::all INTERFACE IMPORTED)
-set_target_properties(ggml::all
-    PROPERTIES
-        INTERFACE_LINK_LIBRARIES "${_ggml_all_targets}")
+    add_library(ggml::all INTERFACE IMPORTED)
+    set_target_properties(ggml::all
+        PROPERTIES
+            INTERFACE_LINK_LIBRARIES "${_ggml_all_targets}")
+
+endif()
 
 check_required_components(ggml)

ggml/include/ggml-webgpu.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_WEBGPU_NAME "WebGPU"
+
+// Needed for examples in ggml
+GGML_BACKEND_API ggml_backend_t ggml_backend_webgpu_init(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_webgpu_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml/src/CMakeLists.txt

@@ -370,6 +370,7 @@ ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
+ggml_add_backend(WebGPU)
 ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)

ggml/src/ggml-alloc.c

@@ -22,21 +22,6 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
     return t->view_src != NULL;
 }
 
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
 // ops that return true for this function must not use restrict pointers for their backend implementations
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {

ggml/src/ggml-backend-reg.cpp

@@ -45,6 +45,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef GGML_USE_WEBGPU
+#include "ggml-webgpu.h"
+#endif
+
 #ifdef GGML_USE_OPENCL
 #include "ggml-opencl.h"
 #endif
@@ -173,6 +177,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_VULKAN
         register_backend(ggml_backend_vk_reg());
 #endif
+#ifdef GGML_USE_WEBGPU
+        register_backend(ggml_backend_webgpu_reg());
+#endif
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif

ggml/src/ggml-backend.cpp

@@ -352,21 +352,6 @@ ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
 
 // backend copy
 
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
 void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 
@@ -662,6 +647,7 @@ struct ggml_backend_sched {
     // pipeline parallelism support
     int n_copies;
     int cur_copy;
+    int next_copy;
     ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
     struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
     int n_graph_inputs;
@@ -1448,8 +1434,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         }
     }
 
-    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
-
     return GGML_STATUS_SUCCESS;
 }
 
@@ -1550,10 +1534,10 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
-    ggml_backend_sched_split_graph(sched, measure_graph);
-
     ggml_backend_sched_synchronize(sched);
 
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
@@ -1565,6 +1549,10 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
+    GGML_ASSERT(!sched->is_alloc);
+
+    sched->cur_copy = sched->next_copy;
+    sched->next_copy = (sched->next_copy + 1) % sched->n_copies;
 
     ggml_backend_sched_split_graph(sched, graph);
 
@@ -1605,7 +1593,7 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
         // if the graph is not already allocated, always use copy 0 after a synchronization
         // this ensures that during generation the same copy is used every time,
         // which avoids changes in the graph that could cause CUDA or other graphs to be disabled
-        sched->cur_copy = 0;
+        sched->next_copy = 0;
     }
 }
 

ggml/src/ggml-cann/acl_tensor.cpp

@@ -77,6 +77,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     for (int i = 0; i < final_dims; i++) {
         acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
     }
+    size_t elem_offset = offset / ggml_element_size(tensor);
+    acl_storage_len += elem_offset;
 
     // Reverse ne and stride.
     std::reverse(acl_ne, acl_ne + final_dims);
@@ -84,7 +86,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
 
     aclTensor* acl_tensor = aclCreateTensor(
         acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
-        offset / ggml_element_size(tensor), format, &acl_storage_len, 1,
+        elem_offset, format, &acl_storage_len, 1,
         tensor->data);
 
     return acl_tensor;

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -68,6 +68,8 @@
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
 #include <aclnnop/aclnn_zero.h>
+#include <aclnnop/aclnn_index_copy.h>
+#include <aclnnop/aclnn_index_select.h>
 #include <float.h>
 
 #include <cmath>
@@ -99,7 +101,7 @@ void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclT
     }
 }
 
-void ggml_cann_unary_op(
+void ggml_cann_op_unary(
     std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
     ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
@@ -111,6 +113,42 @@ void ggml_cann_unary_op(
     ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }
 
+void ggml_cann_op_unary_gated(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor *acl_src0 = nullptr, *acl_src1 = nullptr;
+    if(src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+
+        acl_src0 = ggml_cann_create_tensor(src0);
+        acl_src1 = ggml_cann_create_tensor(src1);
+    } else {
+        int64_t ne[] = {src0->ne[0] / 2, src0->ne[1], src0->ne[2], src0->ne[3]};
+        size_t nb[] = {src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]};
+        acl_src0 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, 0);
+        acl_src1 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, ne[0] * ggml_element_size(src0));
+        if (swapped) {
+            std::swap(acl_src0, acl_src1);
+        }
+    }
+
+    unary_op(ctx, acl_src0, acl_dst);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst, acl_src1);
+
+    ggml_cann_release_resources(ctx, acl_src0, acl_dst);
+    if(src1)
+        ggml_cann_release_resources(ctx, acl_src1);
+}
+
 /**
  * @brief Repeats elements of a tensor along each dimension according to the
  * specified repeat array.
@@ -1578,50 +1616,97 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 }
 
 /**
- * @brief Performs embedding operation on a 4D tensor using the CANN backend.
+ * @brief Performs index select operation on a 4D tensor using the CANN backend.
  *
- * This function extracts slices from the source tensor (`src_buffer`),
- * index tensor (`index`), and destination tensor (`dst`), and performs an
- * embedding operation on them. The embedding operation is applied by iterating
- * over the last two dimensions of the source tensor, creating the necessary
- * tensors for the source, index, and output, and executing the embedding operation.
+ * This function applies the `IndexSelect` operation along a specific dimension
+ * of the source tensor (`src_buffer`) using the indices from the index tensor (`index`).
+ * It iterates over the last two dimensions of the source tensor, creates the corresponding
+ * CANN tensors for the source, index, and output slices, and executes the `IndexSelect`
+ * operation for each slice.
  *
  * @param ctx The context for CANN backend operations.
- * @param src_buffer The source buffer holding the data for the source tensor.
+ * @param src_buffer The source buffer containing the 4D input tensor data.
  * @param src_ne The dimensions of the source tensor.
  * @param src_nb The strides (byte offsets) of the source tensor.
- * @param index The index tensor used in the embedding operation.
- * @param dst The destination tensor where the result will be stored.
+ * @param dst_buffer The destination buffer where the output tensor data will be written.
+ * @param dst_ne The dimensions of the destination tensor.
+ * @param dst_nb The strides (byte offsets) of the destination tensor.
+ * @param index The index tensor specifying the indices to select from the source tensor.
+ * @param type The data type of the source and destination tensors.
  */
-static void aclnn_embedding_4d(ggml_backend_cann_context& ctx, void* src_buffer,
-                            int64_t* src_ne, size_t* src_nb, ggml_tensor* index,
-                            ggml_tensor* dst) {
+static void aclnn_index_select_4d(ggml_backend_cann_context& ctx,
+                                void* src_buffer,int64_t* src_ne, size_t* src_nb,
+                                void* dst_buffer, int64_t* dst_ne, size_t* dst_nb,
+                                ggml_tensor* index, ggml_type type) {
     for (int64_t i = 0; i < src_ne[3]; i++) {
         for (int64_t j = 0; j < src_ne[2]; j++) {
             // src
-            int64_t acl_src_ne[2] = {src_ne[0], src_ne[1]};
-            size_t acl_src_nb[2] = {src_nb[0], src_nb[1]};
             aclTensor* acl_src_tensor = ggml_cann_create_tensor(
                 (char*)src_buffer + i * src_nb[3] + j * src_nb[2],
-                ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
-                acl_src_ne, acl_src_nb, 2);
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                src_ne, src_nb, 2);
 
             // index
-            int64_t acl_index_ne[1] = {index->ne[0]};
-            size_t acl_index_nb[1] = {index->nb[0]};
             aclTensor* acl_index = ggml_cann_create_tensor(
-                (char*)index->data + i * index->nb[2] + j * index->nb[1],
+                (char*)index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
                 ggml_cann_type_mapping(index->type), ggml_element_size(index),
-                acl_index_ne, acl_index_nb, 1);
+                index->ne, index->nb, 1);
 
             // out
-            int64_t acl_out_ne[2] = {dst->ne[0], dst->ne[1]};
-            size_t acl_out_nb[2] = {dst->nb[0], dst->nb[1]};
             aclTensor* acl_out = ggml_cann_create_tensor(
-                (char*)dst->data + i * dst->nb[3] + j * dst->nb[2],
-                ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
-                acl_out_ne, acl_out_nb, 2);
-            GGML_CANN_CALL_ACLNN_OP(ctx, Embedding, acl_src_tensor, acl_index, acl_out);
+                (char*)dst_buffer + i * dst_nb[3] + j * dst_nb[2],
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                dst_ne, dst_nb, 2);
+            GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, acl_src_tensor, 0, acl_index, acl_out);
+            ggml_cann_release_resources(ctx, acl_src_tensor, acl_index, acl_out);
+        }
+    }
+}
+
+/**
+ * @brief Performs inplace index copy operation on a 4D tensor using the CANN backend.
+ *
+ * This function applies the `IndexCopy` operation along a specific dimension of the
+ * destination tensor (`dst_buffer`) by copying elements from the source tensor (`src_buffer`)
+ * to positions specified by the index tensor (`index`).
+ * It iterates over the last two dimensions of the tensors, creates the corresponding
+ * CANN tensors for source, index, and destination slices, and performs the index copy
+ * operation for each slice.
+ *
+ * @param ctx The context for CANN backend operations.
+ * @param src_buffer The source buffer containing the 4D input tensor data to be copied.
+ * @param src_ne The dimensions of the source tensor.
+ * @param src_nb The strides (byte offsets) of the source tensor.
+ * @param dst_buffer The destination buffer where values will be copied to.
+ * @param dst_ne The dimensions of the destination tensor.
+ * @param dst_nb The strides (byte offsets) of the destination tensor.
+ * @param index The index tensor specifying target positions in the destination tensor.
+ * @param type The data type of the source and destination tensors.
+ */
+static void aclnn_index_copy_4d(ggml_backend_cann_context& ctx,
+                                void* src_buffer,int64_t* src_ne, size_t* src_nb,
+                                void* dst_buffer, int64_t* dst_ne, size_t* dst_nb,
+                                ggml_tensor* index, ggml_type type) {
+    for (int64_t i = 0; i < src_ne[3]; i++) {
+        for (int64_t j = 0; j < src_ne[2]; j++) {
+            // src
+            aclTensor* acl_src_tensor = ggml_cann_create_tensor(
+                (char*)src_buffer + i * src_nb[3] + j * src_nb[2],
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                src_ne, src_nb, 2);
+
+            // index
+            aclTensor* acl_index = ggml_cann_create_tensor(
+                (char*)index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
+                ggml_cann_type_mapping(index->type), ggml_element_size(index),
+                index->ne, index->nb, 1);
+
+            // out
+            aclTensor* acl_out = ggml_cann_create_tensor(
+                (char*)dst_buffer + i * dst_nb[3] + j * dst_nb[2],
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                dst_ne, dst_nb, 2);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_out, 0, acl_index, acl_src_tensor);
             ggml_cann_release_resources(ctx, acl_src_tensor, acl_index, acl_out);
         }
     }
@@ -1633,8 +1718,9 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
-            aclnn_embedding_4d(ctx, src0->data, src0->ne, src0->nb, src1,
-                                   dst);
+            aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
             break;
         }
         case GGML_TYPE_F16: {
@@ -1651,8 +1737,9 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 src_trans_buffer, ACL_FLOAT, ggml_type_size(dst->type),
                 src0->ne, src_trans_nb, GGML_MAX_DIMS);
             aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
-            aclnn_embedding_4d(ctx, src_trans_buffer, src0->ne,
-                                   src_trans_nb, src1, dst);
+            aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
             ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
             break;
         }
@@ -1712,8 +1799,10 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
             }
 
-            aclnn_embedding_4d(ctx, dequant_buffer_allocator.get(),
-                                   dequant_ne, dequant_nb, src1, dst);
+            aclnn_index_select_4d(ctx, dequant_buffer_allocator.get(),
+                                   dequant_ne, dequant_nb,
+                                   dst->data, dst->ne, dst->nb,
+                                   src1, dst->type);
 
             ggml_cann_release_resources(ctx, dequant_tensor);
             break;
@@ -1724,6 +1813,43 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     }
 }
 
+void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];  // src
+    ggml_tensor* src1 = dst->src[1];  // index
+
+    switch (dst->type) {
+        case GGML_TYPE_F32: {
+            aclnn_index_copy_4d(ctx, src0->data, src0->ne, src0->nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
+            break;
+        }
+        case GGML_TYPE_F16: {
+            aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+            ggml_cann_pool_alloc src_buffer_allocator(
+                ctx.pool(), ggml_nelements(src0) * sizeof(uint16_t));
+            void* src_trans_buffer = src_buffer_allocator.get();
+            size_t src_trans_nb[GGML_MAX_DIMS];
+            src_trans_nb[0] = sizeof(uint16_t);
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+            }
+            aclTensor* src_trans_tensor = ggml_cann_create_tensor(
+                src_trans_buffer, ACL_FLOAT16, ggml_type_size(dst->type),
+                src0->ne, src_trans_nb, GGML_MAX_DIMS);
+            aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
+            aclnn_index_copy_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
+            ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
+            break;
+        }
+        default:
+            GGML_ABORT("Unsupported tensor type for GGML_OP_SET_ROWS");
+            break;
+    }
+}
+
 /**
  * @brief Repeats elements of a tensor along a specified dimension.
  *
@@ -1785,8 +1911,25 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
                              bcast_weight_nb[2], bcast_weight_nb[3],
                              bcast_weight_nb[4], bcast_weight_nb[5]};
-    aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_weight_tensor;
+
+    // Only check env once.
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
+    if (weight_to_nz && is_matmul_weight(weight)) {
+        int64_t acl_stride[2] = {1, transpose_ne[1]};
+
+        // Reverse ne.
+        std::reverse(transpose_ne, transpose_ne + n_dims);
+
+        std::vector<int64_t> storageDims = {transpose_ne[0], transpose_ne[1]};
+
+        acl_weight_tensor = aclCreateTensor(
+            transpose_ne, n_dims, ggml_cann_type_mapping(weight->type), acl_stride,
+            0, ACL_FORMAT_FRACTAL_NZ, storageDims.data(), 2, weight->data);
+    } else {
+        acl_weight_tensor =
+            ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_ND);
+    }
     aclTensor* acl_dst =
         ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
 

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
+#include <unordered_set>
 #include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
@@ -423,15 +424,25 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @details This function retrieves rows from a source tensor src0 according to
  *          the indices provided in another tensor src1 and stores the result in
- *          a destination tensor (\p dst). It supports different data types
- *          including F32, F16, Q4_0, and Q8_0.
+ *          a destination tensor (\p dst).
  *
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the extracted rows will be stored.
- *            dst->op is `GGML_OP_GET_ROWS`.
  */
 void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Writes specific rows into a tensor at positions specified by indices.
+ *
+ * @details This function copies rows from a source tensor into a destination
+ *          tensor (\p dst) at the positions indicated by the indices in another
+ *          tensor.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor where the specified rows will be updated.
+ */
+void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /**
  * @brief   Executes matrix multiplication for the given tensor.
  *
@@ -1020,6 +1031,37 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  */
 void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
+ *
+ * @details Checks whether the given tensor serves as weight parameters in matrix multiplication operations,
+ *          typically within neural network layers. The function maintains a static set of canonical weight
+ *          naming suffixes from Transformer-based architectures. Uses substring matching to identify weight
+ *          tensors even with hierarchical naming patterns.
+ *
+ * @param tensor Pointer to the target ggml_tensor object (const-qualified).
+ */
+static bool is_matmul_weight(const ggml_tensor* tensor) {
+    std::string name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{
+        "output.weight",
+        "attn_q.weight",
+        "attn_k.weight",
+        "attn_v.weight",
+        "attn_output.weight",
+        "ffn_gate.weight",
+        "ffn_up.weight",
+        "ffn_down.weight"
+    };
+
+    for (const auto& suffix : weight_suffixes) {
+        if (name.find(suffix) != std::string::npos) {
+            return true;
+        }
+    }
+    return false;
+}
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.
@@ -1066,7 +1108,7 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @param dst The destination tensor. Its src[0] is treated as the input tensor.
  */
 template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
-    void ggml_cann_unary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
     aclTensor* acl_src = ggml_cann_create_tensor(src);
@@ -1077,49 +1119,125 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
 }
 
 /**
- * @brief   Applies a unary operation to a ggml tensor using the CANN backend.
+ * @brief Applies a unary operation to a ggml tensor using the CANN backend.
  *
- * @details This function performs a unary operation on the input tensor using
- * a user-provided lambda or callable object `unary_op`, which accepts the CANN
- * context and two ACL tensors (source and destination). Internally, this function
- * creates ACL representations of the ggml tensors and invokes the unary operation.
- * The result is stored in the destination tensor `dst`. This utility abstracts the
- * common boilerplate of tensor conversion and cleanup when implementing unary ops.
+ * @details This function applies a unary operation to the input tensor using
+ * a user-provided lambda or callable `unary_op`. The lambda receives the
+ * CANN backend context and two ACL tensors: the source and the destination.
  *
- * @param unary_op A callable that performs the unary operation using CANN APIs.
- * @param ctx The CANN context used for operations.
- * @param dst The destination tensor where the result will be stored.
- *            The source tensor is retrieved from `dst->src[0]`.
+ * Internally, this function handles the conversion from GGML tensors to ACL tensors,
+ * calls the provided unary op, and manages resource cleanup. The input is assumed
+ * to be `dst->src[0]`, and the result is written to `dst`.
+ *
+ * This utility simplifies writing unary op wrappers by abstracting tensor preparation.
+ *
+ * @param unary_op A callable that performs the unary operation using CANN ACL APIs.
+ * @param ctx The CANN context for operation execution.
+ * @param dst The destination ggml_tensor where the result will be stored.
+ *            The input tensor is assumed to be `dst->src[0]`.
+ *
+ * @see GGML_CANN_CALL_OP_UNARY
  */
-void ggml_cann_unary_op(
+void ggml_cann_op_unary(
     std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
     ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
- * @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op.
+ * @brief Applies a gated (GLU-style) unary operation using the CANN backend.
  *
- * This macro defines an inline lambda wrapping a specific ACL operation name,
- * and passes it to the templated ggml_cann_unary_op function. It simplifies
- * calling unary ops by hiding the lambda boilerplate.
+ * @details This function performs a gated activation such as GEGLU or ReGLU.
+ * It supports two input modes:
+ *
+ * 1. **Dual input mode**: `dst->src[0]` and `dst->src[1]` are both valid tensors.
+ *    These are used directly as the value and gate tensors.
+ *
+ * 2. **Packed input mode**: Only `dst->src[0]` is valid, and it is assumed to
+ *    contain a concatenation of value and gate along the first dimension. This tensor
+ *    will be split into two equal halves to form the value and gate inputs.
+ *
+ * The function applies a user-provided unary operation (e.g., GELU) to the value tensor,
+ * then multiplies the result in-place with the gate tensor:
  *
- * Internally, the lambda will call:
  * @code
- * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * dst = unary_op(value) * gate;
  * @endcode
  *
+ * The `swapped` parameter (from `dst->op_params[1]`) allows flipping the
+ * order of value/gate in the packed input case.
+ *
+ * @param unary_op A callable that performs the unary operation using CANN ACL APIs.
+ *                 It receives (ctx, acl_value_tensor, acl_output_tensor).
+ * @param ctx      The CANN context used for execution.
+ * @param dst      The destination ggml_tensor. Source tensors are in `dst->src[0]` and optionally `src[1]`.
+ *
+ * @see GGML_CANN_CALL_OP_UNARY_GATED
+ */
+void ggml_cann_op_unary_gated(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
+ *
+ * This macro wraps the specified ACLNN unary operator name into a lambda expression,
+ * and passes it to `ggml_cann_op_unary`, which handles the common logic for executing
+ * unary ops in the CANN backend.
+ *
+ * Internally, this macro expands to a lambda like:
+ * @code
+ * [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
+ *     GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * };
+ * @endcode
+ *
+ * This lambda is then passed to `ggml_cann_op_unary`, which applies the operation.
+ *
  * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
  *
- * @see ggml_cann_unary_op
+ * @see ggml_cann_op_unary
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                              \
+#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                              \
     do {                                                              \
         auto lambda = [](ggml_backend_cann_context& ctx,              \
             aclTensor* acl_src,                                       \
             aclTensor* acl_dst) {                                     \
             GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
         };                                                            \
-        ggml_cann_unary_op(lambda, ctx, dst);                         \
+        ggml_cann_op_unary(lambda, ctx, dst);                         \
     }                                                                 \
     while (0)
+
+/**
+ * @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
+ *
+ * This macro wraps the specified ACLNN unary operator name into a lambda expression,
+ * and passes it to `ggml_cann_op_unary_gated`, which handles the common logic for
+ * executing gated unary ops in the CANN backend.
+ *
+ * Internally, this macro expands to a lambda like:
+ * @code
+ * [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
+ *     GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * };
+ * @endcode
+ *
+ * This lambda is then passed to `ggml_cann_op_unary_gated`, which applies the operation.
+ *
+ * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
+ *
+ * @see ggml_cann_op_unary_gated
+ * @see GGML_CANN_CALL_ACLNN_OP
+ */
+#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                        \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_op_unary_gated(lambda, ctx, dst);                   \
+    }                                                                 \
+    while (0)
+
 #endif  // CANN_ACLNN_OPS

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

@@ -24,6 +24,7 @@
 
 #include <acl/acl.h>
 #include <stdarg.h>
+#include <aclnnop/aclnn_trans_matmul_weight.h>
 
 #include <cmath>
 #include <cstdio>
@@ -1115,6 +1116,61 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     return GGML_STATUS_SUCCESS;
 }
 
+// ND to NZ Workspace Cache Management. Thread-safety: Not guaranteed
+namespace {
+    void* g_nz_workspace = nullptr;
+    size_t g_nz_workspace_allocated = 0;
+
+    void release_nz_workspace() {
+        if (g_nz_workspace) {
+            aclrtFree(g_nz_workspace);
+            g_nz_workspace = nullptr;
+            g_nz_workspace_allocated = 0;
+        }
+    }
+
+    void relloc_nz_workspace(size_t new_size) {
+        if (new_size > g_nz_workspace_allocated) {
+        if (g_nz_workspace) {
+            aclrtFree(g_nz_workspace);
+            g_nz_workspace = nullptr;
+        }
+        ACL_CHECK(aclrtMalloc(&g_nz_workspace, new_size, ACL_MEM_MALLOC_HUGE_FIRST));
+        g_nz_workspace_allocated = new_size;
+    }
+    }
+}
+
+/**
+ * @brief Convert tensor weights to NZ format using Ascend CANN API.
+ *
+ * This function creates a transposed tensor descriptor and performs the
+ * TransMatmulWeight operation. Converting tensor formats can significantly
+ * improve performance on certain hardware.
+ *
+ * @param tensor Pointer to the input ggml_tensor containing the weights.
+ * @param data Pointer to the raw data buffer for the tensor weights.
+ * @param offset Byte offset within the tensor data buffer where weights start.
+ *
+ * @note The workspace buffer used in this function is managed globally and reused
+ *       across calls. This reduces overhead from repeated memory allocation and deallocation.
+ */
+static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
+    aclTensor* weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne,
+                                    tensor->nb, 2, ACL_FORMAT_ND, offset);
+    uint64_t workspaceSize = 0;
+    aclOpExecutor *executor;
+
+    // TransMatmulWeight
+    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed,
+                                                    &workspaceSize, &executor));
+    // Avoid frequent malloc/free of the workspace.
+    relloc_nz_workspace(workspaceSize);
+
+    ACL_CHECK(aclnnTransMatmulWeight(g_nz_workspace, workspaceSize, executor, nullptr));
+    ACL_CHECK(aclDestroyTensor(weightTransposed));
+}
+
 // TODO: need handle tensor which has paddings.
 /**
  * @brief Set tensor data in a CANN buffer.
@@ -1139,9 +1195,16 @@ static void ggml_backend_cann_buffer_set_tensor(
     // For acl, synchronous functions use this default stream.
     // Why aclrtSynchronizeDevice?
 
+    // Only check env once.
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
+        if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+            GGML_ASSERT(tensor->ne[3] == 1);
+            weight_format_to_nz(tensor, data, offset);
+        }
     } else {
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
@@ -1375,20 +1438,32 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
+    // Only check env once.
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
+
     // last line must bigger than 32, because every single op deal at
     // least 32 bytes.
     // TODO: quantized type?
     // int64_t line_size = ne0 * ggml_element_size(tensor);
     // int64_t line_size_align_32 = (line_size + 31) & ~31;
     // size += (line_size_align_32 - line_size);
-
-    // TODO: not support quantized yet.
-    // TODO: consider un-continue tensor.
     if (ggml_is_quantized(tensor->type)) {
         if (ne0 % MATRIX_ROW_PADDING != 0) {
             size += ggml_row_size(
                 tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
+    } else if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
+        // NZ format weight are not support quantized yet.
+        // If ND tensor transform to NZ, size may changed.
+        int64_t shape[] = {tensor->ne[1], tensor->ne[0]};
+        GGML_ASSERT(tensor->ne[2] == 1);
+        GGML_ASSERT(tensor->ne[3] == 1);
+        const aclIntArray *acl_shape = aclCreateIntArray(shape, 2);
+        size_t new_size;
+        ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(acl_shape,
+                    ggml_cann_type_mapping(tensor->type), &new_size));
+        ACL_CHECK(aclDestroyIntArray(acl_shape));
+        size = std::max(size, new_size);
     }
 
     return size;
@@ -1594,6 +1669,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
         case GGML_OP_GET_ROWS:
             ggml_cann_get_rows(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_cann_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cann_dup(ctx, dst);
             break;
@@ -1616,16 +1694,18 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(dst)) {
                 case GGML_UNARY_OP_ABS:
-                    GGML_CANN_CALL_UNARY_OP(Abs);
+                    GGML_CANN_CALL_OP_UNARY(Abs);
                     break;
                 case GGML_UNARY_OP_NEG:
-                    GGML_CANN_CALL_UNARY_OP(Neg);
+                    GGML_CANN_CALL_OP_UNARY(Neg);
                     break;
                 case GGML_UNARY_OP_GELU:
-                    GGML_CANN_CALL_UNARY_OP(Gelu);
+                case GGML_UNARY_OP_GELU_ERF:
+                    // aclnnGelu internally uses the erf-based approximation.
+                    GGML_CANN_CALL_OP_UNARY(Gelu);
                     break;
                 case GGML_UNARY_OP_SILU:
-                    GGML_CANN_CALL_UNARY_OP(Silu);
+                    GGML_CANN_CALL_OP_UNARY(Silu);
                     break;
                 case GGML_UNARY_OP_GELU_QUICK: {
                     auto lambda = [](ggml_backend_cann_context& ctx,
@@ -1633,31 +1713,31 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                         aclTensor* acl_dst) {
                         GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
                     };
-                    ggml_cann_unary_op(lambda, ctx, dst);
+                    ggml_cann_op_unary(lambda, ctx, dst);
                 } break;
                 case GGML_UNARY_OP_TANH:
-                    GGML_CANN_CALL_UNARY_OP(Tanh);
+                    GGML_CANN_CALL_OP_UNARY(Tanh);
                     break;
                 case GGML_UNARY_OP_RELU:
-                    GGML_CANN_CALL_UNARY_OP(Relu);
+                    GGML_CANN_CALL_OP_UNARY(Relu);
                     break;
                 case GGML_UNARY_OP_SIGMOID:
-                    GGML_CANN_CALL_UNARY_OP(Sigmoid);
+                    GGML_CANN_CALL_OP_UNARY(Sigmoid);
                     break;
                 case GGML_UNARY_OP_HARDSIGMOID:
-                    GGML_CANN_CALL_UNARY_OP(Hardsigmoid);
+                    GGML_CANN_CALL_OP_UNARY(Hardsigmoid);
                     break;
                 case GGML_UNARY_OP_HARDSWISH:
-                    GGML_CANN_CALL_UNARY_OP(Hardswish);
+                    GGML_CANN_CALL_OP_UNARY(Hardswish);
                     break;
                 case GGML_UNARY_OP_EXP:
-                    GGML_CANN_CALL_UNARY_OP(Exp);
+                    GGML_CANN_CALL_OP_UNARY(Exp);
                     break;
                 case GGML_UNARY_OP_ELU:
                     ggml_cann_elu(ctx, dst);
                     break;
                 case GGML_UNARY_OP_SGN:
-                    GGML_CANN_CALL_UNARY_OP(Sign);
+                    GGML_CANN_CALL_OP_UNARY(Sign);
                     break;
                 case GGML_UNARY_OP_STEP:
                     ggml_cann_step(ctx, dst);
@@ -1666,6 +1746,31 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                     return false;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(dst)) {
+                case GGML_GLU_OP_REGLU:
+                    GGML_CANN_CALL_OP_UNARY_GATED(Relu);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                    // aclnnGelu internally uses the erf-based approximation.
+                    GGML_CANN_CALL_OP_UNARY_GATED(Gelu);
+                    break;
+                case GGML_GLU_OP_SWIGLU:
+                    GGML_CANN_CALL_OP_UNARY_GATED(Silu);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK: {
+                    auto lambda = [](ggml_backend_cann_context& ctx,
+                        aclTensor* acl_src,
+                        aclTensor* acl_dst) {
+                        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
+                    };
+                    ggml_cann_op_unary_gated(lambda, ctx, dst);
+                } break;
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_NORM:
             ggml_cann_norm(ctx, dst);
             break;
@@ -1708,7 +1813,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_binary_op<aclnn_mul>(ctx, dst);
             break;
         case GGML_OP_SQRT:
-            GGML_CANN_CALL_UNARY_OP(Sqrt);
+            GGML_CANN_CALL_OP_UNARY(Sqrt);
             break;
         case GGML_OP_CLAMP:
             ggml_cann_clamp(ctx, dst);
@@ -1753,16 +1858,16 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_argmax(ctx, dst);
             break;
         case GGML_OP_COS:
-            ggml_cann_unary_op<aclnn_cos>(ctx, dst);
+            ggml_cann_op_unary<aclnn_cos>(ctx, dst);
             break;
         case GGML_OP_SIN:
-            ggml_cann_unary_op<aclnn_sin>(ctx, dst);
+            ggml_cann_op_unary<aclnn_sin>(ctx, dst);
             break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cann_conv_transpose_1d(ctx, dst);
             break;
         case GGML_OP_LOG:
-            GGML_CANN_CALL_UNARY_OP(Log);
+            GGML_CANN_CALL_OP_UNARY(Log);
             break;
         case GGML_OP_MEAN:
             ggml_cann_mean(ctx, dst);
@@ -1985,6 +2090,8 @@ static enum ggml_status ggml_backend_cann_graph_compute(
         (ggml_backend_cann_context*)backend->context;
 
     ggml_cann_set_device(cann_ctx->device);
+    //release temp buffer create by set tensor.
+    release_nz_workspace();
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor* node = cgraph->nodes[i];
@@ -2036,10 +2143,23 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_SGN:
                 case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_GELU_ERF:
                     return true;
                 default:
                     return false;
             }
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return true;
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
                 case GGML_TYPE_F16:
@@ -2086,13 +2206,15 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
-        case GGML_OP_SET_ROWS:
-            {
-                // TODO: add support
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-#pragma message("TODO: implement F32, F16, BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
-                return false;
-            } break;
+        case GGML_OP_SET_ROWS: {
+            switch (op->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                default:
+                    return false;
+            }
+        } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||

ggml/src/ggml-cpu/CMakeLists.txt

@@ -70,10 +70,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     if (GGML_OPENMP)
         find_package(OpenMP)
         if (OpenMP_FOUND)
+            set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
             target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
 
             target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
         else()
+            set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
             message(WARNING "OpenMP not found")
         endif()
     endif()
@@ -456,6 +458,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             list(APPEND ARCH_FLAGS -march=z16)
         elseif (${S390X_M} MATCHES "9175|9176")
             # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
             message(STATUS "z17 target")
             list(APPEND ARCH_FLAGS -march=z17)
         else()
@@ -494,9 +497,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.11.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
+        set(KLEIDIAI_ARCHIVE_MD5  "3fe9e5ab964c375c53839296eb71eaa2")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)

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

@@ -1236,44 +1236,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
-
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int sum = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 32; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
-                }
-            }
-        }
-        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 16; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
-                }
-            }
-        }
-
-        for (size_t l = 0; l < 4; ++l) {
-            for (size_t j = 0; j < sizeof(x->qh); ++j) {
-                uint8_t q = x[i].qh[j] * pow3[l];
-                uint16_t xi = ((uint16_t) q * 3) >> 8;
-                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
-            }
-        }
-
-        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1381,25 +1347,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int32_t sumi = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            for (size_t l = 0; l < 4; ++l) {
-                for (size_t k = 0; k < 32; ++k) {
-                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
-                }
-            }
-        }
-
-        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-
-        sumf += (float) sumi * d;
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1729,45 +1680,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2057,68 +1973,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -2431,61 +2291,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2578,66 +2391,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3093,47 +2854,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     }
     *s = sum;
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3229,34 +2953,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * sumf;
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3327,42 +3027,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * sumf;
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3455,45 +3123,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * sumf;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -3553,36 +3186,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.5f * sumf;
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3689,48 +3296,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3793,36 +3362,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3912,52 +3455,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-
-    int sum1[2], sum2[2], delta[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            delta[0] = qh[0] & 0x08 ? -1 : 1;
-            delta[1] = qh[0] & 0x80 ? -1 : 1;
-            delta[2] = qh[1] & 0x08 ? -1 : 1;
-            delta[3] = qh[1] & 0x80 ? -1 : 1;
-            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
-                int lsum1 = 0, lsum2 = 0;
-                for (int j = 0; j < 8; ++j) {
-                    lsum1 += q8[j] * grid[j];
-                    lsum2 += q8[j];
-                }
-                q8 += 8;
-                sum1[l/2] += lsum1;
-                sum2[l/2] += lsum2*delta[l];
-            }
-
-            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
-            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
-
-            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
-            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
-            qs += 4;
-            qh += 2;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(scale);
+    ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -4078,37 +3580,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = sumf;
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

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

@@ -86,35 +86,9 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
         }
     }
 #else
-    // scalar
-    const int blck_size_interleave = 4;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
 #endif
 }
 
@@ -205,35 +179,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -295,29 +243,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
+    ggml_gemv_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -383,29 +309,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
+    ggml_gemv_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -497,31 +401,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif // #if defined(__ARM_FEATURE_SVE)
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -591,31 +471,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1096,40 +952,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     );
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                                }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                            }
-                        }
-                    }
-                }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+    ggml_gemm_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1550,38 +1373,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     );
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    float sumf[4][4];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2019,38 +1811,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2126,38 +1887,5 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
-                                }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                            }
-                        }
-                    }
-                }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+    ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }

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

@@ -544,7 +544,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x1 ));
         const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats
@@ -821,24 +821,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc) + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -883,30 +874,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -954,30 +930,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc) + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1016,18 +977,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1103,45 +1061,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1239,70 +1162,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1391,61 +1257,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1541,66 +1360,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1678,47 +1445,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1815,34 +1545,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1978,42 +1684,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2105,47 +1779,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2209,36 +1847,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2338,48 +1950,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2460,36 +2034,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2603,37 +2151,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = hsum_float_8(accum);
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

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

@@ -201,24 +201,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
-    }
-
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -278,24 +268,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -360,30 +340,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -451,30 +415,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -535,18 +484,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -695,45 +641,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -907,70 +818,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1130,61 +984,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1342,66 +1149,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1556,47 +1311,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1737,34 +1455,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.125f * vec_extract(vsumf0, 0);
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1869,42 +1563,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2030,47 +1692,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2172,36 +1798,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * vec_extract(vsumf0, 0);
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2327,48 +1927,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2481,36 +2043,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2581,17 +2117,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
-        int sumi1 = 0, sumi2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
-            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
-        }
-        sumf += d * (sumi1 + sumi2);
-    }
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2696,37 +2230,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = vec_extract(vsumf0, 0);
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

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

@@ -116,6 +116,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 //===================================== Dot products =================================
 
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -132,7 +133,6 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     int ib = 0;
     float sumf = 0;
 
-#if defined(__riscv_v)
     size_t vl = qk / 2;
 
     for (; ib < nb; ++ib) {
@@ -164,27 +164,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_1;
     const int nb = n / qk;
 
@@ -201,7 +188,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     int ib = 0;
     float sumf = 0;
 
-#if defined(__riscv_v)
     size_t vl = qk / 2;
 
     for (; ib < nb; ++ib) {
@@ -229,27 +215,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -267,7 +240,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q5_0 * GGML_RESTRICT x = vx;
     const block_q8_0 * GGML_RESTRICT y = vy;
 
-#if defined(__riscv_v)
     size_t vl;
     size_t vlenb = __riscv_vlenb();
 
@@ -297,33 +269,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_1;
     const int nb = n / qk;
 
@@ -341,7 +294,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q5_1 * GGML_RESTRICT x = vx;
     const block_q8_1 * GGML_RESTRICT y = vy;
 
-#if defined(__riscv_v)
     size_t vl;
     size_t vlenb = __riscv_vlenb();
 
@@ -370,30 +322,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -431,18 +363,17 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -738,44 +669,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    float sumf = 0;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
 
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1147,68 +1045,14 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
 
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
 
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -1534,60 +1378,15 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(nb);
+    UNUSED(utmp);
 
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1698,65 +1497,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(nb);
+    UNUSED(utmp);
 
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2024,46 +1773,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
 
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

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

@@ -112,31 +112,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #endif
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -361,37 +337,6 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         return;
     }
 
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+#endif
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -172,24 +172,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = acc[0] + acc[1] + acc[2] + acc[3];
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -239,24 +230,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -298,18 +280,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = acc[0] + acc[1] + acc[2] + acc[3];
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -442,70 +421,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -600,61 +522,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -767,66 +642,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -969,47 +792,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1186,17 +972,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
-        int sumi1 = 0, sumi2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
-            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
-        }
-        sumf += d * (sumi1 + sumi2);
-    }
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1264,37 +1048,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = sumf;
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

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

@@ -435,30 +435,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
            wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -545,30 +530,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
            wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -628,18 +598,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
            wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -755,45 +722,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -902,68 +834,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -1089,61 +965,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1279,66 +1108,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1435,47 +1212,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -702,7 +702,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q8_1 * GGML_RESTRICT y = vy;
 
     int ib = 0;
-    float sumf = 0;
 
 #if defined(__AVX2__) || defined(__AVX__)
     // Initialize accumulator with zeros
@@ -737,26 +736,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
     }
 
-    sumf = hsum_float_8(acc) + summs;
-
+    *s = hsum_float_8(acc) + summs;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
-    *s = sumf;
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -764,7 +751,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const int nb = n / qk;
 
     int ib = 0;
-    float sumf = 0;
 
     assert(n % qk == 0);
     assert(qk == QK5_0);
@@ -799,7 +785,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_fmadd_ps(d, q, acc);
     }
 
-    sumf = hsum_float_8(acc);
+    *s = hsum_float_8(acc);
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -830,32 +816,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
     }
 
-    sumf = hsum_float_8(acc);
-
+    *s = hsum_float_8(acc);
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
-    *s = sumf;
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -863,7 +831,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const int nb = n / qk;
 
     int ib = 0;
-    float sumf = 0;
 
     assert(n % qk == 0);
     assert(qk == QK5_1);
@@ -901,7 +868,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
     }
 
-    sumf = hsum_float_8(acc) + summs;
+    *s = hsum_float_8(acc) + summs;
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -935,32 +902,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
     }
 
-    sumf = hsum_float_8(acc) + summs;
-
+    *s = hsum_float_8(acc) + summs;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
-    *s = sumf;
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1017,7 +966,6 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     }
 
     sumf = hsum_float_8(accum);
-
 #endif
     for (; ib < nb; ++ib) {
         int sumi = 0;
@@ -1157,44 +1105,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(sumf);
 
 #else
-    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
-
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int sum = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 32; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
-                }
-            }
-        }
-        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 16; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
-                }
-            }
-        }
-
-        for (size_t l = 0; l < 4; ++l) {
-            for (size_t j = 0; j < sizeof(x->qh); ++j) {
-                uint8_t q = x[i].qh[j] * pow3[l];
-                uint16_t xi = ((uint16_t) q * 3) >> 8;
-                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
-            }
-        }
-
-        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1257,25 +1171,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(sumf);
 
 #else
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int32_t sumi = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            for (size_t l = 0; l < 4; ++l) {
-                for (size_t k = 0; k < 32; ++k) {
-                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
-                }
-            }
-        }
-
-        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-
-        sumf += (float) sumi * d;
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1464,45 +1363,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1769,70 +1633,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2002,61 +1809,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2259,66 +2019,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + summs;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2520,47 +2228,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2712,34 +2383,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3033,42 +2680,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3250,47 +2865,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -3410,36 +2989,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3646,48 +3199,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3811,36 +3326,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -4043,52 +3532,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
 
 #else
-
-    int sum1[2], sum2[2], delta[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            delta[0] = qh[0] & 0x08 ? -1 : 1;
-            delta[1] = qh[0] & 0x80 ? -1 : 1;
-            delta[2] = qh[1] & 0x08 ? -1 : 1;
-            delta[3] = qh[1] & 0x80 ? -1 : 1;
-            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
-                int lsum1 = 0, lsum2 = 0;
-                for (int j = 0; j < 8; ++j) {
-                    lsum1 += q8[j] * grid[j];
-                    lsum2 += q8[j];
-                }
-                q8 += 8;
-                sum1[l/2] += lsum1;
-                sum2[l/2] += lsum2*delta[l];
-            }
-
-            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
-            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
-
-            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
-            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
-            qs += 4;
-            qh += 2;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(scale);
+    ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -4275,37 +3723,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = hsum_float_8(accum);
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 

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

@@ -281,35 +281,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -531,49 +505,9 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK_K];
-    float iscale[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-            float max = 0;
-
-            for (int j = 0; j < QK_K; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
-                // Update the maximum value of the corresponding super block
-                if(amax < fabsf(srcv[row_iter][j])) {
-                    amax = fabsf(srcv[row_iter][j]);
-                    max = srcv[row_iter][j];
-                }
-            }
-
-            iscale[row_iter] = amax ? -127.f/max : 0;
-
-            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
-        }
-
-        for (int j = 0; j < QK_K / 4; j++) {
-            y[i].bsums[j] = 0;
-        }
-
-        // Quants values are interleaved in sequence of eight bytes from corresponding super blocks
-        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
-        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
-        for (int j = 0; j < QK_K * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id     = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-            int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
-
-            float x0 = srcv[src_id][src_offset] * iscale[src_id];
-            y[i].qs[j] = nearest_int(x0);
-            y[i].bsums[index] += y[i].qs[j];
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_K_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -689,31 +623,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     return;
 
 #endif
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -932,61 +842,10 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #else
-
-    float sumf[8];
-    float sum_minf[8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
-
-    const block_q8_K * a_ptr = (const block_q8_K *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
-            sum_minf[j] = 0.0;
-        }
-        for (int l = 0; l < nb; l++) {
-            for (int sb = 0; sb < 8; sb++) {
-                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                utmp[sb * 4 + 2] = uaux_0;
-                utmp[sb * 4 + 0] &= kmask1;
-            }
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
-                uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi1 = 0;
-                    sumi2 = 0;
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]);
-                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]);
-                        sumi1 = sumi1 * scales_0[j];
-                        sumi2 = sumi2 * scales_1[j];
-                        sumi += sumi1 + sumi2;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-            for (int sb = 0; sb < 8; sb++) {
-                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
-        }
-    }
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 #endif
 }
 
@@ -1735,38 +1594,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -3216,70 +3044,9 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #else
-
-    float sumf[4][8];
-    float sum_minf[4][8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j] = 0.0;
-                    sum_minf[m][j] = 0.0;
-                }
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int sb = 0; sb < 8; sb++) {
-                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                    utmp[sb * 4 + 2] = uaux_0;
-                    utmp[sb * 4 + 0] &= kmask1;
-                }
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
-                    uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi1 = 0;
-                            sumi2 = 0;
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]);
-                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
-                                sumi1 = sumi1 * scales_0[j];
-                                sumi2 = sumi2 * scales_1[j];
-                                sumi += sumi1 + sumi2;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-                for (int sb = 0; sb < 8; sb++) {
-                    uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
-                    for(int m = 0; m < 4; m++) {
-                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
-                        for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
-                }
-            }
-        }
-    }
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 #endif
 }

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

@@ -22,9 +22,94 @@
 
 #include "kai_common.h"
 
+#include "simd-mappings.h"
+
 #include "kernels.h"
 
 #define NELEMS(x) sizeof(x) / sizeof(*x)
+
+static const size_t INT4_PER_BYTE = 2;
+static const size_t INT4_BITS     = 4;
+static const int Q4_0_ZERO_POINT  = 8;
+const size_t INT4_PER_UINT16      = 4;
+
+static void dequantize_row_qsi4c32pscalef16(
+    const void *packed_data,
+    int32_t row_idx,
+    int64_t nc,
+    float *out,
+    size_t nr_pack,
+    size_t packed_row_stride,
+    size_t kr,
+    size_t bl,
+    size_t num_bytes_multiplier
+) {
+    size_t group_idx = row_idx / nr_pack;
+    size_t row_in_group = row_idx % nr_pack;
+    const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
+    size_t num_blocks = nc / bl;
+    const uint8_t *block_ptr = packed_group;
+
+    for (size_t b = 0; b < num_blocks; ++b) {
+        uint16_t scale_f16 = *((const uint16_t *)(block_ptr + row_in_group * num_bytes_multiplier));
+        float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
+
+        const uint8_t *segment_ptr = block_ptr + nr_pack * num_bytes_multiplier;
+        size_t num_segments = bl / kr;
+        size_t num_bytes_per_segment = kr / INT4_PER_BYTE;
+
+        for (size_t s = 0; s < num_segments; ++s) {
+            const uint8_t *seg_base = segment_ptr + s * nr_pack * num_bytes_per_segment;
+            const uint8_t *qbytes = seg_base + row_in_group * num_bytes_per_segment;
+            for (size_t k = 0; k < num_bytes_per_segment; ++k) {
+                uint8_t byte = qbytes[k] ^ 0x88;
+                int x0 = (byte & 0x0F) - Q4_0_ZERO_POINT;
+                int x1 = (byte >> INT4_BITS) - Q4_0_ZERO_POINT;
+                out[b * bl + s * num_bytes_per_segment + k] = x0 * scale;
+                out[b * bl + s * num_bytes_per_segment + k + bl/2] = x1 * scale;
+            }
+        }
+        block_ptr += nr_pack * num_bytes_multiplier + num_segments * nr_pack * num_bytes_per_segment;
+    }
+}
+
+static void dequantize_row_qsi4c32ps1s0scalef16(
+    const void *packed_data,
+    int32_t row_idx,
+    int64_t k,
+    float *out,
+    size_t nr,
+    size_t packed_row_stride,
+    size_t kr,
+    size_t bl,
+    size_t num_bytes_multiplier
+) {
+    const size_t num_blocks = k / bl;
+    const size_t bl4 = bl / INT4_PER_UINT16;
+
+    size_t group_idx = row_idx / nr;
+    size_t row_in_group = row_idx % nr;
+
+    const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
+    const uint16_t *qdata = (const uint16_t *)packed_group;
+    const uint16_t *scales = (const uint16_t *)(packed_group + packed_row_stride - (nr * num_blocks * num_bytes_multiplier));
+
+    for (size_t block_idx = 0; block_idx < num_blocks; ++block_idx) {
+        uint16_t scale_f16 = scales[row_in_group + block_idx * nr];
+        float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
+
+        for (size_t bl4_idx = 0; bl4_idx < bl4; ++bl4_idx) {
+            uint16_t q = qdata[(block_idx * bl4 + bl4_idx) * nr + row_in_group];
+
+            for (size_t qidx = 0; qidx < INT4_PER_UINT16; ++qidx) {
+                int v = ((q >> (qidx * 4)) & 0xF) - Q4_0_ZERO_POINT;
+                out[block_idx * bl + bl4_idx * INT4_BITS + qidx] = v * scale;
+            }
+        }
+    }
+    GGML_UNUSED(kr);
+}
+
 static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
 #if defined(__ARM_FEATURE_SME)
     {
@@ -63,8 +148,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .to_float      = */ dequantize_row_qsi4c32ps1s0scalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -107,8 +194,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
-            /* .pack_func   = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .packed_stride = */ NULL,
+            /* .pack_func     = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .to_float      = */ NULL,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -154,8 +243,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -200,8 +291,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -247,8 +340,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -293,8 +388,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
         },
         /* .rhs_info = */ {
-            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,

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

@@ -71,12 +71,15 @@ struct rhs_packing_info {
         std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
         std::function<size_t(size_t n, size_t k)>
     > packed_size;
+    size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
     std::variant<
         std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
             const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
         std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
             const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
     > pack_func;
+    void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out, size_t nr_pack, size_t packed_row_stride,
+          size_t kr, size_t bl, size_t num_bytes_multiplier);
 };
 
 struct ggml_kleidiai_kernels {

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

@@ -40,6 +40,17 @@ struct ggml_kleidiai_context {
     ggml_kleidiai_kernels * kernels;
 } static ctx = { CPU_FEATURE_NONE, NULL };
 
+static const char* cpu_feature_to_string(cpu_feature f) {
+    switch (f) {
+        case CPU_FEATURE_NONE:    return "NONE";
+        case CPU_FEATURE_DOTPROD: return "DOTPROD";
+        case CPU_FEATURE_I8MM:    return "I8MM";
+        case CPU_FEATURE_SVE:     return "SVE";
+        case CPU_FEATURE_SME:     return "SME";
+        default:                  return "UNKNOWN";
+    }
+}
+
 static void init_kleidiai_context(void) {
 
     ggml_critical_section_start();
@@ -62,6 +73,11 @@ static void init_kleidiai_context(void) {
             ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
         }
         ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
+#ifndef NDEBUG
+        if (ctx.kernels) {
+            GGML_LOG_DEBUG("kleidiai: using kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels->required_cpu));
+        }
+#endif
     }
     ggml_critical_section_end();
 }
@@ -102,6 +118,9 @@ static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint1
 
 class tensor_traits : public ggml::cpu::tensor_traits {
     bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
+        if (op->op != GGML_OP_MUL_MAT) {
+            return false;
+        }
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
         GGML_ASSERT(kernels);
         kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
@@ -135,6 +154,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             } else if (dst->src[0]->type == GGML_TYPE_F16) {
                 return compute_forward_kv_cache(params, dst);
             }
+        } else if (dst->op == GGML_OP_GET_ROWS) {
+            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+                return compute_forward_get_rows(params, dst);
+            }
         }
         return false;
     }
@@ -270,6 +293,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
     }
 
     bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
+
         const ggml_tensor * src0 = dst->src[0];
         const ggml_tensor * src1 = dst->src[1];
 
@@ -342,8 +367,49 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         return true;
     }
 
+    bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
+        GGML_ASSERT(ctx.kernels);
+
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
+        kernel_info * kernel        = &ctx.kernels->gemm;
+
+        const int64_t nc     = ne00;
+        const int64_t nr     = ggml_nelements(src1);
+
+        const size_t block_rows = kernel->get_nr();
+        const size_t kr         = kernel->get_kr();
+
+        const size_t num_bytes_multiplier = sizeof(uint16_t);
+        const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, QK4_0);
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const int dr = (nr + nth - 1) / nth;
+        const int ir0 = dr * ith;
+        const int ir1 = MIN(ir0 + dr, nr);
+
+        for (int64_t i = ir0; i < ir1; ++i) {
+            GGML_ASSERT(src1->type == GGML_TYPE_I32);
+            int64_t row_idx = ((const int32_t *)src1->data)[i];
+            GGML_ASSERT(row_idx >= 0 && row_idx < src0->ne[1]);
+
+            float *out = (float *)((char *)dst->data + i * nb1);
+            rhs_info->to_float(src0->data, row_idx, nc, out, block_rows, packed_stride, kr, QK4_0, num_bytes_multiplier);
+        }
+
+        return true;
+    }
+
 public:
     int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
+        GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
         GGML_ASSERT(ctx.kernels);
         const size_t n = tensor->ne[1];
         const size_t k = tensor->ne[0];
@@ -351,17 +417,12 @@ public:
         size_t kr      = ctx.kernels->gemm.get_kr();
         size_t sr      = ctx.kernels->gemm.get_sr();
 
-#ifndef NDEBUG
-        const size_t repacked_size = variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
-        GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
-#endif
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
         variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
 
         return 0;
-
         GGML_UNUSED(data_size);
     }
 };
@@ -375,8 +436,8 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc
 static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
 
-    GGML_UNUSED(buffer);
     return GGML_STATUS_SUCCESS;
+    GGML_UNUSED(buffer);
 }
 
 static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
@@ -418,18 +479,35 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
     GGML_UNUSED(buft);
 }
 
+static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(ctx.kernels);
+
+    const size_t n  = tensor->ne[1];
+    const size_t k  = tensor->ne[0];
+    const size_t nr = ctx.kernels->gemm.get_nr();
+    const size_t kr = ctx.kernels->gemm.get_kr();
+
+    return variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
+
+    GGML_UNUSED(buft);
+}
+
 namespace ggml::cpu::kleidiai {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        if (op->op == GGML_OP_MUL_MAT &&
+        if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
             op->src[0]->type == GGML_TYPE_Q4_0 &&
             op->src[0]->buffer &&
             (ggml_n_dims(op->src[0]) == 2) &&
             op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
+            if (op->op == GGML_OP_GET_ROWS && op->src[1]->ne[0] != 8) {
+                return false;
+            }
             if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                 return false;
             }
-            if (op->src[1]->type == GGML_TYPE_F32 &&
+            if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) &&
                 ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) {
                 return true;
             }
@@ -438,7 +516,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
     }
 
     ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
-        if (op->op == GGML_OP_MUL_MAT) {
+        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }
@@ -469,7 +547,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) {
                            /* .alloc_buffer     = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer,
                            /* .get_alignment    = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment,
                            /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
-                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .get_alloc_size   = */ ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size,
                            /* .is_host          = */ nullptr,
                            },
         /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),

ggml/src/ggml-cpu/ops.cpp

@@ -4015,6 +4015,9 @@ static void ggml_compute_forward_rms_norm_f32(
 
                 const float scale = 1.0f/sqrtf(mean + eps);
 
+                // if you hit this, likely you got an inf somewhere earlier
+                assert(scale > 0.0f);
+
                 ggml_vec_scale_f32(ne00, y, scale);
             }
         }

ggml/src/ggml-cpu/repack.cpp

@@ -14,7 +14,6 @@
 #include <cmath>
 #include <cstring>
 #include <cassert>
-#include <cstdlib> // for qsort
 #include <cstdio>  // for GGML_ASSERT
 
 #include "repack.h"

ggml/src/ggml-cpu/vec.cpp

@@ -221,6 +221,9 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
     for (int i = np; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
+
+    // if you hit this, you are likely running outside the FP range
+    assert(!isnan(sumf) && !isinf(sumf));
 #else
     for (int i = 0; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));

ggml/src/ggml-cuda/CMakeLists.txt

@@ -102,12 +102,12 @@ if (CUDAToolkit_FOUND)
     if (GGML_STATIC)
         if (WIN32)
             # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
-            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas)
         else ()
-            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static)
         endif()
     else()
-        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas)
     endif()
 
     if (GGML_CUDA_NO_VMM)

ggml/src/ggml-cuda/common.cuh

@@ -56,7 +56,7 @@
 #define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
 #define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
 #define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
-#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
+#define GGML_CUDA_CC_CDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x910)  // MI210, minimum acc register renameing
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
 
@@ -72,8 +72,9 @@
 #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
 #define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
-#define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
-#define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
+#define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
 #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
@@ -175,7 +176,7 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 #    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes)                                                       \
         do {                                                                                                   \
             static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = { false };                         \
@@ -190,7 +191,7 @@ static const char * cu_get_error_str(CUresult err) {
         do {                                             \
             GGML_UNUSED(nbytes);                         \
         } while (0)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !(defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
@@ -210,9 +211,9 @@ typedef float2 dfloat2;
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 
-#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
@@ -226,13 +227,17 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
-#define NEW_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#if defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
+#define AMD_MFMA_AVAILABLE
+#endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#define NEW_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #define CP_ASYNC_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
 #if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 #define FLASH_ATTN_AVAILABLE
@@ -254,7 +259,7 @@ static bool fast_fp16_hardware_available(const int cc) {
 
 // Any FP16 tensor core instructions are available for ggml code.
 static bool fp16_mma_available(const int cc) {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+#if defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
     return false;
 #else
     if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
@@ -270,7 +275,7 @@ static bool fp16_mma_available(const int cc) {
     } else {
         return false;
     }
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+#endif // defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -288,6 +293,14 @@ static bool fp32_mma_hardware_available(const int cc) {
     return GGML_CUDA_CC_IS_CDNA(cc);
 }
 
+static bool amd_mfma_available(const int cc) {
+#if !defined(GGML_HIP_NO_MMQ_MFMA)
+    return GGML_CUDA_CC_IS_CDNA(cc);
+#else
+    return false;
+#endif //!defined(GGML_HIP_NO_MMQ_MFMA)
+}
+
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
 static bool new_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
@@ -298,25 +311,25 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+#if defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
     return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+#endif // defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
     printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
            file_name, line, function_name, arch);
     GGML_UNUSED(arch_list);
 #else
     printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
            file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
@@ -353,7 +366,7 @@ struct ggml_cuda_unroll<1> {
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
@@ -361,7 +374,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
         x += __shfl_xor_sync(0xffffffff, x, offset, width);
     }
     return x;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 }
 
 template<int width = WARP_SIZE>
@@ -418,6 +431,20 @@ static __global__ void reduce_rows_f32(const float * x, float * dst, const int n
     dst[row] = norm ? sum / ncols : sum;
 }
 
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ int warp_reduce_all(int x) {
+#ifdef GGML_USE_HIP
+#pragma unroll
+    for (int offset = width/2; offset > 0; offset >>= 1) {
+        x = x && __shfl_xor_sync(0xffffffff, x, offset, width);
+    }
+    return x;
+#else
+    static_assert(width == WARP_SIZE, "width != WARP_SIZE not implemented");
+    return __all_sync(0xffffffff, x);
+#endif // GGML_USE_HIP
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -430,11 +457,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !defined(GGML_USE_HIP) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !defined(GGML_USE_HIP) && CUDART_VERSION < CUDART_HMAX
 
 #else
    NO_DEVICE_CODE;
@@ -462,7 +489,7 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
 #pragma unroll
    for (int offset = width/2; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, width));
@@ -471,7 +498,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
 }
 
 #if CUDART_VERSION < CUDART_HMASK
@@ -483,7 +510,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
 #endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3) || defined(RDNA4)
@@ -509,7 +536,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif
     return c;
 
-#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP)
 
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
     return __dp4a(a, b, c);
@@ -519,7 +546,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
 
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 }
 
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
@@ -765,7 +792,7 @@ struct ggml_tensor_extra_gpu {
 };
 
 
-#if (defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS))
+#if (defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)) || defined(GGML_MUSA_GRAPHS)
 #define USE_CUDA_GRAPH
 #endif
 

ggml/src/ggml-cuda/convert.cu

@@ -6,24 +6,33 @@
 #define CUDA_Q8_0_NE_ALIGN 2048
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x);
+static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    const int64_t i00 = 2 * (int64_t(blockDim.x)*blockIdx.x + threadIdx.x);
 
-    if (i >= k) {
+    if (i00 >= ne00) {
         return;
     }
 
-    const int64_t ib = i/qk; // block index
-    const int64_t iqs = (i%qk)/qr; // quant index
-    const int64_t iybs = i - i%qk; // y block start index
+    const int64_t i01 = blockIdx.y;
+    const int64_t i02 = blockIdx.z % ne02;
+    const int64_t i03 = blockIdx.z / ne02;
+
+    const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+
+    const int64_t ib = ibx0 + i00/qk; // block index
+    const int64_t iqs = (i00%qk)/qr; // quant index
+    const int64_t iybs = i00 - i00%qk; // y block start index
     const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
     dfloat2 v;
     dequantize_kernel(vx, ib, iqs, v);
 
-    y[iybs + iqs + 0]        = v.x;
-    y[iybs + iqs + y_offset] = v.y;
+    const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
+    y[iy0 + 0]        = float(v.x);
+    y[iy0 + y_offset] = float(v.y);
 }
 
 template <bool need_check>
@@ -457,9 +466,17 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
-    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+static void dequantize_block_cuda(const void * vx, dst_t * y,
+        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, cudaStream_t stream) {
+    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, ne02*ne03);
+    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
+        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_cont_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    dequantize_block_cuda<qk, qr, dequantize_kernel, dst_t>(vx, y, k, 1, 1, 1, k/qk, k/qk, k/qk, stream);
 }
 
 static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) {
@@ -624,14 +641,14 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_Q4_1:
             return dequantize_row_q4_1_cuda;
         case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+            return dequantize_block_cont_cuda<QK5_0, QR5_0, dequantize_q5_0>;
         case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+            return dequantize_block_cont_cuda<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
             if (fp16_available(ggml_cuda_info().devices[ggml_cuda_get_device()].cc)) {
                 return dequantize_block_q8_0_f16_cuda;
             }
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+            return dequantize_block_cont_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_cuda;
         case GGML_TYPE_Q3_K:
@@ -676,11 +693,11 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
         case GGML_TYPE_Q4_1:
             return dequantize_row_q4_1_cuda;
         case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+            return dequantize_block_cont_cuda<QK5_0, QR5_0, dequantize_q5_0>;
         case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+            return dequantize_block_cont_cuda<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+            return dequantize_block_cont_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_cuda;
         case GGML_TYPE_Q3_K:
@@ -722,6 +739,16 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16>;
         default:
@@ -733,6 +760,16 @@ to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half, nv_bfloat16>;
         default:
@@ -744,6 +781,16 @@ to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F16:
             return convert_unary_cuda<half, float>;
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16, float>;
         default:

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

@@ -0,0 +1,225 @@
+#pragma once
+
+#include "ggml-common.h"
+
+template<typename src_t, typename dst_t>
+static __device__ __forceinline__ void convert_flt(const src_t * src, dst_t * dst) {
+    if constexpr (std::is_same_v<src_t, dst_t>) {
+        *dst = *src;
+    } else {
+        *dst = float(*src);
+    }
+}
+
+static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static __device__ void quantize_f32_q4_0_block(const float * __restrict__ x, block_q4_0 * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = x[0       + j]*id;
+        const float x1 = x[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
+
+        y->qs[j]  = xi0;
+        y->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void quantize_f32_q4_1_block(const float * __restrict__ x, block_q4_1 * __restrict__ y) {
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = x[j];
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->dm.x = d;
+    y->dm.y = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (x[0       + j] - vmin)*id;
+        const float x1 = (x[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
+
+        y->qs[j]  = xi0;
+        y->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void quantize_f32_q5_0_block(const float * __restrict__ x, block_q5_0 * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK5_0; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = x[0       + j]*id;
+        const float x1 = x[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
+
+        y->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+    memcpy(y->qh, &qh, sizeof(qh));
+}
+
+static __device__ void quantize_f32_q5_1_block(const float * __restrict__ x, block_q5_1 * __restrict__ y) {
+    float min = x[0];
+    float max = x[0];
+
+    for (int j = 1; j < QK5_1; ++j) {
+        const float v = x[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->dm.x = d;
+    y->dm.y = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (x[0       + j] - min)*id;
+        const float x1 = (x[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        y->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+    memcpy(y->qh, &qh, sizeof(qh));
+}
+
+static __device__ void quantize_f32_q8_0_block(const float * __restrict__ x, block_q8_0 * __restrict__ y) {
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = x[j];
+        amax = fmaxf(amax, fabsf(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = x[j]*id;
+        y->qs[j] = roundf(x0);
+    }
+}
+
+static __device__ void quantize_f32_iq4_nl_block(const float * __restrict__ x, block_iq4_nl * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_NL; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    float d = vmax / kvalues_iq4nl[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = x[0        + j]*id;
+        const float x1 = x[QK4_NL/2 + j]*id;
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
+        y->qs[j] = xi0 | (xi1 << 4);
+        const float v0 = kvalues_iq4nl[xi0];
+        const float v1 = kvalues_iq4nl[xi1];
+        const float w0 = x[0        + j]*x[0        + j];
+        const float w1 = x[QK4_NL/2 + j]*x[QK4_NL/2 + j];
+        sumqx += w0*v0*x[j] + w1*v1*x[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+    }
+
+    y->d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+// Wrapper functions for cpy.cu compatibility
+static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    quantize_f32_q4_0_block((const float *)cxi, (block_q4_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    quantize_f32_q4_1_block((const float *)cxi, (block_q4_1 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
+    quantize_f32_q5_0_block((const float *)cxi, (block_q5_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
+    quantize_f32_q5_1_block((const float *)cxi, (block_q5_1 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    quantize_f32_q8_0_block((const float *)cxi, (block_q8_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
+    quantize_f32_iq4_nl_block((const float *)cxi, (block_iq4_nl *)cdsti);
+}
+
+template<typename src_t, typename dst_t>
+static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
+    convert_flt((const src_t *)cxi, (dst_t *)cdsti);
+}

ggml/src/ggml-cuda/cpy.cu

@@ -1,51 +1,17 @@
 #include "cpy.cuh"
 #include "dequantize.cuh"
-#ifdef GGML_USE_MUSA
+#include "cpy-utils.cuh"
+#if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
 #include "ggml-musa/mudnn.cuh"
-#endif // GGML_USE_MUSA
+#endif // GGML_USE_MUSA && GGML_MUSA_MUDNN_COPY
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
-static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    float * dsti = (float *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    half * dsti = (half *) cdsti;
-
-    *dsti = __float2half(*xi);
-}
-
-static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
-    const half * xi = (const half *) cxi;
-    half * dsti = (half *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
-    const half * xi = (const half *) cxi;
-    float * dsti = (float *) cdsti;
-
-    *dsti = *xi;
-}
-
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
-                                   const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-                                   const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                   const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne,
+                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                               const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -71,29 +37,6 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const in
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
-static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q8_0 * dsti = (block_q8_0 *) cdsti;
-
-    float amax = 0.0f; // absolute max
-
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = xi[j];
-        amax = fmaxf(amax, fabsf(v));
-    }
-
-    const float d = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = xi[j]*id;
-
-        dsti->qs[j] = roundf(x0);
-    }
-}
-
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
@@ -106,139 +49,6 @@ static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_0 * dsti = (block_q4_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -8;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK4_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK4_0/2 + j]*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_1 * dsti = (block_q4_1 *) cdsti;
-
-    float vmin = FLT_MAX;
-    float vmax = -FLT_MAX;
-
-    for (int j = 0; j < QK4_1; ++j) {
-        const float v = xi[j];
-
-        if (v < vmin) vmin = v;
-        if (v > vmax) vmax = v;
-    }
-
-    const float d  = (vmax - vmin) / ((1 << 4) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = vmin;
-
-    for (int j = 0; j < QK4_1/2; ++j) {
-        const float x0 = (xi[0       + j] - vmin)*id;
-        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q5_0 * dsti = (block_q5_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK5_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -16;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK5_0/2 + j]*id;
-
-        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
-        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
-
-        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
-static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q5_1 * dsti = (block_q5_1 *) cdsti;
-
-    float min = xi[0];
-    float max = xi[0];
-
-    for (int j = 1; j < QK5_1; ++j) {
-        const float v = xi[j];
-        min = v < min ? v : min;
-        max = v > max ? v : max;
-    }
-
-    const float d  = (max - min) / 31;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = min;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_1/2; ++j) {
-        const float x0 = (xi[0       + j] - min)*id;
-        const float x1 = (xi[QK5_1/2 + j] - min)*id;
-
-        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
 template<dequantize_kernel_t dequant, int qk>
 static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
@@ -252,53 +62,6 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
-static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_NL; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    float d = vmax / kvalues_iq4nl[0];
-    const float id = d ? 1.0f/d : 0.0f;
-
-    float sumqx = 0, sumq2 = 0;
-    for (int j = 0; j < QK4_NL/2; ++j) {
-        const float x0 = xi[0        + j]*id;
-        const float x1 = xi[QK4_NL/2 + j]*id;
-        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
-        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
-        dsti->qs[j] = xi0 | (xi1 << 4);
-        const float v0 = kvalues_iq4nl[xi0];
-        const float v1 = kvalues_iq4nl[xi1];
-        const float w0 = xi[0        + j]*xi[0        + j];
-        const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j];
-        sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j];
-        sumq2 += w0*v0*v0 + w1*v1*v1;
-    }
-
-    dsti->d = sumq2 > 0 ? sumqx/sumq2 : d;
-}
-
 template <cpy_kernel_t cpy_blck, int qk>
 static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -358,7 +121,7 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int
 // Copy destination pointers to GPU to be available when pointer indirection is in use
 
 void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) {
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
     if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers
         CUDA_CHECK(cudaStreamSynchronize(stream));
         if (cuda_graph->dest_ptrs_d != nullptr) {
@@ -376,43 +139,14 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
 #endif
 }
 
-static void ggml_cpy_f16_f32_cuda(
+template<typename src_t, typename dst_t>
+static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_f32_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_bf16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+    cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
@@ -544,16 +278,6 @@ static void ggml_cpy_f32_iq4_nl_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-static void ggml_cpy_f16_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -590,7 +314,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 
     char ** dest_ptrs_d = nullptr;
     int graph_cpynode_index = -1;
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
     if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
@@ -600,20 +324,20 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 #endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
-#ifdef GGML_USE_MUSA
+#if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
         if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
             CUDA_CHECK(mudnnMemcpyAsync(ctx, src1, src0));
         } else
-#endif // GGML_USE_MUSA
+#endif // GGML_USE_MUSA && GGML_MUSA_MUDNN_COPY
         {
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
@@ -640,14 +364,22 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
-#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) || defined(GGML_MUSA_GRAPHS)
     if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
     }
@@ -667,11 +399,11 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         return nullptr;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+        return (void*) cpy_flt<cpy_1_flt<float, float>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
+        return (void*) cpy_flt<cpy_1_flt<float, nv_bfloat16>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+        return (void*) cpy_flt<cpy_1_flt<float, half>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
@@ -695,9 +427,17 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>;
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+        return (void*) cpy_flt<cpy_1_flt<half, half>>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<half, nv_bfloat16>>;
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_f32_f16<cpy_1_f16_f32>;
+        return (void*) cpy_flt<cpy_1_flt<half, float>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, half>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));

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

@@ -15,6 +15,7 @@ typedef void (* fattn_kernel_t)(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -23,31 +24,13 @@ typedef void (* fattn_kernel_t)(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3);
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33);
 
 typedef half (*vec_dot_KQ_f16_t)(
     const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
@@ -518,10 +501,59 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
         nullptr;
 }
 
+template <int ncols1>
+__launch_bounds__(FATTN_KQ_STRIDE/2, 1)
+static __global__ void flash_attn_mask_to_KV_max(
+        const half2 * __restrict__ mask, int * __restrict__ KV_max, const int ne30, const int s31, const int s33) {
+    const int ne31     = gridDim.x;
+    const int tid      = threadIdx.x;
+    const int sequence = blockIdx.y;
+    const int jt       = blockIdx.x;
+
+    mask += sequence*s33 + jt*ncols1*s31;
+
+    __shared__ int buf_iw[WARP_SIZE];
+    if (tid < WARP_SIZE) {
+        buf_iw[tid] = 1;
+    }
+    __syncthreads();
+
+    int KV_max_sj = (ne30 - 1) * FATTN_KQ_STRIDE;
+    for (; KV_max_sj >= 0; KV_max_sj -= FATTN_KQ_STRIDE) {
+        int all_inf = 1;
+
+#pragma unroll
+        for (int j = 0; j < ncols1; ++j) {
+            const float2 tmp = __half22float2(mask[j*s31 + KV_max_sj/2 + tid]);
+            all_inf = all_inf && int(isinf(tmp.x)) && int(isinf(tmp.y));
+        }
+
+        all_inf = warp_reduce_all(all_inf);
+        if (tid % WARP_SIZE == 0) {
+            buf_iw[tid / WARP_SIZE] = all_inf;
+        }
+        __syncthreads();
+        all_inf = buf_iw[tid % WARP_SIZE];
+        __syncthreads();
+        all_inf = warp_reduce_all(all_inf);
+
+        if (!all_inf) {
+            KV_max_sj += FATTN_KQ_STRIDE;
+            break;
+        }
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    KV_max[sequence*ne31 + jt] = KV_max_sj;
+}
+
 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 ne11) {
+        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11) {
     constexpr int ncols = ncols1*ncols2;
 
     const int bidx0 = blockIdx.x;
@@ -535,8 +567,8 @@ static __global__ void flash_attn_stream_k_fixup(
     const int iter_k = ne11 / FATTN_KQ_STRIDE;
     const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
 
-    const int kbc0      = (bidx0 + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
-    const int kbc0_stop = (bidx0 + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+    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;
 
     const bool did_not_have_any_data   = kbc0 == kbc0_stop;
     const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -545,14 +577,15 @@ static __global__ void flash_attn_stream_k_fixup(
         return;
     }
 
-    const int channel = kbc0 / (iter_k*iter_j);
-    const int jt      = (kbc0 - channel*iter_k*iter_j) / iter_k;
+    const int sequence = kbc0 / (iter_k*iter_j*(ne02/ncols2));
+    const int head = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
+    const int jt = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
 
     if (jt*ncols1 + j >= ne01) {
         return;
     }
 
-    dst += jt*ne02*(ncols1*D) + channel*(ncols2*D) + (j*ne02 + c)*D + tid;
+    dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + head*(ncols2*D) + (j*ne02 + c)*D + tid;
 
     // Load the partial result that needs a fixup:
     float dst_val = 0.0f;
@@ -571,7 +604,7 @@ static __global__ void flash_attn_stream_k_fixup(
     int bidx = bidx0 - 1;
     int kbc_stop = kbc0;
     while(true) {
-        const int kbc = bidx*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+        const int kbc = bidx*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
         if (kbc == kbc_stop) { // Did not have any data.
             bidx--;
             kbc_stop = kbc;
@@ -609,24 +642,39 @@ static __global__ void flash_attn_stream_k_fixup(
 }
 
 template<int D> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP)
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,
         float * __restrict__ dst,
         const int parallel_blocks) {
-    VKQ_parts += parallel_blocks*D * gridDim.z*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.z*blockIdx.x;
-    dst       +=                 D * gridDim.z*blockIdx.x;
+    // Dimension 0: threadIdx.x
+    // Dimension 1: blockIdx.x
+    // Dimension 2: blockIdx.y
+    // Dimension 3: blockIdx.z
+    // Memory layout is permuted with [0, 2, 1, 3]
+
+    const int ne01 = gridDim.x;
+    const int ne02 = gridDim.y;
+
+    const int col      = blockIdx.x;
+    const int head     = blockIdx.y;
+    const int sequence = blockIdx.z;
+
+    const int j_dst_unrolled = (sequence*ne01 + col)*ne02 + head;
+
+    VKQ_parts += j_dst_unrolled * parallel_blocks*D;
+    VKQ_meta  += j_dst_unrolled * parallel_blocks;
+    dst       += j_dst_unrolled *                 D;
 
     const int tid = threadIdx.x;
     __builtin_assume(tid < D);
 
     extern __shared__ float2 meta[];
     for (int i = tid; i < 2*parallel_blocks; i += D) {
-        ((float *) meta)[i] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + i];
+        ((float *) meta)[i] = ((const float *)VKQ_meta) [i];
     }
 
     __syncthreads();
@@ -644,11 +692,11 @@ static __global__ void flash_attn_combine_results(
         const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
         *((uint32_t *) &KQ_max_scale) &= ftz_mask;
 
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.z*D + blockIdx.z*D + tid];
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*D + tid];
         VKQ_denominator += KQ_max_scale * meta[l].y;
     }
 
-    dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
+    dst[tid] = VKQ_numerator / VKQ_denominator;
 }
 
 [[noreturn]]
@@ -705,8 +753,6 @@ void launch_fattn(
 
     GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
 
-    GGML_ASSERT(Q->ne[3] == 1);
-
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
     const int id  = ggml_cuda_get_device();
@@ -715,6 +761,7 @@ void launch_fattn(
 
     ggml_cuda_pool_alloc<half>   K_f16(pool);
     ggml_cuda_pool_alloc<half>   V_f16(pool);
+    ggml_cuda_pool_alloc<int>    KV_max(pool);
     ggml_cuda_pool_alloc<float>  dst_tmp(pool);
     ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
 
@@ -729,40 +776,84 @@ void launch_fattn(
     size_t nb23 = V ? V->nb[3] : nb13;
 
     if (need_f16_K && K->type != GGML_TYPE_F16) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(K));
-        K_f16.alloc(ggml_nelements(K));
-        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
-        to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
-        K_data = (char *) K_f16.ptr;
-
         const size_t bs = ggml_blck_size(K->type);
         const size_t ts = ggml_type_size(K->type);
 
-        nb11 = nb11*bs*sizeof(half)/ts;
-        nb12 = nb12*bs*sizeof(half)/ts;
-        nb13 = nb13*bs*sizeof(half)/ts;
+        K_f16.alloc(ggml_nelements(K));
+        if (ggml_is_contiguously_allocated(K)) {
+            to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
+            to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
+
+            nb11 = nb11*bs*sizeof(half)/ts;
+            nb12 = nb12*bs*sizeof(half)/ts;
+            nb13 = nb13*bs*sizeof(half)/ts;
+        } else {
+            GGML_ASSERT(K->nb[0] == ts);
+            to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(K->type);
+            const int64_t s01 = nb11 / ts;
+            const int64_t s02 = nb12 / ts;
+            const int64_t s03 = nb13 / ts;
+            to_fp16(K_data, K_f16.ptr, K->ne[0], K->ne[1], K->ne[2], K->ne[3], s01, s02, s03, main_stream);
+
+            nb11 = K->ne[0] * sizeof(half);
+            nb12 = K->ne[1] * nb11;
+            nb13 = K->ne[2] * nb12;
+        }
+        K_data = (char *) K_f16.ptr;
     }
 
     if (V && need_f16_V && V->type != GGML_TYPE_F16) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(V));
-        V_f16.alloc(ggml_nelements(V));
-        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
-        to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
-        V_data = (char *) V_f16.ptr;
-
         const size_t bs = ggml_blck_size(V->type);
         const size_t ts = ggml_type_size(V->type);
 
-        nb21 = nb21*bs*sizeof(half)/ts;
-        nb22 = nb22*bs*sizeof(half)/ts;
-        nb23 = nb23*bs*sizeof(half)/ts;
-    }
+        V_f16.alloc(ggml_nelements(V));
+        if (ggml_is_contiguously_allocated(V)) {
+            to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
+            to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
+            V_data = (char *) V_f16.ptr;
 
-    int parallel_blocks = 1;
+            nb21 = nb21*bs*sizeof(half)/ts;
+            nb22 = nb22*bs*sizeof(half)/ts;
+            nb23 = nb23*bs*sizeof(half)/ts;
+        } else {
+            GGML_ASSERT(V->nb[0] == ts);
+            to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(V->type);
+            const int64_t s01 = nb21 / ts;
+            const int64_t s02 = nb22 / ts;
+            const int64_t s03 = nb23 / ts;
+            to_fp16(V_data, V_f16.ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3], s01, s02, s03, main_stream);
+
+            nb21 = V->ne[0] * sizeof(half);
+            nb22 = V->ne[1] * nb21;
+            nb23 = V->ne[2] * nb22;
+        }
+        V_data = (char *) V_f16.ptr;
+    }
 
     const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
     const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
 
+    // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
+    // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
+    //     multiple sequences of possibly different lengths.
+    if (mask && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+        const int s31 = mask->nb[1] / sizeof(half2);
+        const int s33 = mask->nb[3] / sizeof(half2);
+
+        const dim3 blocks_num_KV_max(ntiles_x, Q->ne[3], 1);
+        const dim3 block_dim_KV_max(FATTN_KQ_STRIDE/2, 1, 1);
+
+        const int ne_KV_max = blocks_num_KV_max.x*blocks_num_KV_max.y;
+        const int iter_k = K->ne[1] / FATTN_KQ_STRIDE;
+
+        KV_max.alloc(ne_KV_max);
+        flash_attn_mask_to_KV_max<ncols1><<<blocks_num_KV_max, block_dim_KV_max, 0, main_stream>>>
+            ((const half2 *) mask->data, KV_max.ptr, iter_k, s31, s33);
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    int parallel_blocks = 1;
+
     const dim3 block_dim(warp_size, nwarps, 1);
     int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
     CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
@@ -849,16 +940,14 @@ void launch_fattn(
         K_data,
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
+        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],
-        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
-        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
-        Q->nb[1], Q->nb[2], Q->nb[3],
-        nb11, nb12, nb13,
+        Q->ne[0], Q->ne[1], 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,
-        KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0, mask ? mask->ne[3] : 0,
+        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0, mask ? mask->nb[3] : 0
     );
     CUDA_CHECK(cudaGetLastError());
 
@@ -869,11 +958,11 @@ 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], K->ne[1]);
+                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1]);
         }
     } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(DV, 1, 1);
-        const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
+        const dim3 blocks_num_combine(Q->ne[1], Q->ne[2], Q->ne[3]);
         const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
         flash_attn_combine_results<DV>

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

@@ -392,7 +392,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
     }
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles,
+    bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -408,7 +409,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const int stride_K,
         const int stride_V,
         const int stride_mask,
-        const int jt,
         half2        * const __restrict__ tile_Q,
         half2        * const __restrict__ tile_K,
         half2        * const __restrict__ tile_V,
@@ -455,7 +455,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         cp_async_wait_all();
         __syncthreads();
         flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, c::nbatch_fa, use_cp_async>
-            (V_h2 + k_VKQ_0*stride_V, tile_V, nbatch_V2, stride_V);
+            (V_h2 + int64_t(k_VKQ_0)*stride_V, tile_V, nbatch_V2, stride_V);
     } else {
         constexpr bool use_cp_async = nstages == 1;
         if (ncols2 > 1 || mask_h2) {
@@ -471,7 +471,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         if (nstages <= 1) {
             constexpr bool use_cp_async = nstages == 1;
             flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-                (K_h2 + k_VKQ_0*stride_K + k0_start, tile_K, k0_diff, stride_K);
+                (K_h2 + int64_t(k_VKQ_0)*stride_K + k0_start, tile_K, k0_diff, stride_K);
             if (use_cp_async) {
                 cp_async_wait_all();
             }
@@ -715,7 +715,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     (mask_h2 + (k_VKQ_0 + c::nbatch_fa)/2, tile_mask, stride_mask);
             }
             flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-                (K_h2 + (k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K);
+                (K_h2 + int64_t(k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K);
         }
     }
 
@@ -732,7 +732,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         if (nstages <= 1 && i0_start < reusable_cutoff) {
             constexpr bool use_cp_async = nstages == 1;
             flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, c::nbatch_fa, use_cp_async>
-                (V_h2 + k_VKQ_0*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V);
+                (V_h2 + int64_t(k_VKQ_0)*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V);
             if (use_cp_async) {
                 cp_async_wait_all();
             }
@@ -771,8 +771,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
     GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V);
-    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
-    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
+    GGML_UNUSED(stride_mask); GGML_UNUSED(tile_K);
     GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
     GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
     GGML_UNUSED(kb0); GGML_UNUSED(tile_Q);
@@ -920,21 +919,22 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                 (mask_h2 + kb0_start*c::nbatch_fa/2, tile_mask, stride_mask);
         }
         flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-            (K_h2 + kb0_start*c::nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K);
+            (K_h2 + int64_t(kb0_start)*c::nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K);
     }
 
     // Iterate over ne11 == previous tokens:
-    for (int kb0 = kb0_start; kb0 < kb0_stop-1; ++kb0) {
+    int kb0 = kb0_start;
+    for (; kb0 < kb0_stop-1; ++kb0) {
         constexpr bool last_iter = false;
         flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
-             ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
+             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     }
     { // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
         constexpr bool last_iter = true;
         flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
-             ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1);
+             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     }
 
     // With multi-stage loading there is no __syncthreads at the end of the iter,
@@ -1206,6 +1206,7 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -1214,31 +1215,13 @@ static __global__ void flash_attn_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
 #if defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
 
     // Skip unused kernel variants for faster compilation:
@@ -1274,8 +1257,8 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kb_niter = FATTN_KQ_STRIDE / c::nbatch_fa; // Number of kernel iterations per assigned KQ slice.
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int       kbc      = (blockIdx.x + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
-    const int kbc_stop = (blockIdx.x + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+    int       kbc      = (blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc_stop = (blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
 
     // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined.
     // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup).
@@ -1285,21 +1268,26 @@ static __global__ void flash_attn_ext_f16(
     int kb0_start = kbc % iter_k;
     int kb0_stop  = min(iter_k, kb0_start + kbc_stop - kbc);
     while (kbc < kbc_stop && kb0_stop == iter_k) {
-        const int channel = kbc / (iter_k*iter_j);
-        const int jt      = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile.
+        const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
+        const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
+        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
 
-        const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
-        const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
+        const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02*(head*ncols2));
+        const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head*ncols2 / gqa_ratio));
         const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
-            (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
-        float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+            (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
+        float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head*ncols2) * (DV/2);
 
-        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head*ncols2 / gqa_ratio));
 
-        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
+        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
 
         const int kb0_start_kernel = kb0_start * kb_niter;
-        const int kb0_stop_kernel  = kb0_stop  * kb_niter;
+        int       kb0_stop_kernel  = kb0_stop  * kb_niter;
+
+        if (KV_max) {
+            kb0_stop_kernel = min(kb0_stop_kernel, KV_max[sequence*iter_j + jt] / c::nbatch_fa);
+        }
 
         constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         if (kb0_start == 0) {
@@ -1325,21 +1313,26 @@ static __global__ void flash_attn_ext_f16(
         return;
     }
 
-    const int channel = kbc / (iter_k*iter_j);
-    const int jt      = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile.
+    const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
+    const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
+    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
 
-    const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
-    const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
+    const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02*(head*ncols2));
+    const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head*ncols2 / gqa_ratio));
     const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
-        (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
-    float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+        (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
+    float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head*ncols2) * (DV/2);
 
-    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head*ncols2 / gqa_ratio));
 
-    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
+    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
 
     const int kb0_start_kernel = kb0_start * kb_niter;
-    const int kb0_stop_kernel  = kb0_stop  * kb_niter;
+    int       kb0_stop_kernel  = kb0_stop  * kb_niter;
+
+    if (KV_max) {
+        kb0_stop_kernel = min(kb0_stop_kernel, KV_max[sequence*iter_j + jt] / c::nbatch_fa);
+    }
 
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
@@ -1348,15 +1341,16 @@ static __global__ void flash_attn_ext_f16(
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
-    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
+    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
 }
@@ -1408,24 +1402,24 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
         constexpr bool use_logit_softcap = false;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
             CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
             CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     }
 
     launch_fattn<DV, ncols1, ncols2>

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

@@ -5,14 +5,15 @@
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 __launch_bounds__(nwarps*WARP_SIZE, 2)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -21,31 +22,13 @@ static __global__ void flash_attn_tile_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
 #if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 
     // Skip unused kernel variants for faster compilation:
@@ -62,15 +45,17 @@ static __global__ void flash_attn_tile_ext_f16(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const float2 * Q_f2  = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -106,7 +91,8 @@ static __global__ void flash_attn_tile_ext_f16(
 
     __syncthreads();
 
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         half kqmax_new[ncols/nwarps];
@@ -123,7 +109,7 @@ static __global__ void flash_attn_tile_ext_f16(
             for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
                 const int k_KQ = k_KQ_0 + threadIdx.x;
 
-                KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+                KV_tmp[i_KQ][k_KQ] = K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
             }
         }
 
@@ -217,7 +203,7 @@ static __global__ void flash_attn_tile_ext_f16(
             for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
                 const int i = i0 + threadIdx.x;
 
-                KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
+                KV_tmp[k][i] = V_h2[int64_t(k_VKQ_0 + k)*stride_KV2 + i];
             }
         }
 
@@ -255,6 +241,8 @@ static __global__ void flash_attn_tile_ext_f16(
         __syncthreads();
     }
 
+    float2 * dst2 = (float2 *) dst;
+
 #pragma unroll
     for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
         const int j_VKQ = j_VKQ_0 + threadIdx.y;
@@ -266,21 +254,21 @@ static __global__ void flash_attn_tile_ext_f16(
         half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
         kqsum_j = warp_reduce_sum((float)kqsum_j);
 
-#pragma unroll
-        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
-            const int i0 = i00 + 2*threadIdx.x;
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
 
-            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+#pragma unroll
+        for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
+            const int i0 = i00 + threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
             if (gridDim.y == 1) {
                 dst_val /= __half2half2(kqsum_j);
             }
-            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] =  __low2float(dst_val);
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = __high2float(dst_val);
+            dst2[j_dst_unrolled*(D/2) + i0] = __half22float2(dst_val);
         }
 
         if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
@@ -290,12 +278,11 @@ static __global__ void flash_attn_tile_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
     GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
     GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(nb23);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -5,14 +5,15 @@
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 __launch_bounds__(nwarps*WARP_SIZE, 2)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -21,31 +22,13 @@ static __global__ void flash_attn_tile_ext_f32(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
 #ifdef FLASH_ATTN_AVAILABLE
 
     // Skip unused kernel variants for faster compilation:
@@ -55,17 +38,16 @@ static __global__ void flash_attn_tile_ext_f32(
 #endif // FP16_MMA_AVAILABLE
     if (use_logit_softcap && !(D == 128 || D == 256)) {
         GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-        GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-        GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+        GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
+        GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
         GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-        GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-        GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-        GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-        GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-        GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+        GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+        GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+        GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+        GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+        GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
         NO_DEVICE_CODE;
         return;
     }
@@ -74,15 +56,17 @@ static __global__ void flash_attn_tile_ext_f32(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const float2 * Q_f2  = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
 
@@ -116,7 +100,8 @@ static __global__ void flash_attn_tile_ext_f32(
 
     __syncthreads();
 
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         float kqmax_new[ncols/nwarps];
@@ -131,7 +116,7 @@ static __global__ void flash_attn_tile_ext_f32(
 
 #pragma unroll
             for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
-                const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
+                const half2 tmp = K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
                 KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
                 KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
             }
@@ -227,8 +212,9 @@ static __global__ void flash_attn_tile_ext_f32(
             for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
                 const int i = i0 + threadIdx.x;
 
-                KV_tmp2[k*(D/2) + i].x =  __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
-                KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+                const half2 tmp = V_h2[int64_t(k_VKQ_0 + k)*stride_KV2 + i];
+                KV_tmp2[k*(D/2) + i].x =  __low2float(tmp);
+                KV_tmp2[k*(D/2) + i].y = __high2float(tmp);
             }
         }
 
@@ -265,6 +251,8 @@ static __global__ void flash_attn_tile_ext_f32(
         __syncthreads();
     }
 
+    float2 * dst2 = (float2 *) dst;
+
 #pragma unroll
     for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
         const int j_VKQ = j_VKQ_0 + threadIdx.y;
@@ -276,37 +264,36 @@ static __global__ void flash_attn_tile_ext_f32(
         float kqsum_j = kqsum[j_VKQ_0/nwarps];
         kqsum_j = warp_reduce_sum(kqsum_j);
 
-#pragma unroll
-        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
-            const int i0 = i00 + 2*threadIdx.x;
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
 
-            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+#pragma unroll
+        for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
+            const int i0 = i00 + threadIdx.x;
+
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
             if (gridDim.y == 1) {
                 dst_val.x /= kqsum_j;
                 dst_val.y /= kqsum_j;
             }
-            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x;
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y;
+            dst2[j_dst_unrolled*(D/2) + i0] = dst_val;
         }
 
         if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
+    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32);
     GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
     GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
     GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

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

@@ -1,6 +1,12 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
+// Currenlty llvm with the amdgcn target dose not support unrolling loops
+// that contain a break that can not be resolved at compile time.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
@@ -10,6 +16,7 @@ static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -18,31 +25,13 @@ static __global__ void flash_attn_vec_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
 #if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 
     // Skip unused kernel variants for faster compilation:
@@ -65,14 +54,16 @@ static __global__ void flash_attn_vec_ext_f16(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.z              + nb01*ic0;
-    K += nb12*(blockIdx.z / gqa_ratio);
-    V += nb22*(blockIdx.z / gqa_ratio);
+    Q += nb03*sequence + nb02* head              + nb01*ic0;
+    K += nb13*sequence + nb12*(head / gqa_ratio);
+    V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
+    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -187,37 +178,22 @@ static __global__ void flash_attn_vec_ext_f16(
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    K     += blockIdx.y*D * nb11;
+    V     += blockIdx.y*D * nb21;
+    maskh += blockIdx.y*D;
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*D,
+             // Increment pointers after each loop:
+             K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
+
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         if (mask) {
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
-                maskh_shared[j*D + tid] = slopeh*maskh[j*ne11 + k_VKQ_0 + tid];
+                maskh_shared[j*D + tid] = slopeh*maskh[j*ne11 + tid];
             }
-
             __syncthreads();
-
-            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
-            // In such cases, skip the KV slice.
-            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
-#ifndef GGML_USE_HIP
-            bool skip = true;
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                    const int i = i0 + threadIdx.x;
-
-                    const float2 tmp = __half22float2(((const half2 *) maskh_shared)[j*(D/2) + i]);
-                    skip = skip && isinf(tmp.x) && isinf(tmp.y);
-                }
-            }
-            if (__all_sync(0xFFFFFFFF, skip)) {
-                __syncthreads();
-                continue;
-            }
-#endif // GGML_USE_HIP
         }
 
         // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
@@ -240,7 +216,7 @@ static __global__ void flash_attn_vec_ext_f16(
 
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
-                half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
+                half sum = vec_dot_KQ(K + i_KQ*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
                 sum = warp_reduce_sum((float)sum);
 
                 if (use_logit_softcap) {
@@ -296,8 +272,8 @@ static __global__ void flash_attn_vec_ext_f16(
             }
 
             half2 V_k;
-            reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k_VKQ_0 + k0 + 0)*nb21, tid);
-            reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k_VKQ_0 + k0 + 1)*nb21, tid);
+            reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k0 + 0)*nb21, tid);
+            reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k0 + 1)*nb21, tid);
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
                 VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
@@ -330,29 +306,30 @@ static __global__ void flash_attn_vec_ext_f16(
         if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
+        dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + tid] = dst_val;
     }
 
     if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
+        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
+    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
 
 template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -1,6 +1,12 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
+// Currenlty llvm with the amdgcn target dose not support unrolling loops
+// that contain a break that can not be resolved at compile time.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
@@ -10,6 +16,7 @@ static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -18,31 +25,13 @@ static __global__ void flash_attn_vec_ext_f32(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
 #ifdef FLASH_ATTN_AVAILABLE
 
     // Skip unused kernel variants for faster compilation:
@@ -53,12 +42,11 @@ static __global__ void flash_attn_vec_ext_f32(
         GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
         GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
         GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
         GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
         GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-        GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-        GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+        GGML_UNUSED(nb23);
         NO_DEVICE_CODE;
         return;
     }
@@ -77,14 +65,16 @@ static __global__ void flash_attn_vec_ext_f32(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.z              + nb01*ic0;
-    K += nb12*(blockIdx.z / gqa_ratio);
-    V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
+    Q += nb03*sequence + nb02* head              + nb01*ic0;
+    K += nb13*sequence + nb12*(head / gqa_ratio);
+    V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
+    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
     constexpr int nwarps = D / WARP_SIZE;
@@ -194,36 +184,22 @@ static __global__ void flash_attn_vec_ext_f32(
 
     float VKQ[ncols] = {0.0f};
 
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    K     += blockIdx.y*D * nb11;
+    V     += blockIdx.y*D * nb21;
+    maskh += blockIdx.y*D;
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*D,
+             // Increment pointers after each loop:
+             K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
+
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         if (mask) {
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
-                maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + k_VKQ_0 + tid]);
+                maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + tid]);
             }
-
             __syncthreads();
-
-            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
-            // In such cases, skip the KV slice.
-            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
-#ifndef GGML_USE_HIP
-            bool skip = true;
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-                for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
-                    const int i = i0 + threadIdx.x;
-
-                    skip = skip && isinf(maskf_shared[j*D + i]);
-                }
-            }
-            if (__all_sync(0xFFFFFFFF, skip)) {
-                __syncthreads();
-                continue;
-            }
-#endif // GGML_USE_HIP
         }
 
         float kqmax_new_arr[ncols];
@@ -242,7 +218,7 @@ static __global__ void flash_attn_vec_ext_f32(
 
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
-                float sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
+                float sum = vec_dot_KQ(K + i_KQ*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
                 sum = warp_reduce_sum(sum);
 
                 if (use_logit_softcap) {
@@ -293,7 +269,7 @@ static __global__ void flash_attn_vec_ext_f32(
                 break;
             }
 
-            const float V_ki = dequantize_1_v(V + (k_VKQ_0 + k)*nb21, tid);
+            const float V_ki = dequantize_1_v(V + k*nb21, tid);
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
                 VKQ[j] += V_ki*KQ[j*D + k];
@@ -326,12 +302,11 @@ static __global__ void flash_attn_vec_ext_f32(
         if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
+        dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + tid] = dst_val;
     }
 
     if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
+        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
@@ -340,15 +315,17 @@ static __global__ void flash_attn_vec_ext_f32(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
     GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
     GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
     GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
     GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
 
 template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

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

@@ -7,7 +7,7 @@
 #include "fattn-wmma-f16.cuh"
 
 #ifdef FP16_MMA_AVAILABLE
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 #include <mma.h>
 #ifdef GGML_USE_MUSA
 namespace wmma = mtmusa::wmma;
@@ -18,7 +18,7 @@ namespace wmma = nvcuda::wmma;
 #undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
 namespace wmma = rocwmma;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 #endif // FP16_MMA_AVAILABLE
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
@@ -29,6 +29,7 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -37,31 +38,13 @@ static __global__ void flash_attn_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int ne32,
-        const int nb31,
-        const int nb32,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
 #if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(D == 128 || D == 256)) {
@@ -95,17 +78,19 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
     constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q    + nb02* blockIdx.z              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K    + nb12*(blockIdx.z / gqa_ratio));
-    const half  * V_h   = (const half  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half  * maskh = (const half  *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const float * Q_f   = (const float *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half  * V_h   = (const half  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half  * maskh = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
     const half2 * mask2 = (const half2 *)  maskh;
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
     const half2 slope2 = make_half2(slopef, slopef);
 
@@ -181,7 +166,8 @@ static __global__ void flash_attn_ext_f16(
     __syncthreads();
 
     // Iterate over ne11 == previous tokens:
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
         // Calculate tile of KQ:
 #pragma unroll
         for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
@@ -193,7 +179,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
             for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
                 frag_a_K K_a;
-                wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+                wmma::load_matrix_sync(K_a, K_h + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
 #pragma unroll
                 for (int j = 0; j < ncols/frag_n; ++j) {
                     wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
@@ -340,7 +326,7 @@ static __global__ void flash_attn_ext_f16(
                 const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
 
                 frag_a_V v_a;
-                wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+                wmma::load_matrix_sync(v_a, V_h + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
 #pragma unroll
                 for (int j = 0; j < ncols/frag_n; ++j) {
                     wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
@@ -400,7 +386,6 @@ static __global__ void flash_attn_ext_f16(
         if (ic0 + j_VKQ >= ne01) {
             return;
         }
-        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
 
         float KQ_rowsum_j;
         if (std::is_same<KQ_acc_t, float>::value) {
@@ -409,6 +394,8 @@ 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;
+
 #pragma unroll
         for (int i0 = 0; i0 < D; i0 += warp_size) {
             const int i = i0 + threadIdx.x;
@@ -419,7 +406,7 @@ static __global__ void flash_attn_ext_f16(
             if (gridDim.y == 1) {
                 dst_val /= KQ_rowsum_j;
             }
-            dst[j_dst*gridDim.z*D + blockIdx.z*D + i] = dst_val;
+            dst[j_dst_unrolled*D + i] = dst_val;
         }
 
         if (gridDim.y == 1 || threadIdx.x != 0) {
@@ -433,7 +420,7 @@ static __global__ void flash_attn_ext_f16(
             dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
         }
         dst_meta_val.y = KQ_rowsum_j;
-        dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val;
+        dst_meta[j_dst_unrolled] = dst_meta_val;
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
@@ -442,10 +429,10 @@ static __global__ void flash_attn_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
     GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33); GGML_UNUSED(nb31);
+    GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
     GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
 }
@@ -561,7 +548,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten
         return;
     }
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
     if (Q->ne[1] <= 8 && Q->ne[0] % warp_size == 0) {
         constexpr int cols_per_block = 8;
         switch (Q->ne[0]) {
@@ -583,7 +570,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten
         }
         return;
     }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 
     if (Q->ne[1] <= 32) {
         constexpr int cols_per_block = 16;

ggml/src/ggml-cuda/fattn.cu

@@ -280,22 +280,12 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
-    if (GGML_CUDA_CC_IS_AMD(cc)) {
 #if defined(GGML_HIP_ROCWMMA_FATTN)
-        if (fp16_mma_available(cc)) {
-            ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
-            return;
-        }
-#endif // defined(GGML_HIP_ROCWMMA_FATTN)
-
-        // On AMD the tile kernels perform poorly, use the vec kernel instead:
-        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
-            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-        } else {
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-        }
+    if (GGML_CUDA_CC_IS_AMD(cc) && fp16_mma_available(cc)) {
+        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
         return;
     }
+#endif // defined(GGML_HIP_ROCWMMA_FATTN)
 
     if (!fast_fp16_available(cc)) {
         if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
@@ -325,7 +315,8 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
     const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
-    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies &&
+        (Q->ne[3] > 1 || cc < GGML_CUDA_CC_ADA_LOVELACE) && !mma_needs_data_conversion;
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
     if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {

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

@@ -31,7 +31,9 @@
 #include "ggml-cuda/pool2d.cuh"
 #include "ggml-cuda/quantize.cuh"
 #include "ggml-cuda/rope.cuh"
+#include "ggml-cuda/roll.cuh"
 #include "ggml-cuda/scale.cuh"
+#include "ggml-cuda/softcap.cuh"
 #include "ggml-cuda/softmax.cuh"
 #include "ggml-cuda/ssm-conv.cuh"
 #include "ggml-cuda/ssm-scan.cuh"
@@ -55,6 +57,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
+#include <initializer_list>
 #include <limits>
 #include <map>
 #include <memory>
@@ -125,7 +128,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     return err;
 }
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 static int ggml_cuda_parse_id(char devName[]) {
     // A list of possible Target IDs can be found under the rocclr/clr repo in device.cpp
     // these values are not stable so this is susceptible to breakage
@@ -172,10 +175,10 @@ static int ggml_cuda_parse_id(char devName[]) {
     archNum += archMinor;
     return archNum;
 }
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 
 static ggml_cuda_device_info ggml_cuda_init() {
-#ifdef __HIP_PLATFORM_AMD__
+#if defined(GGML_USE_HIP)
     // Workaround for a rocBLAS bug when using multiple graphics cards:
     // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
     {
@@ -248,7 +251,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].nsm        = prop.multiProcessorCount;
         info.devices[id].smpb       = prop.sharedMemPerBlock;
         info.devices[id].warp_size  = prop.warpSize;
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
         info.devices[id].cc = ggml_cuda_parse_id(prop.gcnArchName);
@@ -278,7 +281,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
         GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                         id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -2418,6 +2421,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ROPE_BACK:
             ggml_cuda_op_rope_back(ctx, dst);
             break;
+        case GGML_OP_ROLL:
+            ggml_cuda_op_roll(ctx, dst);
+            break;
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
@@ -2590,6 +2596,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
     cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
+    const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
+    const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
+
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
@@ -2611,9 +2620,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
-            // disable CUDA graphs for batch size > 1 for now.
-            // Changes in batch size or context size can cause changes to the grid size of some kernels.
+        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) && (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true)) {
+            // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
+            // by means of matching node names. See
+            // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
+            // https://github.com/huggingface/transformers/blob/bda75b4011239d065de84aa3e744b67ebfa7b245/src/transformers/models/gemma3n/modeling_gemma3n.py#L1773,
+            // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.
             use_cuda_graph = false;
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
@@ -2759,6 +2771,67 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif
 
+static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops, std::initializer_list<enum ggml_unary_op> unary_ops) {
+#ifndef NDEBUG
+    const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
+    GGML_ASSERT(unary_ops.size() == num_unary);
+#endif
+
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        const ggml_tensor *rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul      = cgraph->nodes[node_idx+1];
+
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+
+        //rms norm only supports F32
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        //if rms norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+            return false;
+        }
+
+        //rms_norm kernel assumes contigous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+
+        return true;
+    }
+
+    if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SCALE && ops.begin()[1] == GGML_OP_UNARY && ops.begin()[2] == GGML_OP_SCALE
+     && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_TANH) {
+        const ggml_tensor *scale  = cgraph->nodes[node_idx];
+        const ggml_tensor *tanh   = cgraph->nodes[node_idx+1];
+        const ggml_tensor *scale2 = cgraph->nodes[node_idx+2];
+
+        GGML_ASSERT(scale->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(scale->type == GGML_TYPE_F32);
+
+        if (ggml_get_unary_op(tanh) != GGML_UNARY_OP_TANH) {
+            return false;
+        }
+
+        // Check for bias
+        if (ggml_get_op_params_f32(scale, 1) != 0.0f || ggml_get_op_params_f32(scale2, 1) != 0.0f) {
+            return false;
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
     // flag used to determine whether it is an integrated_gpu
@@ -2768,6 +2841,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         // 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) {
+
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
 
@@ -2775,6 +2849,20 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     continue;
                 }
 
+                static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
+                if (!disable_fusion) {
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL }, {})) {
+                        ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
+                        i += 2;
+                        ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i], node);
+                        continue;
+                    }
+                }
 #ifndef NDEBUG
                 assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -3226,8 +3314,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             } break;
         case GGML_OP_SET_ROWS:
             {
-#pragma message("TODO: implement Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
-                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16) &&
+                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
+                       op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
+                       op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
                        op->src[0]->type == GGML_TYPE_F32 &&
                        op->src[1]->type == GGML_TYPE_I64;
             } break;
@@ -3235,13 +3324,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                if ((src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_F16) &&
+                    (src1_type == GGML_TYPE_F32 || src1_type == GGML_TYPE_BF16 || src1_type == GGML_TYPE_F16)
+                ) {
                     return true;
                 }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
@@ -3277,12 +3362,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }
@@ -3363,7 +3442,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return op->src[0]->ne[1] % 128 == 0;
         }
         case GGML_OP_CONT:
-            return op->src[0]->type != GGML_TYPE_BF16;
+            return true;
         case GGML_OP_DIAG_MASK_INF:
             return true;
         case GGML_OP_SOFT_MAX:
@@ -3373,6 +3452,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
             return max_bias == 0.0f;
         }
+        case GGML_OP_ROLL:
+            if(op->src[0]->type == GGML_TYPE_F32) {
+                return true;
+            }
+            return false;
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK: {
             return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
@@ -3413,12 +3497,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
-            // TODO: support broadcast
-            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
-            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
             if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                 return false;
             }
@@ -3431,6 +3509,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
                 return true;
             }
+            if (op->src[3] && op->src[3]->ne[2] != 1) {
+                return false;
+            }
             return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
                 op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
         }

ggml/src/ggml-cuda/im2col.cu

@@ -10,7 +10,7 @@ static  __global__ void im2col_kernel(
         return;
     }
 
-    const int64_t  ksize = OW * (KH > 1 ? KW : 1);
+    const int64_t  ksize = OW * KH;
     const int64_t  kx = i / ksize;
     const int64_t  kd = kx * ksize;
     const int64_t  ky = (i - kd) / OW;

ggml/src/ggml-cuda/mma.cuh

@@ -12,7 +12,8 @@
 // The methods get_i and get_j can be used to get the physical 32 bit index of the lth element of a thread within a tile.
 // All matrix tiles have ne physical 32 bit elements per warp.
 //
-// As described in the documentation, all pointers for load_ldmatrix must be to shared memory and aligned to 16 bytes.
+// As described in the PTX documentation, all pointers for load_ldmatrix must be to shared memory and aligned to 16 bytes.
+// The API in this file also assumes that the pointers for load_generic are aligned to 16 bytes, unaligned pointers are considered undefined behavior.
 
 #include "common.cuh"
 
@@ -66,7 +67,44 @@ namespace ggml_cuda_mma {
     struct tile {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
-        static constexpr int ne = I * J / WARP_SIZE;
+
+#if defined(GGML_USE_HIP)
+        static constexpr int ne = I * J / 64;
+        T x[ne] = {0};
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+                return threadIdx.x % 16;
+            } else if constexpr (I == 16 && J == 8) {
+                return threadIdx.x % 16;
+            } else if constexpr (I == 32 && J == 4) {
+                return threadIdx.x % 32;
+            } else if constexpr (I == 16 && J == 16) {
+                return 4 * (threadIdx.x / 16) + l;
+            } else if constexpr (I == 32 && J == 32) {
+                return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
+            } else {
+                static_assert(I == -1 && J == -1, "template specialization not implemented");
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+                return (2 * ((threadIdx.x / 16) % 2) + l);
+            } else if constexpr (I == 16 && J == 8) {
+                return 2 * (threadIdx.x / 16) + l;
+            } else if constexpr (I == 32 && J == 4) {
+                return 2 * (threadIdx.x / 32) + l;
+            } else if constexpr (I == 16 && J == 16) {
+                return threadIdx.x % 16;
+            } else if constexpr (I == 32 && J == 32) {
+                return threadIdx.x % 32;
+            } else {
+                static_assert(I == -1 && J == -1, "template specialization not implemented");
+            }
+        }
+#else
+        static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
         static __device__ __forceinline__ int get_i(const int l) {
@@ -94,6 +132,7 @@ namespace ggml_cuda_mma {
                 static_assert(I == -1 && J == -1, "template specialization not implemented");
             }
         }
+#endif // defined(GGML_USE_HIP)
     };
 
     template <int I_, int J_>
@@ -148,10 +187,23 @@ namespace ggml_cuda_mma {
 
     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) {
+#if defined(AMD_MFMA_AVAILABLE)
+        if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+#pragma unroll
+            for (int l = 0; l < t.ne; ++l) {
+                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
+            }
+        } else {
+            int64_t * xi = (int64_t *) t.x;
+            const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
+            xi[0] = xs[0];
+        }
+#else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
             t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
         }
+#endif // defined(AMD_MFMA_AVAILABLE)
     }
 
     template <typename T>
@@ -186,7 +238,7 @@ namespace ggml_cuda_mma {
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
-#ifdef NEW_MMA_AVAILABLE
+#if defined(NEW_MMA_AVAILABLE)
         int * xi = (int * ) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
@@ -393,4 +445,60 @@ namespace ggml_cuda_mma {
         NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
     }
+
+    static __device__ __forceinline__ void mma(
+            tile<16, 16, int> & D, const tile<16, 8, int> & A, const tile<16, 8, int> & B) {
+#if defined(AMD_MFMA_AVAILABLE)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * acc = (int32x4_t *) D.x;
+#if defined(CDNA3)
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0],
+                                                       ((int64_t *) B.x)[0],
+                                                       acc[0],
+                                                       0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA)
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0],
+                                                      B.x[0],
+                                                      acc[0],
+                                                      0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1],
+                                                      B.x[1],
+                                                      acc[0],
+                                                      0, 0, 0);
+#endif // defined(CDNA3)
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif // AMD_MFMA_AVAILABLE
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 32, int> & D, const tile<32, 4, int> & A, const tile<32, 4, int> & B) {
+#if defined(AMD_MFMA_AVAILABLE)
+        using int32x16_t = __attribute__((__vector_size__(16 * sizeof(int)))) int;
+        int32x16_t * acc = (int32x16_t *) D.x;
+#if defined(CDNA3)
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0],
+                                                       ((int64_t *) B.x)[0],
+                                                       acc[0],
+                                                       0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA)
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0],
+                                                     B.x[0],
+                                                     acc[0],
+                                                     0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1],
+                                                     B.x[1],
+                                                     acc[0],
+                                                     0, 0, 0);
+#endif // defined(CDNA3)
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif // AMD_MFMA_AVAILABLE
+    }
 }

ggml/src/ggml-cuda/mmq.cu

@@ -109,7 +109,8 @@ void ggml_cuda_mul_mat_q(
     const int64_t s03 = src0->nb[3] / ts_src0;
     const int64_t s3  =  dst->nb[3] / ts_dst;
 
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
+    const bool use_stream_k = (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
+                            || GGML_CUDA_CC_IS_CDNA(cc);
 
     if (!ids) {
         const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
@@ -250,8 +251,9 @@ void ggml_cuda_op_mul_mat_q(
     // The stream-k decomposition is only faster for recent NVIDIA GPUs.
     // Also its fixup needs to allocate a temporary buffer in the memory pool.
     // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
+    const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
+                            || GGML_CUDA_CC_IS_CDNA(cc))
+                            && src1_ncols == ne11;
     const mmq_args args = {
         src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
         ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst,
@@ -320,5 +322,21 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
+    if (amd_mfma_available(cc)) {
+        // As of ROCM 7.0 rocblas/tensile performs very poorly on CDNA3 and hipblaslt (via ROCBLAS_USE_HIPBLASLT)
+        // performs better but is currently suffering from a crash on this architecture.
+        // TODO: Revisit when hipblaslt is fixed on CDNA3
+        if (GGML_CUDA_CC_IS_CDNA3(cc)) {
+            return true;
+        }
+        if (ne11 <= 128 || type == GGML_TYPE_Q4_0 || type == GGML_TYPE_Q4_1 || type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1) {
+            return true;
+        }
+        if (ne11 <= 256 && (type == GGML_TYPE_Q4_K || type == GGML_TYPE_Q5_K)) {
+            return true;
+        }
+        return false;
+    }
+
     return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ggml/src/ggml-cuda/norm.cu

@@ -104,10 +104,12 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
     }
 }
 
-template <int block_size>
+template <int block_size, bool do_multiply = false>
 static __global__ void rms_norm_f32(
         const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
-        const int64_t stride_sample, const float eps) {
+        const int64_t stride_sample, const float eps, const float * mul = nullptr, const int64_t mul_stride_row = 0,
+        const int64_t mul_stride_channel = 0, const int64_t mul_stride_sample = 0, const int mul_ncols = 0,
+        const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {
     const int nrows     = gridDim.x;
     const int nchannels = gridDim.y;
 
@@ -119,6 +121,13 @@ static __global__ void rms_norm_f32(
     x   += sample*stride_sample + channel*stride_channel + row*stride_row;
     dst += ((sample*nchannels + channel)*nrows + row)*ncols;
 
+    if constexpr (do_multiply) {
+        const int mul_row = row % mul_nrows;
+        const int mul_channel = channel % mul_nchannels;
+        const int mul_sample = sample % mul_nsamples;
+        mul += mul_sample*mul_stride_sample + mul_channel*mul_stride_channel + mul_row*mul_stride_row;
+    }
+
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
@@ -145,7 +154,12 @@ static __global__ void rms_norm_f32(
     const float scale = rsqrtf(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = scale * x[col];
+        if constexpr (do_multiply) {
+            const int mul_col = col % mul_ncols;
+            dst[col] = scale * x[col] * mul[mul_col];
+        } else {
+            dst[col] = scale * x[col];
+        }
     }
 }
 
@@ -310,10 +324,30 @@ static void rms_norm_f32_cuda(
     const dim3 blocks_num(nrows, nchannels, nsamples);
     if (ncols < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
-        rms_norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        rms_norm_f32<WARP_SIZE, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        rms_norm_f32<1024><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        rms_norm_f32<1024, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+    }
+}
+
+static void rms_norm_mul_f32_cuda(
+        const float * x, const float * mul, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample,
+        const int64_t mul_stride_row, const int64_t mul_stride_channel, const int64_t mul_stride_sample,
+        const int mul_ncols, const int mul_nrows, const int mul_nchannels, const int mul_nsamples,
+        const float eps, cudaStream_t stream) {
+    const dim3 blocks_num(nrows, nchannels, nsamples);
+    if (mul == nullptr) {
+        rms_norm_f32_cuda(x, dst, ncols, nrows, nchannels, nsamples, stride_row, stride_channel, stride_sample, eps, stream);
+        return;
+    }
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE, true><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel, mul_stride_sample, mul_ncols, mul_nrows, mul_nchannels, mul_nsamples);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024, true><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel, mul_stride_sample, mul_ncols, mul_nrows, mul_nchannels, mul_nsamples);
     }
 }
 
@@ -407,6 +441,59 @@ void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     rms_norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream);
 }
 
+void ggml_cuda_op_rms_norm_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * mul_tensor) {
+    const ggml_tensor * rms_norm_src = (ggml_tensor *) dst->src[0];
+    float eps = 0.0f;
+
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    const float * src0_d = (const float *) rms_norm_src->data;
+    const float * mul_d = nullptr;
+    const ggml_tensor * mul_src = nullptr;
+
+    if (mul_tensor->src[0] == dst) {
+        mul_d = (float *) mul_tensor->src[1]->data;
+        mul_src = mul_tensor->src[1];
+    } else if(mul_tensor->src[1] == dst) {
+        mul_d = (float *) mul_tensor->src[0]->data;
+        mul_src = mul_tensor->src[0];
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    float * dst_d = (float *) mul_tensor->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(rms_norm_src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(mul_tensor->type == GGML_TYPE_F32);
+    GGML_ASSERT(eps >= 0.0f);
+
+    const int64_t ne00 = rms_norm_src->ne[0];
+    const int64_t ne01 = rms_norm_src->ne[1];
+    const int64_t ne02 = rms_norm_src->ne[2];
+    const int64_t ne03 = rms_norm_src->ne[3];
+
+    const size_t ts0 = ggml_type_size(rms_norm_src->type);
+    GGML_ASSERT(rms_norm_src->nb[0] == ts0);
+    const int64_t s01 = rms_norm_src->nb[1] / ts0;
+    const int64_t s02 = rms_norm_src->nb[2] / ts0;
+    const int64_t s03 = rms_norm_src->nb[3] / ts0;
+
+    const size_t ts_mul = ggml_type_size(mul_src->type);
+    GGML_ASSERT(mul_src->nb[0] == ts_mul);
+    const int64_t mul_s01 = mul_src->nb[1] / ts_mul;
+    const int64_t mul_s02 = mul_src->nb[2] / ts_mul;
+    const int64_t mul_s03 = mul_src->nb[3] / ts_mul;
+
+    const int mul_ncols     = mul_src->ne[0];
+    const int mul_nrows     = mul_src->ne[1];
+    const int mul_nchannels = mul_src->ne[2];
+    const int mul_nsamples  = mul_src->ne[3];
+
+    rms_norm_mul_f32_cuda(src0_d, mul_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, mul_s01, mul_s02, mul_s03, mul_ncols, mul_nrows, mul_nchannels, mul_nsamples, eps, stream);
+}
+
 void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * grad  = dst->src[0]; // gradients
     const ggml_tensor * src0f = dst->src[1]; // src0 from forward pass

ggml/src/ggml-cuda/norm.cuh

@@ -6,6 +6,8 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_rms_norm_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * mul_tensor);
+
 void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/roll.cu

@@ -0,0 +1,67 @@
+#include "ggml-cuda/common.cuh"
+#include "roll.cuh"
+
+static __forceinline__ __device__ int64_t wrap_index(const int64_t idx, const int64_t ne) {
+    if (idx < 0) {
+        return idx + ne;
+    }
+    if (idx >= ne) {
+        return idx - ne;
+    }
+    return idx;
+}
+
+static __global__ void roll_f32_cuda(const float * __restrict__ src,
+                                     float * __restrict__ dst,
+                                     const int64_t ne00,
+                                     const int64_t ne01,
+                                     const int64_t ne02,
+                                     const int64_t ne03,
+                                     const int     s0,
+                                     const int     s1,
+                                     const int     s2,
+                                     const int     s3) {
+    const int64_t idx        = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t n_elements = ne00 * ne01 * ne02 * ne03;
+
+    if (idx >= n_elements) {
+        return;
+    }
+
+    const int64_t i0 = idx % ne00;
+    const int64_t i1 = (idx / ne00) % ne01;
+    const int64_t i2 = (idx / (ne00 * ne01)) % ne02;
+    const int64_t i3 = (idx / (ne00 * ne01 * ne02)) % ne03;
+
+    const int64_t d0 = wrap_index(i0 - s0, ne00);
+    const int64_t d1 = wrap_index(i1 - s1, ne01);
+    const int64_t d2 = wrap_index(i2 - s2, ne02);
+    const int64_t d3 = wrap_index(i3 - s3, ne03);
+
+    dst[i3 * (ne00 * ne01 * ne02) + i2 * (ne01 * ne00) + i1 * ne00 + i0] =
+        src[d3 * (ne00 * ne01 * ne02) + d2 * (ne01 * ne00) + d1 * ne00 + d0];
+}
+
+void ggml_cuda_op_roll(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    int s0 = dst->op_params[0];
+    int s1 = dst->op_params[1];
+    int s2 = dst->op_params[2];
+    int s3 = dst->op_params[3];
+
+    const ggml_tensor * src0   = dst->src[0];
+    const float *       src0_d = (const float *) dst->src[0]->data;
+    float *             dst_d  = (float *) dst->data;
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_are_same_shape(dst->src[0], dst));
+
+    cudaStream_t stream = ctx.stream();
+
+    int64_t sz         = (ne00 * ne01 * ne02 * ne03);
+    int64_t num_blocks = (sz + CUDA_ROLL_BLOCK_SIZE - 1) / CUDA_ROLL_BLOCK_SIZE;
+
+    roll_f32_cuda<<<num_blocks, CUDA_ROLL_BLOCK_SIZE, 0, stream>>>(
+        src0_d, dst_d, ne00, ne01, ne02, ne03, s0, s1, s2, s3);
+}

ggml/src/ggml-cuda/roll.cuh

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

ggml/src/ggml-cuda/set-rows.cu

@@ -1,26 +1,90 @@
 #include "set-rows.cuh"
+#include "cpy-utils.cuh"
 
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
 template<typename src_t, typename dst_t>
-__device__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
-    GGML_UNUSED(src_f);
-    GGML_UNUSED(dst_f);
+__device__ __forceinline__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
+    convert_flt(src_f, dst_f);
 }
 
-template<>
-__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
-    *dst_h = __float2half(*src_f);
+// Generic quantized set_rows kernel template
+template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+static __global__ void k_set_rows_quant(
+        const float * __restrict__ src0, const int64_t * __restrict__ src1, block_type * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s10, const int64_t s11, const int64_t s12,
+        const int64_t s1, const int64_t s2, const int64_t s3) {
+
+    const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
+
+    if (i >= ne_total) {
+        return;
+    }
+
+    const int64_t i_base = i * qk;
+    const int64_t i03 = i_base / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i_base - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
+
+    const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+
+    const float * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
+    block_type * dst_row_ptr = dst + (dst_row*s1 + i02*s2 + i03*s3) / sizeof(block_type);
+
+    const float * src_block = src0_row + i00;
+    block_type * dst_block = dst_row_ptr + i00 / qk;
+
+    quantize_func(src_block, dst_block);
+
+    GGML_UNUSED(ne10);
+    GGML_UNUSED(ne13);
 }
 
-template<>
-__device__ __forceinline__ void set_rows_1<float, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
-    *dst_b = *src_f;
-}
+// Template dispatch function for quantized set_rows
+template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+static void set_rows_cuda_quant(
+        const float * src0_d, const int64_t * src1_d, block_type * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
 
-template<>
-__device__ __forceinline__ void set_rows_1<float, float>(const float * src_f, float * dst_f) {
-    *dst_f = *src_f;
+    GGML_ASSERT(ne00 % qk == 0);
+    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
+    const int num_blocks = (ne_total + CUDA_SET_ROWS_BLOCK_SIZE - 1) / CUDA_SET_ROWS_BLOCK_SIZE;
+    const dim3 block_size(CUDA_SET_ROWS_BLOCK_SIZE);
+    const dim3 grid_size(num_blocks);
+
+    const int64_t s01 = nb01/sizeof(float);
+    const int64_t s02 = nb02/sizeof(float);
+    const int64_t s03 = nb03/sizeof(float);
+    const int64_t s10 = nb10/sizeof(int64_t);
+    const int64_t s11 = nb11/sizeof(int64_t);
+    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s1  = nb1;
+    const int64_t s2  = nb2;
+    const int64_t s3  = nb3;
+
+    if (ne_total > 0) {
+        k_set_rows_quant<block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            s01, s02, s03,
+            s10, s11, s12,
+            s1, s2, s3);
+    }
 }
 
 template<typename src_t, typename dst_t>
@@ -145,7 +209,67 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             nb1, nb2, nb3,
             stream
         );
+    } else if (dst->type == GGML_TYPE_Q4_0) {
+        set_rows_cuda_quant<block_q4_0, QK4_0, quantize_f32_q4_0_block>(
+            src0_d, src1_d, (block_q4_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q4_1) {
+        set_rows_cuda_quant<block_q4_1, QK4_1, quantize_f32_q4_1_block>(
+            src0_d, src1_d, (block_q4_1*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q5_0) {
+        set_rows_cuda_quant<block_q5_0, QK5_0, quantize_f32_q5_0_block>(
+            src0_d, src1_d, (block_q5_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q5_1) {
+        set_rows_cuda_quant<block_q5_1, QK5_1, quantize_f32_q5_1_block>(
+            src0_d, src1_d, (block_q5_1*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q8_0) {
+        set_rows_cuda_quant<block_q8_0, QK8_0, quantize_f32_q8_0_block>(
+            src0_d, src1_d, (block_q8_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_IQ4_NL) {
+        set_rows_cuda_quant<block_iq4_nl, QK4_NL, quantize_f32_iq4_nl_block>(
+            src0_d, src1_d, (block_iq4_nl*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
     } else {
-        GGML_ABORT("unsupported type");
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
     }
 }

ggml/src/ggml-cuda/softcap.cu

@@ -0,0 +1,34 @@
+#include "softcap.cuh"
+
+static __global__ void softcap_f32(const float * x, float * dst, const float scale, const float softcap, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = tanhf(scale * x[i]) * softcap;
+}
+
+static void softcap_f32_cuda(const float * x, float * dst, const float scale, const float softcap, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SOFTCAP_BLOCK_SIZE - 1) / CUDA_SOFTCAP_BLOCK_SIZE;
+    softcap_f32<<<num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream>>>(x, dst, scale, softcap, k);
+}
+
+// fused GGML_OP_SCALE + GGML_UNARY_OP_TANH + GGML_OP_SCALE
+void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * src) {
+    const ggml_tensor * src0 = src->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float scale;
+    float softcap;
+    memcpy(&scale,   (float *) src->op_params + 0, sizeof(float));
+    memcpy(&softcap, (float *) dst->op_params + 0, sizeof(float));
+
+    softcap_f32_cuda(src0_d, dst_d, scale, softcap, ggml_nelements(src0), stream);
+}

ggml/src/ggml-cuda/softcap.cuh

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

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

@@ -5,10 +5,8 @@
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
 #include <hip/hip_bfloat16.h>
-#ifdef __HIP_PLATFORM_AMD__
 // for rocblas_initialize()
 #include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
 
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
@@ -139,7 +137,7 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION >= 70000000
+#if HIP_VERSION >= 70000000
 #define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
 #define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
@@ -151,7 +149,11 @@
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define cublasComputeType_t hipblasDatatype_t
 #define cudaDataType_t hipblasDatatype_t
-#endif
+#endif // HIP_VERSION >= 7000000
+
+#if !defined(__HIP_PLATFORM_AMD__)
+#error "The HIP backend supports only AMD targets"
+#endif // !defined(__HIP_PLATFORM_AMD__)
 
 #define __CUDA_ARCH__ 1300
 
@@ -160,7 +162,19 @@
 #endif
 
 #if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__)
-#define CDNA
+#define CDNA // For the entire family
+#endif
+
+#if defined(__gfx942__)
+#define CDNA3
+#endif
+
+#if defined(__gfx90a__)
+#define CDNA2
+#endif
+
+#if defined(__gfx908__)
+#define CDNA1
 #endif
 
 #if defined(__GFX12__)
@@ -237,7 +251,7 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne
     return c;
 }
 
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+#if HIP_VERSION < 50600000
 // __shfl_xor() for half2 was added in ROCm 5.6
 static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
     typedef union half2_b32 {
@@ -249,4 +263,4 @@ static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int
     tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
     return tmp.val;
 }
-#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+#endif // HIP_VERSION < 50600000

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

@@ -13,7 +13,7 @@
 #define CUBLAS_OP_N MUBLAS_OP_N
 #define CUBLAS_OP_T MUBLAS_OP_T
 #define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
-#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_TENSOR_OP_MATH
 #define CUDA_R_16F  MUSA_R_16F
 #define CUDA_R_16BF MUSA_R_16BF
 #define CUDA_R_32F  MUSA_R_32F
@@ -29,7 +29,7 @@
 #define cublasSgemm mublasSgemm
 #define cublasStatus_t mublasStatus_t
 #define cublasOperation_t mublasOperation_t
-#define cublasGetStatusString mublasStatus_to_string
+#define cublasGetStatusString mublasGetStatusString
 #define cudaDataType_t musaDataType_t
 #define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess

ggml/src/ggml-hip/CMakeLists.txt

@@ -113,6 +113,10 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (NOT GGML_HIP_MMQ_MFMA)
+    add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
+endif()
+
 if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
 endif()

ggml/src/ggml-impl.h

@@ -73,6 +73,22 @@ static inline int ggml_up(int n, int m) {
     return (n + m - 1) & ~(m - 1);
 }
 
+// TODO: move to ggml.h?
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
 //
 // logging
 //

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

@@ -126,6 +126,7 @@ typedef struct {
     uint64_t nb2;
     uint64_t nb3;
     uint64_t offs;
+    uint64_t o1[8];
 } ggml_metal_kargs_bin;
 
 typedef struct {
@@ -240,7 +241,7 @@ typedef struct {
     float    max_bias;
     float    m0;
     float    m1;
-    uint16_t n_head_log2;
+    int32_t  n_head_log2;
     float    logit_softcap;
 } ggml_metal_kargs_flash_attn_ext;
 
@@ -377,8 +378,16 @@ typedef struct {
 typedef struct {
     int32_t  ne00;
     int32_t  ne00_4;
-    uint64_t nb01;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    eps;
+    int32_t  nef1[3];
+    int32_t  nef2[3];
+    int32_t  nef3[3];
+    uint64_t nbf1[3];
+    uint64_t nbf2[3];
+    uint64_t nbf3[3];
 } ggml_metal_kargs_rms_norm;
 
 typedef struct {
@@ -484,7 +493,7 @@ typedef struct {
     float    max_bias;
     float    m0;
     float    m1;
-    uint32_t n_head_log2;
+    int32_t  n_head_log2;
 } ggml_metal_kargs_soft_max;
 
 typedef struct {
@@ -519,6 +528,7 @@ typedef struct {
     int64_t  n_group;
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
+    int64_t  s_off;
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;

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

@@ -55,6 +55,12 @@ static struct ggml_backend_metal_device_context {
     bool has_residency_sets;
     bool has_bfloat;
     bool use_bfloat;
+    bool use_fusion;
+
+    int debug_fusion;
+
+    // how many times a given op was fused
+    uint64_t fuse_cnt[GGML_OP_COUNT];
 
     size_t max_size;
 
@@ -69,6 +75,9 @@ static struct ggml_backend_metal_device_context {
     /*.has_residency_sets      =*/ false,
     /*.has_bfloat              =*/ false,
     /*.use_bfloat              =*/ false,
+    /*.use_fusion              =*/ true,
+    /*.debug_fusion            =*/ 0,
+    /*.fuse_cnt                =*/ { 0 },
     /*.max_size                =*/ 0,
     /*.name                    =*/ "",
 };
@@ -83,16 +92,14 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
 
     if (ctx->mtl_device == nil) {
         ctx->mtl_device = MTLCreateSystemDefaultDevice();
-    }
 
-    if (ctx->mtl_device) {
         ctx->has_simdgroup_reduction  = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
         ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
 
         ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
 
 #if defined(GGML_METAL_HAS_RESIDENCY_SETS)
-        ctx->has_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == NULL;
+        ctx->has_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil;
 #endif
 
         ctx->has_bfloat  = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
@@ -103,6 +110,14 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
 #else
         ctx->use_bfloat = false;
 #endif
+        ctx->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
+
+        {
+            const char * val = getenv("GGML_METAL_FUSION_DEBUG");
+            ctx->debug_fusion = val ? atoi(val) : 0;
+        }
+
+        memset(ctx->fuse_cnt, 0, sizeof(ctx->fuse_cnt));
 
         ctx->max_size = ctx->mtl_device.maxBufferLength;
 
@@ -122,6 +137,18 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     ctx->mtl_device_ref_count--;
 
     if (ctx->mtl_device_ref_count == 0) {
+        if (ctx->debug_fusion > 0) {
+            fprintf(stderr, "%s: fusion stats:\n", __func__);
+            for (int i = 0; i < GGML_OP_COUNT; i++) {
+                if (ctx->fuse_cnt[i] == 0) {
+                    continue;
+                }
+
+                // note: cannot use ggml_log here
+                fprintf(stderr, "%s: - %s: %" PRIu64 "\n", __func__, ggml_op_name((enum ggml_op) i), ctx->fuse_cnt[i]);
+            }
+        }
+
         if (ctx->mtl_lock) {
             [ctx->mtl_lock release];
             ctx->mtl_lock = nil;
@@ -147,13 +174,27 @@ struct ggml_metal_kernel {
 
 enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_ADD,
-    GGML_METAL_KERNEL_TYPE_ADD_ROW,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_2,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_3,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_4,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_5,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_6,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_7,
+    GGML_METAL_KERNEL_TYPE_ADD_FUSE_8,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_2,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_3,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_4,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_5,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_6,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_7,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_8,
     GGML_METAL_KERNEL_TYPE_SUB,
-    GGML_METAL_KERNEL_TYPE_SUB_ROW,
+    GGML_METAL_KERNEL_TYPE_SUB_ROW_C4,
     GGML_METAL_KERNEL_TYPE_MUL,
-    GGML_METAL_KERNEL_TYPE_MUL_ROW,
+    GGML_METAL_KERNEL_TYPE_MUL_ROW_C4,
     GGML_METAL_KERNEL_TYPE_DIV,
-    GGML_METAL_KERNEL_TYPE_DIV_ROW,
+    GGML_METAL_KERNEL_TYPE_DIV_ROW_C4,
     GGML_METAL_KERNEL_TYPE_REPEAT_F32,
     GGML_METAL_KERNEL_TYPE_REPEAT_F16,
     GGML_METAL_KERNEL_TYPE_REPEAT_I32,
@@ -218,6 +259,8 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,
     GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
+    GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL,
+    GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL_ADD,
     GGML_METAL_KERNEL_TYPE_L2_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
     GGML_METAL_KERNEL_TYPE_NORM,
@@ -1135,13 +1178,27 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         // simd_sum and simd_max requires MTLGPUFamilyApple7
 
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                             add,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                         add_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_2,                      add_fuse_2,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_3,                      add_fuse_3,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_4,                      add_fuse_4,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_5,                      add_fuse_5,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_6,                      add_fuse_6,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_7,                      add_fuse_7,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_8,                      add_fuse_8,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4,                      add_row_c4,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_2,               add_row_c4_fuse_2,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_3,               add_row_c4_fuse_3,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_4,               add_row_c4_fuse_4,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_5,               add_row_c4_fuse_5,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_6,               add_row_c4_fuse_6,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_7,               add_row_c4_fuse_7,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_8,               add_row_c4_fuse_8,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB,                             sub,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW,                         sub_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW_C4,                      sub_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                             mul,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                         mul_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW_C4,                      mul_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                             div,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                         div_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW_C4,                      div_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32,                      repeat_f32,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16,                      repeat_f16,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32,                      repeat_i32,                      true);
@@ -1206,6 +1263,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,                   set_rows_q5_1,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,                 set_rows_iq4_nl,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                        rms_norm,                        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL,                    rms_norm_mul,                    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL_ADD,                rms_norm_mul_add,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_L2_NORM,                         l2_norm,                         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                      group_norm,                      has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                            norm,                            true);
@@ -1893,9 +1952,10 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
     }
 }
 
-static bool ggml_metal_encode_node(
+static int ggml_metal_encode_node(
                         ggml_backend_t   backend,
                                    int   idx,
+                                   int   idx_end,
           id<MTLComputeCommandEncoder>   encoder,
             struct ggml_metal_mem_pool * mem_pool) {
     struct ggml_backend_metal_context        * ctx     = backend->context;
@@ -1903,7 +1963,10 @@ static bool ggml_metal_encode_node(
 
     struct ggml_cgraph * gf = ctx->gf;
 
-    struct ggml_tensor * node = ggml_graph_node(gf, idx);
+    enum ggml_op ops[8];
+
+    struct ggml_tensor ** nodes = ggml_graph_nodes(gf) + idx;
+    struct ggml_tensor *  node  = nodes[0];
 
     //GGML_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, idx, ggml_op_name(node->op));
 
@@ -1913,7 +1976,7 @@ static bool ggml_metal_encode_node(
     struct ggml_tensor * dst  = node;
 
     if (ggml_is_empty(dst)) {
-        return true;
+        return 1;
     }
 
     switch (dst->op) {
@@ -1924,7 +1987,7 @@ static bool ggml_metal_encode_node(
         case GGML_OP_PERMUTE:
             {
                 // noop -> next node
-            } return true;
+            } return 1;
         default:
             {
             } break;
@@ -1991,6 +2054,8 @@ static bool ggml_metal_encode_node(
     id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
     id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
+    int n_fuse = 1;
+
 #if 0
     GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
     if (src0) {
@@ -2062,37 +2127,15 @@ static bool ggml_metal_encode_node(
                 GGML_ASSERT(src0t == GGML_TYPE_F32);
                 GGML_ASSERT(src1t == GGML_TYPE_F32);
 
+                GGML_ASSERT(ggml_is_contiguous_rows(src0));
+                GGML_ASSERT(ggml_is_contiguous_rows(src1));
+
                 const size_t offs = 0;
 
                 bool bcast_row = false;
 
                 id<MTLComputePipelineState> pipeline = nil;
 
-                if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-                    GGML_ASSERT(ggml_is_contiguous(src0));
-
-                    // src1 is a row
-                    GGML_ASSERT(ne11 == 1);
-
-                    switch (dst->op) {
-                        case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
-                        case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break;
-                        case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
-                        case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
-                        default: GGML_ABORT("fatal error");
-                    }
-
-                    bcast_row = true;
-                } else {
-                    switch (dst->op) {
-                        case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
-                        case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break;
-                        case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
-                        case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
-                        default: GGML_ABORT("fatal error");
-                    }
-                }
-
                 ggml_metal_kargs_bin args = {
                     /*.ne00 =*/ ne00,
                     /*.ne01 =*/ ne01,
@@ -2119,12 +2162,119 @@ static bool ggml_metal_encode_node(
                     /*.nb2  =*/ nb2,
                     /*.nb3  =*/ nb3,
                     /*.offs =*/ offs,
+                    /*.o1   =*/ { offs_src1 },
                 };
 
+                // c[0] = add(a,    b[0])
+                // c[1] = add(c[0], b[1])
+                // c[2] = add(c[1], b[2])
+                // ...
+                if (ctx_dev->use_fusion) {
+                    ops[0] = GGML_OP_ADD;
+                    ops[1] = GGML_OP_ADD;
+                    ops[2] = GGML_OP_ADD;
+                    ops[3] = GGML_OP_ADD;
+                    ops[4] = GGML_OP_ADD;
+                    ops[5] = GGML_OP_ADD;
+                    ops[6] = GGML_OP_ADD;
+                    ops[7] = GGML_OP_ADD;
+
+                    size_t offs_fuse;
+                    id<MTLBuffer> id_fuse;
+
+                    // note: in metal, we sometimes encode the graph in parallel so we have to avoid fusing nodes
+                    //       across splits. idx_end indicates the last node in the current split
+                    for (n_fuse = 0; n_fuse <= 6 && idx + n_fuse + 1 < idx_end; ++n_fuse) {
+                        if (!ggml_can_fuse(gf, idx + n_fuse, ops + n_fuse, 2)) {
+                            break;
+                        }
+
+                        if (nodes[n_fuse] != nodes[n_fuse + 1]->src[0]) {
+                            break;
+                        }
+
+                        // b[0] === b[1] === ...
+                        if (!ggml_are_same_layout(nodes[n_fuse]->src[1], nodes[n_fuse + 1]->src[1])) {
+                            break;
+                        }
+
+                        // only fuse nodes if src1 is in the same Metal buffer
+                        id_fuse = ggml_metal_get_buffer(nodes[n_fuse + 1]->src[1], &offs_fuse);
+                        if (id_fuse != id_src1) {
+                            break;
+                        }
+
+                        ctx_dev->fuse_cnt[nodes[n_fuse + 1]->op]++;
+
+                        args.o1[n_fuse + 1] = offs_fuse;
+                    }
+
+                    ++n_fuse;
+
+                    if (ctx_dev->debug_fusion > 1 && n_fuse > 1) {
+                        GGML_LOG_DEBUG("%s: fuse: ADD x %d\n", __func__, n_fuse);
+                    }
+                }
+
+                if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+                    GGML_ASSERT(ggml_is_contiguous(src0));
+
+                    // src1 is a row
+                    GGML_ASSERT(ne11 == 1);
+
+                    switch (dst->op) {
+                        case GGML_OP_ADD:
+                            {
+                                switch (n_fuse) {
+                                    case 1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4       ].pipeline; break;
+                                    case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_2].pipeline; break;
+                                    case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_3].pipeline; break;
+                                    case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_4].pipeline; break;
+                                    case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_5].pipeline; break;
+                                    case 6: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_6].pipeline; break;
+                                    case 7: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_7].pipeline; break;
+                                    case 8: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_8].pipeline; break;
+                                    default: GGML_ABORT("fatal error");
+                                }
+                            } break;
+                        case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW_C4].pipeline; break;
+                        case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW_C4].pipeline; break;
+                        case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW_C4].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    }
+
+                    bcast_row = true;
+                } else {
+                    switch (dst->op) {
+                        case GGML_OP_ADD:
+                            {
+                                switch (n_fuse) {
+                                    case 1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD       ].pipeline; break;
+                                    case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_2].pipeline; break;
+                                    case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_3].pipeline; break;
+                                    case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_4].pipeline; break;
+                                    case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_5].pipeline; break;
+                                    case 6: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_6].pipeline; break;
+                                    case 7: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_7].pipeline; break;
+                                    case 8: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_FUSE_8].pipeline; break;
+                                    default: GGML_ABORT("fatal error");
+                                }
+                            } break;
+                        case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break;
+                        case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
+                        case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    }
+                }
+
+                if (n_fuse > 1) {
+                    id_dst = ggml_metal_get_buffer(nodes[n_fuse - 1], &offs_dst);
+                }
+
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBytes:&args length:sizeof(args) atIndex:0];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_src1 offset:0         atIndex:2];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 if (bcast_row) {
@@ -2132,7 +2282,11 @@ static bool ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } else {
-                    const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
+                    int nth = 32;
+
+                    while (16*nth < ne0 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                        nth *= 2;
+                    }
 
                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                 }
@@ -2257,12 +2411,13 @@ static bool ggml_metal_encode_node(
                     /*.nb2  =*/ pnb2,
                     /*.nb3  =*/ pnb3,
                     /*.offs =*/ offs,
+                    /*.o1   =*/ { offs_src1},
                 };
 
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBytes:&args length:sizeof(args) atIndex:0];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_src1 offset:0         atIndex:2];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
@@ -2764,7 +2919,7 @@ static bool ggml_metal_encode_node(
                 id<MTLBuffer> h_src0 = h_src0 = ggml_metal_mem_pool_alloc(mem_pool, ggml_nbytes(src0));
                 if (!h_src0) {
                     GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, ggml_nbytes(src0));
-                    return false;
+                    return 0;
                 }
 
                 offs_src0 = 0;
@@ -2986,6 +3141,7 @@ static bool ggml_metal_encode_node(
                     /*.n_group      =*/ n_group,
                     /*.n_seq_tokens =*/ n_seq_tokens,
                     /*.n_seqs       =*/ n_seqs,
+                    /*.s_off        =*/ ggml_nelements(src1) * sizeof(float),
                     /*.nb01         =*/ nb01,
                     /*.nb02         =*/ nb02,
                     /*.nb03         =*/ nb03,
@@ -3014,12 +3170,22 @@ static bool ggml_metal_encode_node(
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:7];
                 [encoder setBytes:&args    length:sizeof(args) atIndex:8];
 
+                // One shared memory bucket for each simd group in the threadgroup
+                // NOTE: Metal kernels require the buffer size to be multiple of 16 bytes
+                //  https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
+                if (d_state >= 32) {
+                    GGML_ASSERT((int64_t)(d_state / 32) <= 32);
+                    const int64_t shmem_size = 32;
+                    GGML_ASSERT(d_state <= (int64_t)pipeline.maxTotalThreadsPerThreadgroup);
+                    [encoder setThreadgroupMemoryLength:(shmem_size)*sizeof(float) atIndex:0];
+                }
+
                 if (ne30 == 1) {
                     // Mamba-2
-                    [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_head, n_seqs) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_head, n_seqs) threadsPerThreadgroup:MTLSizeMake(d_state, 1, 1)];
                 } else {
                     GGML_ASSERT(d_inner == 1);
-                    [encoder dispatchThreadgroups:MTLSizeMake(n_head, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    [encoder dispatchThreadgroups:MTLSizeMake(n_head, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(d_state, 1, 1)];
                 }
             } break;
         case GGML_OP_RWKV_WKV6:
@@ -3640,7 +3806,7 @@ static bool ggml_metal_encode_node(
                     id<MTLBuffer> h_src1 = ggml_metal_mem_pool_alloc(mem_pool, s_src1);
                     if (!h_src1) {
                         GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_src1);
-                        return false;
+                        return 0;
                     }
 
                     const int64_t neh0 = ne0;
@@ -3656,7 +3822,7 @@ static bool ggml_metal_encode_node(
                     id<MTLBuffer> h_dst = ggml_metal_mem_pool_alloc(mem_pool, s_dst);
                     if (!h_dst) {
                         GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_dst);
-                        return false;
+                        return 0;
                     }
 
                     // tokens per expert
@@ -3664,7 +3830,7 @@ static bool ggml_metal_encode_node(
                     id<MTLBuffer> h_tpe = ggml_metal_mem_pool_alloc(mem_pool, s_tpe);
                     if (!h_tpe) {
                         GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_tpe);
-                        return false;
+                        return 0;
                     }
 
                     // id map
@@ -3673,7 +3839,7 @@ static bool ggml_metal_encode_node(
                     id<MTLBuffer> h_ids = ggml_metal_mem_pool_alloc(mem_pool, s_ids);
                     if (!h_ids) {
                         GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_ids);
-                        return false;
+                        return 0;
                     }
 
                     {
@@ -4105,12 +4271,95 @@ static bool ggml_metal_encode_node(
         case GGML_OP_RMS_NORM:
             {
                 GGML_ASSERT(ne00 % 4 == 0);
-                GGML_ASSERT(ggml_is_contiguous_1(src0));
+                GGML_ASSERT(ggml_is_contiguous_rows(src0));
 
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
+                ggml_metal_kargs_rms_norm args = {
+                    /*.ne00   =*/ ne00,
+                    /*.ne00_4 =*/ ne00/4,
+                    /*.nb1    =*/ nb1,
+                    /*.nb2    =*/ nb2,
+                    /*.nb3    =*/ nb3,
+                    /*.eps    =*/ eps,
+                    /*.nef1   =*/ { ne01 },
+                    /*.nef2   =*/ { ne02 },
+                    /*.nef3   =*/ { ne03 },
+                    /*.nbf1   =*/ { nb01 },
+                    /*.nbf2   =*/ { nb02 },
+                    /*.nbf3   =*/ { nb03 },
+                };
+
+                size_t offs_fuse[2] = { 0, 0 };
+                id<MTLBuffer> id_fuse[2] = { id_src0, id_src0 };
+
+                // d[0] = rms_norm(a)
+                // d[1] = mul(d[0], b)
+                // d[2] = add(d[1], c)
+                if (ctx_dev->use_fusion) {
+                    ops[0] = GGML_OP_RMS_NORM;
+                    ops[1] = GGML_OP_MUL;
+                    ops[2] = GGML_OP_ADD;
+
+                    for (n_fuse = 0; n_fuse <= 1 && idx + n_fuse + 1 < idx_end; ++n_fuse) {
+                        if (!ggml_can_fuse(gf, idx + n_fuse, ops + n_fuse, 2)) {
+                            break;
+                        }
+
+                        if (nodes[n_fuse] != nodes[n_fuse + 1]->src[0]) {
+                            break;
+                        }
+
+                        if (nodes[n_fuse + 1]->src[1]->ne[0] != node->ne[0]) {
+                            break;
+                        }
+
+                        if (!ggml_is_contiguous_rows(nodes[n_fuse + 1]->src[1])) {
+                            break;
+                        }
+
+                        if (nodes[n_fuse + 1]->type != GGML_TYPE_F32) {
+                            break;
+                        }
+
+                        ctx_dev->fuse_cnt[nodes[n_fuse + 1]->op]++;
+
+                        id_fuse[n_fuse] = ggml_metal_get_buffer(nodes[n_fuse + 1]->src[1], &offs_fuse[n_fuse]);
+
+                        args.nef1[n_fuse + 1] = nodes[n_fuse + 1]->src[1]->ne[1];
+                        args.nef2[n_fuse + 1] = nodes[n_fuse + 1]->src[1]->ne[2];
+                        args.nef3[n_fuse + 1] = nodes[n_fuse + 1]->src[1]->ne[3];
+
+                        args.nbf1[n_fuse + 1] = nodes[n_fuse + 1]->src[1]->nb[1];
+                        args.nbf2[n_fuse + 1] = nodes[n_fuse + 1]->src[1]->nb[2];
+                        args.nbf3[n_fuse + 1] = nodes[n_fuse + 1]->src[1]->nb[3];
+                    }
+
+                    ++n_fuse;
+
+                    if (ctx_dev->debug_fusion > 1 && n_fuse > 1) {
+                        if (n_fuse == 2) {
+                            GGML_LOG_DEBUG("%s: fuse: RMS_NORM + MUL\n", __func__);
+                        }
+                        if (n_fuse == 3) {
+                            GGML_LOG_DEBUG("%s: fuse: RMS_NORM + MUL + ADD\n", __func__);
+                        }
+                    }
+                }
+
+                if (n_fuse > 1) {
+                    id_dst = ggml_metal_get_buffer(nodes[n_fuse - 1], &offs_dst);
+                }
+
+                id<MTLComputePipelineState> pipeline;
+
+                switch (n_fuse) {
+                    case 1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM        ].pipeline; break;
+                    case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL    ].pipeline; break;
+                    case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL_ADD].pipeline; break;
+                    default: GGML_ABORT("unsupported n_fuse = %d\n", n_fuse);
+                }
 
                 int nth = 32; // SIMD width
 
@@ -4121,23 +4370,16 @@ static bool ggml_metal_encode_node(
                 nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
-                ggml_metal_kargs_rms_norm args = {
-                    /*.ne00   =*/ ne00,
-                    /*.ne00_4 =*/ ne00/4,
-                    /*.nb01   =*/ nb01,
-                    /*.eps    =*/ eps,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&args length:sizeof(args)       atIndex:0];
+                [encoder setBuffer:id_src0    offset:offs_src0    atIndex:1];
+                [encoder setBuffer:id_fuse[0] offset:offs_fuse[0] atIndex:2];
+                [encoder setBuffer:id_fuse[1] offset:offs_fuse[1] atIndex:3];
+                [encoder setBuffer:id_dst     offset:offs_dst     atIndex:4];
 
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
-                const int64_t nrows = ggml_nrows(src0);
-
-                [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_L2_NORM:
             {
@@ -5532,7 +5774,7 @@ static bool ggml_metal_encode_node(
             }
     }
 
-    return true;
+    return n_fuse;
 }
 
 static enum ggml_status ggml_metal_graph_compute(
@@ -6038,20 +6280,26 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
         struct ggml_metal_mem_pool * mem_pool = ctx->cmd_bufs[cb_idx].mem_pool;
         ggml_metal_mem_pool_reset(mem_pool);
 
-        for (int idx = node_start; idx < node_end; ++idx) {
+        for (int idx = node_start; idx < node_end;) {
             if (should_capture) {
                 [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(ggml_graph_node(ctx->gf, idx)) encoding:NSUTF8StringEncoding]];
             }
 
-            const bool res = ggml_metal_encode_node(backend, idx, encoder, mem_pool);
+            const int res = ggml_metal_encode_node(backend, idx, node_end, encoder, mem_pool);
+            if (idx + res > node_end) {
+                GGML_ABORT("fusion error: nodes spanning multiple encoders have been fused. this indicates a bug in the fusion logic %s",
+                        "https://github.com/ggml-org/llama.cpp/pull/14849");
+            }
 
             if (should_capture) {
                 [encoder popDebugGroup];
             }
 
-            if (!res) {
+            if (res == 0) {
                 break;
             }
+
+            idx += res;
         }
 
         [encoder endEncoding];

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

@@ -832,7 +832,8 @@ enum ggml_sort_order {
 // general-purpose kernel for addition, subtraction, multiplication and division of two tensors
 // pros: works for non-contiguous tensors, supports broadcast across all dims
 // cons: not very efficient
-kernel void kernel_add(
+template <int F>
+kernel void kernel_add_fuse_impl(
         constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
@@ -848,16 +849,39 @@ kernel void kernel_add(
     const int i12 = i02%args.ne12;
     const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
-    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+    device       float * dst_ptr  = (device       float *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+    device const float * src1_ptr[F];
+    for (short j = 0; j < F; ++j) {
+        src1_ptr[j] = (device const float *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+    }
 
     for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
         const int i10 = i0%args.ne10;
-        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) + *((device float *)(src1_ptr + i10*args.nb10));
+
+        float res = src0_ptr[i0];
+
+#pragma unroll
+        for (short j = 0; j < F; ++j) {
+            res += src1_ptr[j][i10];
+        }
+
+        dst_ptr[i0] = res;
     }
 }
 
+typedef decltype(kernel_add_fuse_impl<2>) kernel_add_fuse_t;
+
+template [[host_name("kernel_add")]]        kernel kernel_add_fuse_t kernel_add_fuse_impl<1>;
+template [[host_name("kernel_add_fuse_2")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<2>;
+template [[host_name("kernel_add_fuse_3")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<3>;
+template [[host_name("kernel_add_fuse_4")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<4>;
+template [[host_name("kernel_add_fuse_5")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<5>;
+template [[host_name("kernel_add_fuse_6")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<6>;
+template [[host_name("kernel_add_fuse_7")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<7>;
+template [[host_name("kernel_add_fuse_8")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<8>;
+
 kernel void kernel_sub(
         constant ggml_metal_kargs_bin & args,
         device const char * src0,
@@ -875,7 +899,7 @@ kernel void kernel_sub(
     const int i11 = i01%args.ne11;
 
     device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
     device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
     for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
@@ -900,9 +924,9 @@ kernel void kernel_mul(
     const int i12 = i02%args.ne12;
     const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
-    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
     for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
         const int i10 = i0%args.ne10;
@@ -926,9 +950,9 @@ kernel void kernel_div(
     const int i12 = i02%args.ne12;
     const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
-    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
     for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
         const int i10 = i0%args.ne10;
@@ -970,46 +994,145 @@ template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat
 
 // assumption: src1 is a row
 // broadcast src1 into src0
-kernel void kernel_add_row(
+template <short F>
+kernel void kernel_add_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] + src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res += src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
-kernel void kernel_sub_row(
+typedef decltype(kernel_add_row_c4_fuse_impl<1>) kernel_add_row_c4_fuse_t;
+
+template [[host_name("kernel_add_row_c4")]]        kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<1>;
+template [[host_name("kernel_add_row_c4_fuse_2")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<2>;
+template [[host_name("kernel_add_row_c4_fuse_3")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<3>;
+template [[host_name("kernel_add_row_c4_fuse_4")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<4>;
+template [[host_name("kernel_add_row_c4_fuse_5")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<5>;
+template [[host_name("kernel_add_row_c4_fuse_6")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<6>;
+template [[host_name("kernel_add_row_c4_fuse_7")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<7>;
+template [[host_name("kernel_add_row_c4_fuse_8")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<8>;
+
+template <short F>
+kernel void kernel_sub_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] - src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res -= src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
-kernel void kernel_mul_row(
+typedef decltype(kernel_sub_row_c4_fuse_impl<1>) kernel_sub_row_c4_fuse_t;
+
+template [[host_name("kernel_sub_row_c4")]] kernel kernel_sub_row_c4_fuse_t kernel_sub_row_c4_fuse_impl<1>;
+
+template <short F>
+kernel void kernel_mul_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] * src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res *= src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
-kernel void kernel_div_row(
+typedef decltype(kernel_mul_row_c4_fuse_impl<1>) kernel_mul_row_c4_fuse_t;
+
+template [[host_name("kernel_mul_row_c4")]] kernel kernel_mul_row_c4_fuse_t kernel_mul_row_c4_fuse_impl<1>;
+
+template <short F>
+kernel void kernel_div_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] / src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res /= src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
+typedef decltype(kernel_div_row_c4_fuse_impl<1>) kernel_div_row_c4_fuse_t;
+
+template [[host_name("kernel_div_row_c4")]] kernel kernel_div_row_c4_fuse_t kernel_div_row_c4_fuse_impl<1>;
+
 kernel void kernel_scale(
         device const float * src0,
         device       float * dst,
@@ -1700,10 +1823,16 @@ kernel void kernel_ssm_scan_f32(
         device const void * src5,
         device const void * src6,
         device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
         constant ggml_metal_kargs_ssm_scan & args,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        uint3  tpitg[[thread_position_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgptg[[simdgroups_per_threadgroup]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    const int64_t i0 = tpitg.x;
     const int64_t i1 = 0;
     const int64_t ir = tgpig.x; // current head
     const int64_t i3 = tgpig.y; // current seq
@@ -1718,41 +1847,88 @@ kernel void kernel_ssm_scan_f32(
     const int64_t ng  = args.n_group;
     const int64_t n_t = args.n_seq_tokens;
 
-    const int64_t s_off = nr * nh * n_t * args.n_seqs * sizeof(float);
+    const int64_t s_off = args.s_off;
 
     device const int32_t * ids = (device const int32_t *) src6;
 
-    device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
-    device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+    const int64_t i = i0 + i1*nc;
+    float s0 = s0_buff[i];
+    float s  = s_buff[i];
+
+        device const float * A        = (device const float *) ((device const char *) src3 + ir*args.nb31);
+        device const float * x_block  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i3*args.nb13);
+        device const float * dt_block = (device const float *) ((device const char *) src2 + ir*nb20 + i3*args.nb22);
+        device const float * B_block  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i3*args.nb43);
+        device const float * C_block  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i3*args.nb53);
+        device       float * y_block  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i3*(n_t*nh*nr))*nb00);
 
     for (int64_t i2 = 0; i2 < n_t; ++i2) {
-        device const float * x  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i2*args.nb12 + i3*args.nb13); // {dim, nh, nt, ns}
-        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*args.nb21 + i3*args.nb22); // {nh, nt, ns}
-        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31); // {d_state, nh}
-        device const float * B  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i2*args.nb42 + i3*args.nb43); // {d_state, ng, nt, ns}
-        device const float * C  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i2*args.nb52 + i3*args.nb53); // {d_state, ng, nt, ns}
-        device       float * y  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i2*(nh*nr) + i3*(n_t*nh*nr))*nb00); // {dim, nh, nt, ns}
+        device const float * x  = (device const float *) ((device const char *) x_block + i2*args.nb12);    // {dim, nh, nt, ns}
+        device const float * dt = (device const float *) ((device const char *) dt_block + i2*args.nb21);   // {nh, nt, ns}
+        device const float * B  = (device const float *) ((device const char *) B_block + i2*args.nb42);    // {d_state, ng, nt, ns}
+        device const float * C  = (device const float *) ((device const char *) C_block + i2*args.nb52);    // {d_state, ng, nt, ns}
+        device       float * y  = (device       float *) ((device       char *) y_block + i2*(nh*nr*nb00)); // {dim, nh, nt, ns}
 
         const float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
         const float x_dt = x[0] * dt_soft_plus;
-        float sumf = 0.0f;
 
-        for (int64_t i0 = 0; i0 < nc; ++i0) {
-            const int64_t i = i0 + i1*nc;
-            const float state = (s0[i] * exp(dt_soft_plus * A[i0])) + (B[i0] * x_dt);
-            sumf += state * C[i0];
-            s[i] = state;
+        const float state = (s0 * exp(dt_soft_plus * A[i0])) + (B[i0] * x_dt);
+        s = state;
+
+        // Parallel sum: This relies on the fact that this kernel will be
+        // dispatched with each threadgroup having (d_state, 1, 1) threads which
+        // are subdivided into SIMD groups of size `sgptg`. The goal is to
+        // compute y = sum({state * C[i] for i in range(d_state)}).
+        // To parallelize this effectively, we first use simd_sum over each SIMD
+        // group to compute the sum of each SIMD group, then place the result in
+        // the SIMD group's indexed bucket in the shared memory. We then sum
+        // over the individual group sums to compute the final sum.
+
+        // Computed for each thread
+        float sumf = state * C[i0];
+
+        // Sum the threads in the simd group => simd sum
+        sumf = simd_sum(sumf);
+
+        if (sgptg > 1) {
+
+            // Once per simd group, place the group sum into the shared buffer
+            if (tiisg == 0) {
+                shared[sgitg] = sumf;
+            }
+
+            // Wait for all threads in the threadgroup to reach this point. This
+            // ensures that all elements of the shared buffer are populated with the
+            // sum of the individual simd groups.
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // For simd group 0 at indices < num simd groups, extract the shared
+            // simd sum
+            sumf = 0.0f;
+            if (sgitg == 0) {
+                if (tiisg < sgptg) {
+                    sumf = shared[tiisg];
+                }
+                sumf = simd_sum(sumf);
+                if (tiisg == 0) {
+                    y[0] = sumf;
+                }
+            }
+        } else if (tiisg == 0) {
+            y[0] = sumf;
         }
 
-        y[0] = sumf;
-
         // recurse
         s0 = s;
     }
+
+    // Assign the final state to the output buffer
+    s_buff[i] = s;
 }
 
 // ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
-// TODO: optimize (e.g. by parallelizing over d_state)
 kernel void kernel_ssm_scan_f32_group(
         device const void * src0,
         device const void * src1,
@@ -1762,10 +1938,16 @@ kernel void kernel_ssm_scan_f32_group(
         device const void * src5,
         device const void * src6,
         device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
         constant ggml_metal_kargs_ssm_scan & args,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        uint3  tpitg[[thread_position_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgptg[[simdgroups_per_threadgroup]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    const int64_t i0 = tpitg.x;
     const int64_t i1 = tgpig.x;
     const int64_t ir = tgpig.y; // current head
     const int64_t i3 = tgpig.z; // current seq
@@ -1780,38 +1962,81 @@ kernel void kernel_ssm_scan_f32_group(
     const int64_t ng  = args.n_group;
     const int64_t n_t = args.n_seq_tokens;
 
-    const int64_t s_off = nr * nh * n_t * args.n_seqs * sizeof(float);
+    const int64_t s_off = args.s_off;
 
     device const int32_t * ids = (device const int32_t *) src6;
 
-    device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
-    device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+    const int64_t i = i0 + i1*nc;
+    float s0 = s0_buff[i];
+    float s  = s_buff[i];
+
+    device const float * A        = (device const float *) ((device const char *) src3 + ir*args.nb31); // {1, nh}
+    device const float * x_block  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i3*args.nb13);
+    device const float * dt_block = (device const float *) ((device const char *) src2 + ir*nb20 + i3*args.nb22);
+    device const float * B_block  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i3*args.nb43);
+    device const float * C_block  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i3*args.nb53);
+    device       float * y_block  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i3*(n_t*nh*nr))*nb00);
 
     for (int64_t i2 = 0; i2 < n_t; ++i2) {
-        device const float * x  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i2*args.nb12 + i3*args.nb13); // {dim, nh, nt, ns}
-        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*args.nb21 + i3*args.nb22); // {nh, nt, ns}
-        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31); // {1, nh}
-        device const float * B  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i2*args.nb42 + i3*args.nb43); // {d_state, ng, nt, ns}
-        device const float * C  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i2*args.nb52 + i3*args.nb53); // {d_state, ng, nt, ns}
-        device       float * y  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i2*(nh*nr) + i3*(n_t*nh*nr))*nb00); // {dim, nh, nt, ns}
+        device const float * x  = (device const float *) ((device const char *) x_block  + i2*args.nb12);    // {dim, nh, nt, ns}
+        device const float * dt = (device const float *) ((device const char *) dt_block + i2*args.nb21);    // {nh, nt, ns}
+        device const float * B  = (device const float *) ((device const char *) B_block  + i2*args.nb42);    // {d_state, ng, nt, ns}
+        device const float * C  = (device const float *) ((device const char *) C_block  + i2*args.nb52);    // {d_state, ng, nt, ns}
+        device       float * y  = (device       float *) ((device       char *) y_block  + i2*(nh*nr*nb00)); // {dim, nh, nt, ns}
 
         const float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
         const float x_dt = x[0] * dt_soft_plus;
         const float dA = exp(dt_soft_plus * A[0]);
-        float sumf = 0.0f;
 
-        for (int64_t i0 = 0; i0 < nc; ++i0) {
-            const int64_t i = i0 + i1*nc;
-            const float state = (s0[i] * dA) + (B[i0] * x_dt);
-            sumf += state * C[i0];
-            s[i] = state;
+        const float state = (s0 * dA) + (B[i0] * x_dt);
+        s = state;
+
+        // Parallel sum: This relies on the fact that this kernel will be
+        // dispatched with each threadgroup having (d_state, 1, 1) threads which
+        // are subdivided into SIMD groups of size `sgptg`. The goal is to
+        // compute y = sum({state * C[i] for i in range(d_state)}).
+        // To parallelize this effectively, we first use simd_sum over each SIMD
+        // group to compute the sum of each SIMD group, then place the result in
+        // the SIMD group's indexed bucket in the shared memory. We then sum
+        // over the individual group sums to compute the final sum.
+
+        // Computed for each thread
+        float sumf = state * C[i0];
+
+        // Sum the threads in the simd group => simd sum
+        sumf = simd_sum(sumf);
+
+        // Once per simd group, place the group sum into the shared buffer
+        if (tiisg == 0) {
+            shared[sgitg] = sumf;
         }
 
-        y[0] = sumf;
+        // Wait for all threads in the threadgroup to reach this point. This
+        // ensures that all elements of the shared buffer are populated with the
+        // sum of the individual simd groups.
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // For simd group 0 at indices < num simd groups, extract the shared
+        // simd sum
+        sumf = 0.0f;
+        if (sgitg == 0) {
+            if (tiisg < sgptg) {
+                sumf = shared[tiisg];
+            }
+            sumf = simd_sum(sumf);
+            if (tiisg == 0) {
+                y[0] = sumf;
+            }
+        }
 
         // recurse
         s0 = s;
     }
+
+    // Assign the final state to the output buffer
+    s_buff[i] = s;
 }
 
 kernel void kernel_rwkv_wkv6_f32(
@@ -2116,26 +2341,39 @@ kernel void kernel_norm(
     }
 }
 
-kernel void kernel_rms_norm(
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <short F>
+kernel void kernel_rms_norm_fuse_impl(
         constant ggml_metal_kargs_rms_norm & args,
         device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
         device       char * dst,
         threadgroup float * shmem_f32 [[threadgroup(0)]],
-        uint   tgpig[[threadgroup_position_in_grid]],
-        ushort tpitg[[thread_position_in_threadgroup]],
-        ushort sgitg[[simdgroup_index_in_threadgroup]],
-        ushort tiisg[[thread_index_in_simdgroup]],
-        ushort   ntg[[threads_per_threadgroup]]) {
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
     if (sgitg == 0) {
         shmem_f32[tiisg] = 0.0f;
     }
 
-    device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01);
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const float4 * x = (device const float4 *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const float4 * f0 = (device const float4 *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const float4 * f1 = (device const float4 *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
 
     float sumf = 0.0f;
 
     // parallel sum
-    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00_4; i00 += ntg.x) {
         sumf += dot(x[i00], x[i00]);
     }
     sumf = simd_sum(sumf);
@@ -2154,12 +2392,26 @@ kernel void kernel_rms_norm(
     const float mean  = sumf/args.ne00;
     const float scale = 1.0f/sqrt(mean + args.eps);
 
-    device float4 * y = (device float4 *) dst + tgpig*args.ne00_4;
-    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
-        y[i00] = x[i00] * scale;
+    device float4 * y = (device float4 *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_4; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
     }
 }
 
+typedef decltype(kernel_rms_norm_fuse_impl<1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<1>;
+template [[host_name("kernel_rms_norm_mul")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<2>;
+template [[host_name("kernel_rms_norm_mul_add")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<3>;
+
 kernel void kernel_l2_norm(
         constant ggml_metal_kargs_l2_norm & args,
         device const char * src0,

ggml/src/ggml-musa/CMakeLists.txt

@@ -34,8 +34,12 @@ if (MUSAToolkit_FOUND)
     list(APPEND GGML_SOURCES_MUSA ${SRCS})
     file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
     list(APPEND GGML_SOURCES_MUSA ${SRCS})
-    file(GLOB   SRCS "../ggml-musa/*.cu")
-    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+
+    if (GGML_MUSA_MUDNN_COPY)
+        file(GLOB   SRCS "../ggml-musa/*.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        add_compile_definitions(GGML_MUSA_MUDNN_COPY)
+    endif()
 
     if (GGML_CUDA_FA_ALL_QUANTS)
         file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
@@ -72,6 +76,10 @@ if (MUSAToolkit_FOUND)
     add_compile_definitions(GGML_USE_MUSA)
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
+    if (GGML_MUSA_GRAPHS)
+        add_compile_definitions(GGML_MUSA_GRAPHS)
+    endif()
+
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
@@ -97,10 +105,16 @@ if (MUSAToolkit_FOUND)
     endif()
 
     if (GGML_STATIC)
-        # TODO: mudnn has not provided static libraries yet
         target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
+        # TODO: mudnn has not provided static libraries yet
+        # if (GGML_MUSA_MUDNN_COPY)
+        #     target_link_libraries(ggml-musa PRIVATE mudnn_static)
+        # endif()
     else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas mudnn)
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
+        if (GGML_MUSA_MUDNN_COPY)
+            target_link_libraries(ggml-musa PRIVATE mudnn)
+        endif()
     endif()
 
     if (GGML_CUDA_NO_VMM)